footers
· 1 year ago
17f444c4be70b39d717103504f8041d42399efd3
Parent:
a794fa476
2 files changed +2402 −865
- ai-frontier.html +1248 −374
- airisk.html +1154 −491
Diff
--- a/ai-frontier.html +++ b/ai-frontier.html @@ -5,8 +5,10 @@ <meta name="viewport" content="width=device-width, initial-scale=1.0" /> <title>AI Frontier Model Builders Cheatsheet</title> - <link rel="icon" href="data:image/svg+xml,<svg xmlns=%22http://www.w3.org/2000/svg%22 viewBox=%220 0 100 100%22><text y=%22.9em%22 font-size=%2290%22>🤖</text></svg>"> - + <link + rel="icon" + href="data:image/svg+xml,<svg xmlns=%22http://www.w3.org/2000/svg%22 viewBox=%220 0 100 100%22><text y=%22.9em%22 font-size=%2290%22>🤖</text></svg>" + /> <!-- SEO Meta Description --> <meta @@ -24,9 +26,9 @@ content="Explore the philosophies, origins, approaches, and AGI/ASI goals of leading AI companies like OpenAI, Google DeepMind, Anthropic, Meta AI, Cohere, Mistral AI, and AI21 Labs." /> <meta property="og:type" content="article" /> - <meta property="og:url" content="https://cheatsheets.davidveksler.com/ai-frontier.html"> - <meta property="og:image" content="https://cheatsheets.davidveksler.com//images/ai-frontiers.png"> - <meta property="og:image:alt" content="AI Frontier Model Builders Cheatsheet Preview"> + <meta property="og:url" content="https://cheatsheets.davidveksler.com/ai-frontier.html" /> + <meta property="og:image" content="https://cheatsheets.davidveksler.com//images/ai-frontiers.png" /> + <meta property="og:image:alt" content="AI Frontier Model Builders Cheatsheet Preview" /> <meta name="twitter:card" content="summary_large_image" /> <meta name="twitter:title" content="AI Frontier Model Builders Cheatsheet (May 2025 Update)" /> @@ -34,8 +36,8 @@ name="twitter:description" content="Explore the philosophies, origins, approaches, and AGI/ASI goals of leading AI companies like OpenAI, Google DeepMind, Anthropic, Meta AI, Cohere, Mistral AI, and AI21 Labs." /> - <meta name="twitter:image" content="https://cheatsheets.davidveksler.com//images/ai-frontiers.png"> - <meta name="twitter:image:alt" content="AI Frontier Model Builders Cheatsheet Preview"> + <meta name="twitter:image" content="https://cheatsheets.davidveksler.com//images/ai-frontiers.png" /> + <meta name="twitter:image:alt" content="AI Frontier Model Builders Cheatsheet Preview" /> <link href="https://cdn.jsdelivr.net/npm/[email protected]/dist/css/bootstrap.min.css" rel="stylesheet" /> <link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/font/bootstrap-icons.min.css" /> @@ -137,7 +139,7 @@ max-width: 900px; margin: auto; } - .page-header .last-updated { + .page-header .last-updated { font-size: 0.9rem; color: var(--ai-text-secondary); margin-top: 0.5rem; @@ -329,13 +331,12 @@ text-decoration: underline; } - .row > * { margin-bottom: 2.5rem; /* Default, adjust if needed */ } /* Ensure last row doesn't have excessive margin */ .schema-container .row:last-child > * { - margin-bottom: 1.5rem; /* Reduced margin for the last items in a section */ + margin-bottom: 1.5rem; /* Reduced margin for the last items in a section */ } /* Key Info Specific Styles */ @@ -362,11 +363,11 @@ .key-info-list a:hover { text-decoration: underline; } - .info-card.type-info .card-content-wrapper { /* Ensure no toggle button for key info */ - padding-bottom: 1.2rem; + .info-card.type-info .card-content-wrapper { + /* Ensure no toggle button for key info */ + padding-bottom: 1.2rem; } - footer { padding-top: 3rem; font-size: 0.9em; @@ -422,55 +423,125 @@ } /* --- Company/Category Color Assignments --- */ - .cat-openai { --ai-category-color: var(--ai-color-openai); } - .cat-deepmind { --ai-category-color: var(--ai-color-deepmind); } - .cat-anthropic { --ai-category-color: var(--ai-color-anthropic); } - .cat-meta { --ai-category-color: var(--ai-color-meta); } - .cat-cohere { --ai-category-color: var(--ai-color-cohere); } - .cat-mistral { --ai-category-color: var(--ai-color-mistral); } - .cat-ai21 { --ai-category-color: var(--ai-color-ai21); } - - .cat-openai .section-title { color: var(--ai-color-openai); } - .cat-openai .info-card.is-highlighted { box-shadow: 0 0 0 3px var(--ai-color-openai), 0 8px 16px rgba(0,0,0,0.25) !important; } + .cat-openai { + --ai-category-color: var(--ai-color-openai); + } + .cat-deepmind { + --ai-category-color: var(--ai-color-deepmind); + } + .cat-anthropic { + --ai-category-color: var(--ai-color-anthropic); + } + .cat-meta { + --ai-category-color: var(--ai-color-meta); + } + .cat-cohere { + --ai-category-color: var(--ai-color-cohere); + } + .cat-mistral { + --ai-category-color: var(--ai-color-mistral); + } + .cat-ai21 { + --ai-category-color: var(--ai-color-ai21); + } - .cat-deepmind .section-title { color: var(--ai-color-deepmind); } - .cat-deepmind .info-card.is-highlighted { box-shadow: 0 0 0 3px var(--ai-color-deepmind), 0 8px 16px rgba(0,0,0,0.25) !important; } + .cat-openai .section-title { + color: var(--ai-color-openai); + } + .cat-openai .info-card.is-highlighted { + box-shadow: 0 0 0 3px var(--ai-color-openai), 0 8px 16px rgba(0, 0, 0, 0.25) !important; + } - .cat-anthropic .section-title { color: var(--ai-color-anthropic); } - .cat-anthropic .info-card.is-highlighted { box-shadow: 0 0 0 3px var(--ai-color-anthropic), 0 8px 16px rgba(0,0,0,0.25) !important; } + .cat-deepmind .section-title { + color: var(--ai-color-deepmind); + } + .cat-deepmind .info-card.is-highlighted { + box-shadow: 0 0 0 3px var(--ai-color-deepmind), 0 8px 16px rgba(0, 0, 0, 0.25) !important; + } - .cat-meta .section-title { color: var(--ai-color-meta); } - .cat-meta .info-card.is-highlighted { box-shadow: 0 0 0 3px var(--ai-color-meta), 0 8px 16px rgba(0,0,0,0.25) !important; } + .cat-anthropic .section-title { + color: var(--ai-color-anthropic); + } + .cat-anthropic .info-card.is-highlighted { + box-shadow: 0 0 0 3px var(--ai-color-anthropic), 0 8px 16px rgba(0, 0, 0, 0.25) !important; + } - .cat-cohere .section-title { color: var(--ai-color-cohere); } - .cat-cohere .info-card.is-highlighted { box-shadow: 0 0 0 3px var(--ai-color-cohere), 0 8px 16px rgba(0,0,0,0.25) !important; } + .cat-meta .section-title { + color: var(--ai-color-meta); + } + .cat-meta .info-card.is-highlighted { + box-shadow: 0 0 0 3px var(--ai-color-meta), 0 8px 16px rgba(0, 0, 0, 0.25) !important; + } - .cat-mistral .section-title { color: var(--ai-color-mistral); } - .cat-mistral .info-card.is-highlighted { box-shadow: 0 0 0 3px var(--ai-color-mistral), 0 8px 16px rgba(0,0,0,0.25) !important; } + .cat-cohere .section-title { + color: var(--ai-color-cohere); + } + .cat-cohere .info-card.is-highlighted { + box-shadow: 0 0 0 3px var(--ai-color-cohere), 0 8px 16px rgba(0, 0, 0, 0.25) !important; + } - .cat-ai21 .section-title { color: var(--ai-color-ai21); } - .cat-ai21 .info-card.is-highlighted { box-shadow: 0 0 0 3px var(--ai-color-ai21), 0 8px 16px rgba(0,0,0,0.25) !important; } + .cat-mistral .section-title { + color: var(--ai-color-mistral); + } + .cat-mistral .info-card.is-highlighted { + box-shadow: 0 0 0 3px var(--ai-color-mistral), 0 8px 16px rgba(0, 0, 0, 0.25) !important; + } + .cat-ai21 .section-title { + color: var(--ai-color-ai21); + } + .cat-ai21 .info-card.is-highlighted { + box-shadow: 0 0 0 3px var(--ai-color-ai21), 0 8px 16px rgba(0, 0, 0, 0.25) !important; + } /* --- Aspect Type Color Assignments for Card Elements --- */ - .info-card.type-origin { --ai-aspect-color-current: var(--ai-aspect-color-origin); } - .info-card.type-philosophy { --ai-aspect-color-current: var(--ai-aspect-color-philosophy); } - .info-card.type-leadership { --ai-aspect-color-current: var(--ai-aspect-color-leadership); } - .info-card.type-funding { --ai-aspect-color-current: var(--ai-aspect-color-funding); } - .info-card.type-approach { --ai-aspect-color-current: var(--ai-aspect-color-approach); } - .info-card.type-models { --ai-aspect-color-current: var(--ai-aspect-color-models); } - .info-card.type-safety { --ai-aspect-color-current: var(--ai-aspect-color-safety); } - .info-card.type-opensource { --ai-aspect-color-current: var(--ai-aspect-color-opensource); } - .info-card.type-agi { --ai-aspect-color-current: var(--ai-aspect-color-agi); } - .info-card.type-audience { --ai-aspect-color-current: var(--ai-aspect-color-audience); } - .info-card.type-differentiators { --ai-aspect-color-current: var(--ai-aspect-color-differentiators); } - .info-card.type-developments { --ai-aspect-color-current: var(--ai-aspect-color-developments); } - .info-card.type-info { --ai-aspect-color-current: var(--ai-aspect-color-info); } - + .info-card.type-origin { + --ai-aspect-color-current: var(--ai-aspect-color-origin); + } + .info-card.type-philosophy { + --ai-aspect-color-current: var(--ai-aspect-color-philosophy); + } + .info-card.type-leadership { + --ai-aspect-color-current: var(--ai-aspect-color-leadership); + } + .info-card.type-funding { + --ai-aspect-color-current: var(--ai-aspect-color-funding); + } + .info-card.type-approach { + --ai-aspect-color-current: var(--ai-aspect-color-approach); + } + .info-card.type-models { + --ai-aspect-color-current: var(--ai-aspect-color-models); + } + .info-card.type-safety { + --ai-aspect-color-current: var(--ai-aspect-color-safety); + } + .info-card.type-opensource { + --ai-aspect-color-current: var(--ai-aspect-color-opensource); + } + .info-card.type-agi { + --ai-aspect-color-current: var(--ai-aspect-color-agi); + } + .info-card.type-audience { + --ai-aspect-color-current: var(--ai-aspect-color-audience); + } + .info-card.type-differentiators { + --ai-aspect-color-current: var(--ai-aspect-color-differentiators); + } + .info-card.type-developments { + --ai-aspect-color-current: var(--ai-aspect-color-developments); + } + .info-card.type-info { + --ai-aspect-color-current: var(--ai-aspect-color-info); + } /* Apply the current aspect color to relevant card elements */ - .info-card { border-left-color: var(--ai-aspect-color-current); } - .info-card h5 .bi { color: var(--ai-aspect-color-current); } + .info-card { + border-left-color: var(--ai-aspect-color-current); + } + .info-card h5 .bi { + color: var(--ai-aspect-color-current); + } .info-card .details-toggle { color: var(--ai-aspect-color-current); border-color: var(--ai-aspect-color-current); @@ -479,8 +550,9 @@ background-color: var(--ai-aspect-color-current); color: var(--ai-body-bg); /* Ensure contrast on hover */ } - .info-card .collapse-content li::before { color: var(--ai-aspect-color-current); } - + .info-card .collapse-content li::before { + color: var(--ai-aspect-color-current); + } </style> </head> <body> @@ -502,13 +574,28 @@ <h5><i class="bi bi-info-circle-fill"></i> Key Information</h5> <div class="card-content-wrapper"> <ul class="key-info-list"> - <li><strong>Founded:</strong> December 2015, by Elon Musk, Sam Altman, Greg Brockman, Ilya Sutskever, Wojciech Zaremba, John Schulman, and others.</li> + <li> + <strong>Founded:</strong> December 2015, by Elon Musk, Sam Altman, Greg Brockman, Ilya Sutskever, + Wojciech Zaremba, John Schulman, and others. + </li> <li><strong>Headquarters:</strong> San Francisco, California, USA</li> - <li><strong>Valuation:</strong> $157 billion (as of October 2024). [3, 13] Reported talks for $300 billion valuation (April 2025).</li> + <li> + <strong>Valuation:</strong> $157 billion (as of October 2024). [3, 13] Reported talks for $300 + billion valuation (April 2025). + </li> <li><strong>Flagship Models:</strong> GPT-4o, GPT-4, DALL-E 3, Sora, Whisper, o1.</li> <li><strong>Main Products:</strong> ChatGPT, OpenAI API, various specialized models.</li> - <li><strong>Official Website:</strong> <a href="https://openai.com" target="_blank" rel="noopener noreferrer">openai.com</a></li> - <li><strong>Documentation:</strong> <a href="https://platform.openai.com/docs" target="_blank" rel="noopener noreferrer">platform.openai.com/docs</a> [6, 43]</li> + <li> + <strong>Official Website:</strong> + <a href="https://openai.com" target="_blank" rel="noopener noreferrer">openai.com</a> + </li> + <li> + <strong>Documentation:</strong> + <a href="https://platform.openai.com/docs" target="_blank" rel="noopener noreferrer" + >platform.openai.com/docs</a + > + [6, 43] + </li> </ul> </div> </div> @@ -520,7 +607,9 @@ <h5><i class="bi bi-flag-fill"></i> Origin & Founding Vision</h5> <div class="card-content-wrapper"> <p class="summary"> - Founded Dec 2015 as a non-profit, later adopted a "capped-profit" model. Aims to ensure AGI benefits all humanity. Learn more on their <a href="https://openai.com/about" target="_blank" rel="noopener noreferrer">about page</a>. + Founded Dec 2015 as a non-profit, later adopted a "capped-profit" model. Aims to ensure AGI benefits + all humanity. Learn more on their + <a href="https://openai.com/about" target="_blank" rel="noopener noreferrer">about page</a>. </p> <button class="btn btn-sm details-toggle" @@ -537,16 +626,23 @@ <h6>Key Details</h6> <ul> <li> - <strong>Founding Goal:</strong> To build Artificial General Intelligence (AGI) that is safe and broadly beneficial. + <strong>Founding Goal:</strong> To build Artificial General Intelligence (AGI) that is safe and + broadly beneficial. </li> <li><strong>Initial Structure:</strong> Non-profit research company (OpenAI, Inc.).</li> <li> - <strong>Key Founders:</strong> Included Elon Musk, Sam Altman, Greg Brockman, Ilya Sutskever, Wojciech Zaremba, John Schulman. + <strong>Key Founders:</strong> Included Elon Musk, Sam Altman, Greg Brockman, Ilya Sutskever, + Wojciech Zaremba, John Schulman. </li> <li> - <strong>Transition:</strong> In 2019, created OpenAI LP, a "capped-profit" company, to raise capital for compute-intensive research, while the non-profit OpenAI, Inc. remains the overall governing body and its mission is primary. + <strong>Transition:</strong> In 2019, created OpenAI LP, a "capped-profit" company, to raise capital + for compute-intensive research, while the non-profit OpenAI, Inc. remains the overall governing body + and its mission is primary. + </li> + <li> + <strong>Recent Structure:</strong> As of 2025, involves OpenAI, Inc. (non-profit) and for-profit + subsidiaries like OpenAI Global, LLC. </li> - <li><strong>Recent Structure:</strong> As of 2025, involves OpenAI, Inc. (non-profit) and for-profit subsidiaries like OpenAI Global, LLC.</li> </ul> </div> </div> @@ -557,7 +653,9 @@ <h5><i class="bi bi-lightbulb-fill"></i> Philosophy & Culture</h5> <div class="card-content-wrapper"> <p class="summary"> - Balances ambitious research towards AGI with a stated emphasis on safety, responsibility, and broad benefit. Iterative deployment of increasingly powerful systems. Read their <a href="https://openai.com/research" target="_blank" rel="noopener noreferrer">research</a>. + Balances ambitious research towards AGI with a stated emphasis on safety, responsibility, and broad + benefit. Iterative deployment of increasingly powerful systems. Read their + <a href="https://openai.com/research" target="_blank" rel="noopener noreferrer">research</a>. </p> <button class="btn btn-sm details-toggle" @@ -575,22 +673,27 @@ <ul> <li><strong>Beneficial AGI:</strong> Primary mission is to ensure AGI benefits all of humanity.</li> <li> - <strong>Safety Research:</strong> Significant investment in AI safety research and mitigating risks from powerful AI. Developed a "Preparedness Framework" to assess and mitigate catastrophic risks. + <strong>Safety Research:</strong> Significant investment in AI safety research and mitigating risks + from powerful AI. Developed a "Preparedness Framework" to assess and mitigate catastrophic risks. </li> <li> <strong>Long-term Perspective:</strong> Commitment to long, challenging research projects for AGI. </li> <li> - <strong>Iterative Deployment:</strong> Believes in deploying increasingly powerful (but still limited) AI systems to learn from real-world use and adapt, enabling societal adaptation. + <strong>Iterative Deployment:</strong> Believes in deploying increasingly powerful (but still + limited) AI systems to learn from real-world use and adapt, enabling societal adaptation. </li> <li> - <strong>Collaboration & Openness (Evolving):</strong> Started with a strong open-source ethos. Now more selective about model releases, citing safety and competitive concerns, but still releases some models and research (e.g., on <a href="https://github.com/openai" target="_blank" rel="noopener noreferrer">GitHub</a>). + <strong>Collaboration & Openness (Evolving):</strong> Started with a strong open-source ethos. Now + more selective about model releases, citing safety and competitive concerns, but still releases some + models and research (e.g., on + <a href="https://github.com/openai" target="_blank" rel="noopener noreferrer">GitHub</a>). </li> </ul> </div> </div> </div> - <div class="col-lg-4 col-md-6"> + <div class="col-lg-4 col-md-6"> <div class="info-card type-leadership" id="card-openai-leadership"> <div class="card-body"> <h5><i class="bi bi-person-badge"></i> Leadership</h5> @@ -617,9 +720,12 @@ <li><strong>Mira Murati:</strong> Chief Technology Officer (CTO).</li> <li><strong>Bret Taylor:</strong> Chairman of the Board of Directors (OpenAI, Inc. nonprofit).</li> <li><strong>Sarah Friar:</strong> Chief Financial Officer (CFO).</li> - <li><strong>Jakub Pachocki:</strong> Chief Scientist Officer.</li> + <li><strong>Jakub Pachocki:</strong> Chief Scientist Officer.</li> </ul> - <p>Note: Leadership can change. There was a significant leadership shuffle in November 2023, with Altman briefly removed and then reinstated with a new initial board.</p> + <p> + Note: Leadership can change. There was a significant leadership shuffle in November 2023, with Altman + briefly removed and then reinstated with a new initial board. + </p> </div> </div> </div> @@ -629,7 +735,11 @@ <h5><i class="bi bi-boxes"></i> Flagship Models & Products</h5> <div class="card-content-wrapper"> <p class="summary"> - GPT series (GPT-4, GPT-4o), DALL-E 3 (images), Sora (video), Whisper (speech-to-text), <a href="https://chat.openai.com" target="_blank" rel="noopener noreferrer">ChatGPT</a> interface. Recent models like o1 focus on reasoning. Access models via the <a href="https://platform.openai.com" target="_blank" rel="noopener noreferrer">OpenAI API</a>. [6, 10, 25, 40, 43] + GPT series (GPT-4, GPT-4o), DALL-E 3 (images), Sora (video), Whisper (speech-to-text), + <a href="https://chat.openai.com" target="_blank" rel="noopener noreferrer">ChatGPT</a> interface. + Recent models like o1 focus on reasoning. Access models via the + <a href="https://platform.openai.com" target="_blank" rel="noopener noreferrer">OpenAI API</a>. [6, + 10, 25, 40, 43] </p> <button class="btn btn-sm details-toggle" @@ -650,30 +760,39 @@ <ul> <li><code>GPT-3.5</code>: Powers many applications and the free version of ChatGPT.</li> <li> - <code>GPT-4</code>: Highly capable model with improved reasoning, creativity, and longer context. + <code>GPT-4</code>: Highly capable model with improved reasoning, creativity, and longer + context. </li> <li> - <code>GPT-4o ("omni")</code>: Latest flagship (as of May 2024), enhanced multimodality (text, audio, vision), speed, and interaction capabilities. + <code>GPT-4o ("omni")</code>: Latest flagship (as of May 2024), enhanced multimodality (text, + audio, vision), speed, and interaction capabilities. </li> - <li><code>o1</code>: A model focused on enhanced reasoning capabilities.</li> - <li>Development pipeline includes models like <code>o3</code> and <code>o4-mini</code>.</li> + <li><code>o1</code>: A model focused on enhanced reasoning capabilities.</li> + <li>Development pipeline includes models like <code>o3</code> and <code>o4-mini</code>.</li> </ul> </li> <li> - <strong>DALL-E Series (e.g., DALL-E 3):</strong> AI system creating realistic images and art from natural language. - </li> - <li> - <strong>Sora:</strong> AI model generating realistic and imaginative video scenes from text. + <strong>DALL-E Series (e.g., DALL-E 3):</strong> AI system creating realistic images and art from + natural language. </li> + <li><strong>Sora:</strong> AI model generating realistic and imaginative video scenes from text.</li> <li><strong>Whisper:</strong> Versatile speech recognition (ASR) and translation model.</li> </ul> <h6>Access & Products</h6> <ul> <li> - <span class="term"><a href="https://chat.openai.com" target="_blank" rel="noopener noreferrer">ChatGPT</a>:</span> Conversational AI interface (free, Plus, Team, Enterprise tiers). + <span class="term" + ><a href="https://chat.openai.com" target="_blank" rel="noopener noreferrer">ChatGPT</a>:</span + > + Conversational AI interface (free, Plus, Team, Enterprise tiers). </li> <li> - <span class="term"><a href="https://platform.openai.com" target="_blank" rel="noopener noreferrer">OpenAI API</a>:</span> Allows developer integration of models into applications. Includes new Responses API and Agents SDK for building AI agents. [6, 10, 25, 40, 43] + <span class="term" + ><a href="https://platform.openai.com" target="_blank" rel="noopener noreferrer">OpenAI API</a + >:</span + > + Allows developer integration of models into applications. Includes new Responses API and Agents SDK + for building AI agents. [6, 10, 25, 40, 43] </li> <li>Partnerships (e.g., Microsoft Azure, Apple Intelligence).</li> </ul> @@ -686,7 +805,8 @@ <h5><i class="bi bi-bullseye"></i> AGI/ASI Goals & Approach</h5> <div class="card-content-wrapper"> <p class="summary"> - Explicitly aims to build Artificial General Intelligence (AGI) that is safe and beneficial. Pursues this through scaling models and iterative deployment. + Explicitly aims to build Artificial General Intelligence (AGI) that is safe and beneficial. Pursues + this through scaling models and iterative deployment. </p> <button class="btn btn-sm details-toggle" @@ -703,28 +823,37 @@ <h6>Stated Ambition</h6> <ul> <li> - <strong>Core Mission:</strong> The development of AGI is central to OpenAI's charter. Defines AGI as "highly autonomous systems that outperform humans at most economically valuable work." + <strong>Core Mission:</strong> The development of AGI is central to OpenAI's charter. Defines AGI as + "highly autonomous systems that outperform humans at most economically valuable work." </li> <li> - <strong>Safety Emphasis:</strong> AGI development is coupled with a strong focus on ensuring alignment with human values and intentions, and responsible power usage. This includes the "Preparedness Framework" and past projects like Superalignment (though the specific team saw departures). + <strong>Safety Emphasis:</strong> AGI development is coupled with a strong focus on ensuring + alignment with human values and intentions, and responsible power usage. This includes the + "Preparedness Framework" and past projects like Superalignment (though the specific team saw + departures). </li> <li> - <strong>Path to AGI:</strong> Primarily through scaling current deep learning architectures (transformers), combined with new research breakthroughs and continuous safety improvements. Iterative deployment of more capable systems is a key part of the strategy. + <strong>Path to AGI:</strong> Primarily through scaling current deep learning architectures + (transformers), combined with new research breakthroughs and continuous safety improvements. + Iterative deployment of more capable systems is a key part of the strategy. </li> <li> - <strong>ASI Considerations:</strong> Acknowledges the potential for Artificial Superintelligence (ASI) beyond AGI and the profound societal implications, emphasizing the need for careful management and governance. + <strong>ASI Considerations:</strong> Acknowledges the potential for Artificial Superintelligence + (ASI) beyond AGI and the profound societal implications, emphasizing the need for careful management + and governance. </li> </ul> </div> </div> </div> - <div class="col-lg-4 col-md-6"> + <div class="col-lg-4 col-md-6"> <div class="info-card type-funding" id="card-openai-funding"> <div class="card-body"> <h5><i class="bi bi-piggy-bank"></i> Funding & Valuation</h5> <div class="card-content-wrapper"> <p class="summary"> - Substantial backing from Microsoft (~$13B). Raised $6.6B in Oct 2024 (valuing at $157B) and reported talks for $40B in Apr 2025 (valuing at $300B). [3, 13, 23] + Substantial backing from Microsoft (~$13B). Raised $6.6B in Oct 2024 (valuing at $157B) and reported + talks for $40B in Apr 2025 (valuing at $300B). [3, 13, 23] </p> <button class="btn btn-sm details-toggle" @@ -740,11 +869,28 @@ <div class="collapse collapse-content" id="collapseOpenAIFunding"> <h6>Key Investments</h6> <ul> - <li><strong>Microsoft Partnership:</strong> Multi-year, multi-billion dollar investment (reportedly around $13 billion total), including significant Azure cloud computing resources. Microsoft is entitled to a share of profits from OpenAI's for-profit arm. [23]</li> - <li><strong>October 2024 Round:</strong> Secured $6.6 billion, valuing OpenAI at $157 billion. Major investors included Microsoft, Nvidia, and SoftBank. [3, 13]</li> - <li><strong>April 2025 Round:</strong> Reported talks to raise up to $40 billion at a $300 billion post-money valuation, potentially led by SoftBank, with participation from Microsoft, Coatue, Altimeter, and Thrive Capital. This would mark one of the largest private technology deals.</li> - <li><strong>Early Backers:</strong> Initial funding came from Sam Altman, Greg Brockman, Elon Musk, Reid Hoffman, Peter Thiel, AWS, Infosys, and YC Research.</li> - <li><strong>Path to Profitability:</strong> Reports suggest some funding tranches are contingent on OpenAI transitioning to a more conventional for-profit structure. [13]</li> + <li> + <strong>Microsoft Partnership:</strong> Multi-year, multi-billion dollar investment (reportedly + around $13 billion total), including significant Azure cloud computing resources. Microsoft is + entitled to a share of profits from OpenAI's for-profit arm. [23] + </li> + <li> + <strong>October 2024 Round:</strong> Secured $6.6 billion, valuing OpenAI at $157 billion. Major + investors included Microsoft, Nvidia, and SoftBank. [3, 13] + </li> + <li> + <strong>April 2025 Round:</strong> Reported talks to raise up to $40 billion at a $300 billion + post-money valuation, potentially led by SoftBank, with participation from Microsoft, Coatue, + Altimeter, and Thrive Capital. This would mark one of the largest private technology deals. + </li> + <li> + <strong>Early Backers:</strong> Initial funding came from Sam Altman, Greg Brockman, Elon Musk, Reid + Hoffman, Peter Thiel, AWS, Infosys, and YC Research. + </li> + <li> + <strong>Path to Profitability:</strong> Reports suggest some funding tranches are contingent on + OpenAI transitioning to a more conventional for-profit structure. [13] + </li> </ul> </div> </div> @@ -755,7 +901,9 @@ <h5><i class="bi bi-newspaper"></i> Recent Developments (2024-2025)</h5> <div class="card-content-wrapper"> <p class="summary"> - GPT-4o launch, Sora video model access expanded, o1 reasoning model. New Responses API and Agents SDK. Apple partnership. Major funding rounds. Stay updated via their <a href="https://openai.com/blog" target="_blank" rel="noopener noreferrer">blog</a>. + GPT-4o launch, Sora video model access expanded, o1 reasoning model. New Responses API and Agents + SDK. Apple partnership. Major funding rounds. Stay updated via their + <a href="https://openai.com/blog" target="_blank" rel="noopener noreferrer">blog</a>. </p> <button class="btn btn-sm details-toggle" @@ -771,11 +919,26 @@ <div class="collapse collapse-content" id="collapseOpenAIDevelopments"> <h6>Key Announcements</h6> <ul> - <li><strong>Model Releases:</strong> GPT-4o (May 2024) as new flagship. Sora text-to-video model made available to ChatGPT Plus/Pro users (Dec 2024). OpenAI o1 reasoning model launched (Dec 2024). Preview of o3 models.</li> - <li><strong>Product Enhancements:</strong> ChatGPT Pro introduced ($200/month with o1 access). New image generation capabilities in API (Apr 2025).</li> - <li><strong>Developer Tools:</strong> New Responses API and Agents SDK for building AI agents, aiming to simplify agentic AI development (Mar 2025).</li> - <li><strong>Partnerships:</strong> Integration of ChatGPT into Apple Intelligence (announced June 2024).</li> - <li><strong>Corporate:</strong> Major funding rounds (Oct 2024, Apr 2025). [3, 13] Acquired domain Chat.com. Some high-profile departures and new board members (e.g., former NSA head Paul Nakasone).</li> + <li> + <strong>Model Releases:</strong> GPT-4o (May 2024) as new flagship. Sora text-to-video model made + available to ChatGPT Plus/Pro users (Dec 2024). OpenAI o1 reasoning model launched (Dec 2024). + Preview of o3 models. + </li> + <li> + <strong>Product Enhancements:</strong> ChatGPT Pro introduced ($200/month with o1 access). New image + generation capabilities in API (Apr 2025). + </li> + <li> + <strong>Developer Tools:</strong> New Responses API and Agents SDK for building AI agents, aiming to + simplify agentic AI development (Mar 2025). + </li> + <li> + <strong>Partnerships:</strong> Integration of ChatGPT into Apple Intelligence (announced June 2024). + </li> + <li> + <strong>Corporate:</strong> Major funding rounds (Oct 2024, Apr 2025). [3, 13] Acquired domain + Chat.com. Some high-profile departures and new board members (e.g., former NSA head Paul Nakasone). + </li> <li><strong>Safety Framework:</strong> Updated Preparedness Framework (Apr 2025).</li> </ul> </div> @@ -795,14 +958,29 @@ <h5><i class="bi bi-info-circle-fill"></i> Key Information</h5> <div class="card-content-wrapper"> <ul class="key-info-list"> - <li><strong>Founded:</strong> DeepMind Technologies in 2010 (merged with Google Brain in April 2023 to form Google DeepMind). [1]</li> + <li> + <strong>Founded:</strong> DeepMind Technologies in 2010 (merged with Google Brain in April 2023 to + form Google DeepMind). [1] + </li> <li><strong>Founders (DeepMind):</strong> Demis Hassabis, Shane Legg, Mustafa Suleyman. [1]</li> <li><strong>Headquarters:</strong> London, UK (with global research centres). [1]</li> <li><strong>Parent Company:</strong> Alphabet Inc. (Market Cap of Alphabet is relevant). [1]</li> <li><strong>Flagship Models:</strong> Gemini family (1.5 Pro, Ultra, Nano), Gemma. [42]</li> - <li><strong>Main Products/Technologies:</strong> AlphaFold, Imagen, Lyria, RoboCat, contributions to Google products (Search, Cloud AI, Android). [1, 44]</li> - <li><strong>Official Website:</strong> <a href="https://deepmind.google" target="_blank" rel="noopener noreferrer">deepmind.google</a></li> - <li><strong>Documentation:</strong> Primarily via <a href="https://ai.google/research/pubs" target="_blank" rel="noopener noreferrer">ai.google/research/pubs</a> and specific product docs (e.g., Vertex AI).</li> + <li> + <strong>Main Products/Technologies:</strong> AlphaFold, Imagen, Lyria, RoboCat, contributions to + Google products (Search, Cloud AI, Android). [1, 44] + </li> + <li> + <strong>Official Website:</strong> + <a href="https://deepmind.google" target="_blank" rel="noopener noreferrer">deepmind.google</a> + </li> + <li> + <strong>Documentation:</strong> Primarily via + <a href="https://ai.google/research/pubs" target="_blank" rel="noopener noreferrer" + >ai.google/research/pubs</a + > + and specific product docs (e.g., Vertex AI). + </li> </ul> </div> </div> @@ -814,7 +992,8 @@ <h5><i class="bi bi-flag-fill"></i> Origin & Structure</h5> <div class="card-content-wrapper"> <p class="summary"> - DeepMind founded 2010 to "solve intelligence." Acquired by Google 2014. Merged with Google Brain in April 2023 to form Google DeepMind under Alphabet Inc. [1, 44] + DeepMind founded 2010 to "solve intelligence." Acquired by Google 2014. Merged with Google Brain in + April 2023 to form Google DeepMind under Alphabet Inc. [1, 44] </p> <button class="btn btn-sm details-toggle" @@ -831,16 +1010,22 @@ <h6>Key Milestones</h6> <ul> <li> - <strong>DeepMind Technologies (2010):</strong> Founded in London by Demis Hassabis, Shane Legg, and Mustafa Suleyman. Goal: "Solve intelligence" and use it to make the world better. [1] + <strong>DeepMind Technologies (2010):</strong> Founded in London by Demis Hassabis, Shane Legg, and + Mustafa Suleyman. Goal: "Solve intelligence" and use it to make the world better. [1] </li> <li> - <strong>Google Acquisition (2014):</strong> Acquired by Google for a reported $400-$650 million, operating with significant research autonomy. An ethics board was part of the acquisition terms. [1, 38] + <strong>Google Acquisition (2014):</strong> Acquired by Google for a reported $400-$650 million, + operating with significant research autonomy. An ethics board was part of the acquisition terms. [1, + 38] </li> <li> - <strong>Google Brain:</strong> A separate leading AI research team within Google, known for TensorFlow, Transformers, and other breakthroughs. [44] + <strong>Google Brain:</strong> A separate leading AI research team within Google, known for + TensorFlow, Transformers, and other breakthroughs. [44] </li> <li> - <strong>Google DeepMind (April 2023):</strong> Formal merger of DeepMind and the Google Brain team, consolidating Google's AI research under Demis Hassabis's leadership as CEO of Google DeepMind. Part of Alphabet Inc. [1, 44] + <strong>Google DeepMind (April 2023):</strong> Formal merger of DeepMind and the Google Brain team, + consolidating Google's AI research under Demis Hassabis's leadership as CEO of Google DeepMind. Part + of Alphabet Inc. [1, 44] </li> </ul> </div> @@ -852,7 +1037,11 @@ <h5><i class="bi bi-lightbulb-fill"></i> Philosophy & Approach</h5> <div class="card-content-wrapper"> <p class="summary"> - Science-led approach to AGI, emphasizing fundamental research (see their <a href="https://deepmind.google/research/publications/" target="_blank" rel="noopener noreferrer">publications</a>), responsible AI development (guided by Google's AI Principles), and applying AI to global scientific and societal challenges. [44] + Science-led approach to AGI, emphasizing fundamental research (see their + <a href="https://deepmind.google/research/publications/" target="_blank" rel="noopener noreferrer" + >publications</a + >), responsible AI development (guided by Google's AI Principles), and applying AI to global + scientific and societal challenges. [44] </p> <button class="btn btn-sm details-toggle" @@ -869,16 +1058,22 @@ <h6>Core Beliefs & Strategy</h6> <ul> <li> - <strong>Solving Intelligence:</strong> A long-term commitment to understanding and building AGI. [44] + <strong>Solving Intelligence:</strong> A long-term commitment to understanding and building AGI. + [44] </li> <li> - <strong>Science & Research Driven:</strong> Strong emphasis on publishing research, advancing the field through scientific discovery, and tackling grand scientific challenges (e.g., protein folding with AlphaFold, fusion energy, materials science with GNoME). [1, 44] + <strong>Science & Research Driven:</strong> Strong emphasis on publishing research, advancing the + field through scientific discovery, and tackling grand scientific challenges (e.g., protein folding + with AlphaFold, fusion energy, materials science with GNoME). [1, 44] </li> <li> - <strong>Responsible Innovation:</strong> Adherence to Google's AI Principles, with a focus on safety, ethics, fairness, transparency, and societal benefit. Includes a dedicated Responsibility & Safety team. + <strong>Responsible Innovation:</strong> Adherence to Google's AI Principles, with a focus on + safety, ethics, fairness, transparency, and societal benefit. Includes a dedicated Responsibility & + Safety team. </li> - <li> - <strong>Real-world Impact:</strong> Aiming to translate AI breakthroughs into applications that benefit humanity, from scientific tools to enhancing Google products. + <li> + <strong>Real-world Impact:</strong> Aiming to translate AI breakthroughs into applications that + benefit humanity, from scientific tools to enhancing Google products. </li> </ul> </div> @@ -889,9 +1084,7 @@ <div class="card-body"> <h5><i class="bi bi-person-badge"></i> Leadership</h5> <div class="card-content-wrapper"> - <p class="summary"> - Led by co-founder and CEO Demis Hassabis. Lila Ibrahim serves as COO. [1] - </p> + <p class="summary">Led by co-founder and CEO Demis Hassabis. Lila Ibrahim serves as COO. [1]</p> <button class="btn btn-sm details-toggle" type="button" @@ -906,9 +1099,15 @@ <div class="collapse collapse-content" id="collapseDeepMindLeadership"> <h6>Key Figures (as of early 2025)</h6> <ul> - <li><strong>Demis Hassabis:</strong> Co-founder and Chief Executive Officer (CEO) of Google DeepMind. Also co-founder of Isomorphic Labs. Nobel Laureate 2024 in Chemistry for AlphaFold. [1]</li> + <li> + <strong>Demis Hassabis:</strong> Co-founder and Chief Executive Officer (CEO) of Google DeepMind. + Also co-founder of Isomorphic Labs. Nobel Laureate 2024 in Chemistry for AlphaFold. [1] + </li> <li><strong>Lila Ibrahim:</strong> Chief Operating Officer (COO). [1]</li> - <li>Shane Legg and Mustafa Suleyman were co-founders of DeepMind. Suleyman left in 2019 and is now CEO of Microsoft AI. [1]</li> + <li> + Shane Legg and Mustafa Suleyman were co-founders of DeepMind. Suleyman left in 2019 and is now CEO + of Microsoft AI. [1] + </li> </ul> </div> </div> @@ -919,7 +1118,9 @@ <h5><i class="bi bi-boxes"></i> Flagship Models & Technologies</h5> <div class="card-content-wrapper"> <p class="summary"> - Gemini family (1.5 Pro, Ultra, Nano) as the leading multimodal model. [42] Known for AlphaFold, AlphaGo, Imagen (text-to-image), and Lyria (text-to-music). [1, 44] Explore more at <a href="https://labs.google/" target="_blank" rel="noopener noreferrer">Google Labs</a>. + Gemini family (1.5 Pro, Ultra, Nano) as the leading multimodal model. [42] Known for AlphaFold, + AlphaGo, Imagen (text-to-image), and Lyria (text-to-music). [1, 44] Explore more at + <a href="https://labs.google/" target="_blank" rel="noopener noreferrer">Google Labs</a>. </p> <button class="btn btn-sm details-toggle" @@ -935,29 +1136,48 @@ <div class="collapse collapse-content" id="collapseDeepMindModels"> <h6>Current Flagship</h6> <ul> - <li> - <strong>Gemini:</strong> Google DeepMind's most capable and general multimodal model family. [42] - <ul> - <li><code>Gemini 1.5 Pro</code>: State-of-the-art performance with a long context window.</li> - <li><code>Gemini Ultra</code>: Largest and most capable model for highly complex tasks.</li> - <li><code>Gemini Nano</code>: Efficient model for on-device tasks.</li> - <li>Powers features in Google Search, Google Assistant, Google Cloud AI (Vertex AI), Android, and experimental products like Project Astra.</li> - </ul> - </li> - <li><strong>Gemma:</strong> Family of lightweight, state-of-the-art open models built from the same research and technology used to create Gemini models. [42]</li> + <li> + <strong>Gemini:</strong> Google DeepMind's most capable and general multimodal model family. [42] + <ul> + <li><code>Gemini 1.5 Pro</code>: State-of-the-art performance with a long context window.</li> + <li><code>Gemini Ultra</code>: Largest and most capable model for highly complex tasks.</li> + <li><code>Gemini Nano</code>: Efficient model for on-device tasks.</li> + <li> + Powers features in Google Search, Google Assistant, Google Cloud AI (Vertex AI), Android, and + experimental products like Project Astra. + </li> + </ul> + </li> + <li> + <strong>Gemma:</strong> Family of lightweight, state-of-the-art open models built from the same + research and technology used to create Gemini models. [42] + </li> </ul> <h6>Groundbreaking AI Systems</h6> <ul> - <li><strong>AlphaGo / AlphaZero:</strong> Defeated world champion Go player; generalized to master chess and shogi from self-play. [1, 44]</li> - <li><strong>AlphaFold:</strong> Revolutionized biology by accurately predicting protein structures for nearly all known proteins. [1, 44]</li> + <li> + <strong>AlphaGo / AlphaZero:</strong> Defeated world champion Go player; generalized to master chess + and shogi from self-play. [1, 44] + </li> + <li> + <strong>AlphaFold:</strong> Revolutionized biology by accurately predicting protein structures for + nearly all known proteins. [1, 44] + </li> <li><strong>Imagen:</strong> Advanced text-to-image diffusion model.</li> <li><strong>Lyria:</strong> Text-to-music generation model.</li> - <li><strong>GNoME (Graph Networks for Materials Exploration):</strong> Discovered millions of new stable crystalline materials.</li> + <li> + <strong>GNoME (Graph Networks for Materials Exploration):</strong> Discovered millions of new stable + crystalline materials. + </li> <li>Contributions to core technologies like Transformers.</li> </ul> <h6>Integration</h6> <p> - AI research and models are deeply integrated into Google's products (Search, Ads, Cloud, Android, Pixel, Photos, Workspace) and power new experimental AI experiences. Follow their progress on the <a href="https://deepmind.google/blog" target="_blank" rel="noopener noreferrer">Google DeepMind Blog</a>. + AI research and models are deeply integrated into Google's products (Search, Ads, Cloud, Android, + Pixel, Photos, Workspace) and power new experimental AI experiences. Follow their progress on the + <a href="https://deepmind.google/blog" target="_blank" rel="noopener noreferrer" + >Google DeepMind Blog</a + >. </p> </div> </div> @@ -968,7 +1188,8 @@ <h5><i class="bi bi-bullseye"></i> AGI/ASI Goals & Approach</h5> <div class="card-content-wrapper"> <p class="summary"> - AGI is the foundational long-term research goal ("solve intelligence"). [1, 44] Pursued via scientific breakthroughs, responsible development, and scaling general-purpose systems. + AGI is the foundational long-term research goal ("solve intelligence"). [1, 44] Pursued via + scientific breakthroughs, responsible development, and scaling general-purpose systems. </p> <button class="btn btn-sm details-toggle" @@ -985,18 +1206,28 @@ <h6>Approach to Advanced AI</h6> <ul> <li> - <strong>Long-term Aspiration:</strong> The original and ongoing mission is to "solve intelligence," culminating in AGI. [1, 44] Demis Hassabis believes AGI could arrive this decade. + <strong>Long-term Aspiration:</strong> The original and ongoing mission is to "solve intelligence," + culminating in AGI. [1, 44] Demis Hassabis believes AGI could arrive this decade. + </li> + <li> + <strong>Responsible & Safe AGI:</strong> Strong emphasis on developing AGI safely and ethically, + ensuring it is beneficial and controllable. This includes research into alignment, governance, and + societal impact, guided by Google's AI Principles. </li> <li> - <strong>Responsible & Safe AGI:</strong> Strong emphasis on developing AGI safely and ethically, ensuring it is beneficial and controllable. This includes research into alignment, governance, and societal impact, guided by Google's AI Principles. + <strong>Pathways:</strong> Focus on areas like reinforcement learning, neuroscience-inspired AI, + large-scale multimodal modeling, and developing more general and capable systems like Gemini. + Project Astra explores universal AI assistants. </li> <li> - <strong>Pathways:</strong> Focus on areas like reinforcement learning, neuroscience-inspired AI, large-scale multimodal modeling, and developing more general and capable systems like Gemini. Project Astra explores universal AI assistants. + <strong>Scientific Application for Progress:</strong> Belief that tackling complex scientific + problems (like AlphaFold) drives progress towards more general intelligence and demonstrates AI's + potential benefits. [1, 44] </li> <li> - <strong>Scientific Application for Progress:</strong> Belief that tackling complex scientific problems (like AlphaFold) drives progress towards more general intelligence and demonstrates AI's potential benefits. [1, 44] + <strong>Societal Readiness:</strong> Hassabis has expressed concerns that society may not be ready + for AGI and advocates for international cooperation and standards. </li> - <li><strong>Societal Readiness:</strong> Hassabis has expressed concerns that society may not be ready for AGI and advocates for international cooperation and standards.</li> </ul> </div> </div> @@ -1007,7 +1238,8 @@ <h5><i class="bi bi-piggy-bank"></i> Funding & Resources</h5> <div class="card-content-wrapper"> <p class="summary"> - Operates as a subsidiary of Alphabet Inc. (Google), with access to its extensive resources. [1] Original acquisition in 2014. [1, 38] + Operates as a subsidiary of Alphabet Inc. (Google), with access to its extensive resources. [1] + Original acquisition in 2014. [1, 38] </p> <button class="btn btn-sm details-toggle" @@ -1023,10 +1255,25 @@ <div class="collapse collapse-content" id="collapseDeepMindFunding"> <h6>Resource Allocation</h6> <ul> - <li><strong>Subsidiary of Alphabet:</strong> As part of Google (Alphabet Inc.), Google DeepMind has access to vast computational resources, infrastructure, and funding. Specific internal budget allocations are not typically public. [1]</li> - <li><strong>Original Acquisition:</strong> Acquired by Google in 2014 for a sum reported to be between $400 million and $650 million. [1, 38]</li> - <li><strong>Google.org Support:</strong> Google.org has committed funds (e.g., $20 million in Nov 2024) to support external academic and nonprofit organizations using AI for science, often in collaboration with Google DeepMind expertise.</li> - <li><strong>Isomorphic Labs:</strong> A separate Alphabet company, also led by Demis Hassabis and built on AlphaFold's success for drug discovery, raised $600 million in external funding in early 2025, demonstrating investor interest in DeepMind-related ventures.</li> + <li> + <strong>Subsidiary of Alphabet:</strong> As part of Google (Alphabet Inc.), Google DeepMind has + access to vast computational resources, infrastructure, and funding. Specific internal budget + allocations are not typically public. [1] + </li> + <li> + <strong>Original Acquisition:</strong> Acquired by Google in 2014 for a sum reported to be between + $400 million and $650 million. [1, 38] + </li> + <li> + <strong>Google.org Support:</strong> Google.org has committed funds (e.g., $20 million in Nov 2024) + to support external academic and nonprofit organizations using AI for science, often in + collaboration with Google DeepMind expertise. + </li> + <li> + <strong>Isomorphic Labs:</strong> A separate Alphabet company, also led by Demis Hassabis and built + on AlphaFold's success for drug discovery, raised $600 million in external funding in early 2025, + demonstrating investor interest in DeepMind-related ventures. + </li> </ul> </div> </div> @@ -1037,7 +1284,8 @@ <h5><i class="bi bi-newspaper"></i> Recent Developments (2024-2025)</h5> <div class="card-content-wrapper"> <p class="summary"> - Gemini 1.5 Pro advancements, Project Astra reveal (universal AI assistant), Nobel Prize for AlphaFold work. [1, 42, 44] Gemma open models released. [42] Focus on AI for science. + Gemini 1.5 Pro advancements, Project Astra reveal (universal AI assistant), Nobel Prize for + AlphaFold work. [1, 42, 44] Gemma open models released. [42] Focus on AI for science. </p> <button class="btn btn-sm details-toggle" @@ -1053,14 +1301,35 @@ <div class="collapse collapse-content" id="collapseDeepMindDevelopments"> <h6>Key Announcements & Progress</h6> <ul> - <li><strong>Gemini Model Suite:</strong> Continued advancements and rollout of Gemini 1.5 Pro with its large context window and improved capabilities. Integration across Google products. [42] - </li> - <li><strong>Gemma Open Models:</strong> Release of Gemma, a family of lightweight, state-of-the-art open models. [42]</li> - <li><strong>Project Astra:</strong> Showcased progress on a universal AI assistant capable of multimodal understanding and interaction.</li> - <li><strong>Nobel Prize:</strong> Demis Hassabis and John Jumper awarded the 2024 Nobel Prize in Chemistry for their work on AlphaFold. [1]</li> - <li><strong>AI for Science:</strong> Continued breakthroughs in applying AI to scientific discovery, including materials science (GNoME), weather forecasting, and fusion research. Google.org funding for AI in science.</li> - <li><strong>Responsible AI:</strong> Ongoing work on AI safety, ethics, and governance, including contributions to international discussions and standards.</li> - <li><strong>Lyria & Imagen:</strong> Continued development and integration of text-to-music (Lyria) and text-to-image (Imagen 2 & 3) models. [44]</li> + <li> + <strong>Gemini Model Suite:</strong> Continued advancements and rollout of Gemini 1.5 Pro with its + large context window and improved capabilities. Integration across Google products. [42] + </li> + <li> + <strong>Gemma Open Models:</strong> Release of Gemma, a family of lightweight, state-of-the-art open + models. [42] + </li> + <li> + <strong>Project Astra:</strong> Showcased progress on a universal AI assistant capable of multimodal + understanding and interaction. + </li> + <li> + <strong>Nobel Prize:</strong> Demis Hassabis and John Jumper awarded the 2024 Nobel Prize in + Chemistry for their work on AlphaFold. [1] + </li> + <li> + <strong>AI for Science:</strong> Continued breakthroughs in applying AI to scientific discovery, + including materials science (GNoME), weather forecasting, and fusion research. Google.org funding + for AI in science. + </li> + <li> + <strong>Responsible AI:</strong> Ongoing work on AI safety, ethics, and governance, including + contributions to international discussions and standards. + </li> + <li> + <strong>Lyria & Imagen:</strong> Continued development and integration of text-to-music (Lyria) and + text-to-image (Imagen 2 & 3) models. [44] + </li> </ul> </div> </div> @@ -1079,13 +1348,30 @@ <h5><i class="bi bi-info-circle-fill"></i> Key Information</h5> <div class="card-content-wrapper"> <ul class="key-info-list"> - <li><strong>Founded:</strong> 2021, by Dario Amodei, Daniela Amodei, Tom Brown, Chris Olah, Sam McCandlish, Jack Clark, Jared Kaplan.</li> + <li> + <strong>Founded:</strong> 2021, by Dario Amodei, Daniela Amodei, Tom Brown, Chris Olah, Sam + McCandlish, Jack Clark, Jared Kaplan. + </li> <li><strong>Headquarters:</strong> San Francisco, California, USA</li> <li><strong>Valuation:</strong> $61.5 billion (as of March-May 2025). [12, 15, 18, 20, 26]</li> - <li><strong>Flagship Models:</strong> Claude 3 family (Opus, Sonnet, Haiku), Claude 3.5 Sonnet. [11, 26]</li> - <li><strong>Main Products:</strong> Claude.ai (chat interface), Anthropic API, models for enterprise.</li> - <li><strong>Official Website:</strong> <a href="https://www.anthropic.com" target="_blank" rel="noopener noreferrer">anthropic.com</a></li> - <li><strong>Documentation:</strong> <a href="https://docs.anthropic.com" target="_blank" rel="noopener noreferrer">docs.anthropic.com</a> [8, 11, 17, 30]</li> + <li> + <strong>Flagship Models:</strong> Claude 3 family (Opus, Sonnet, Haiku), Claude 3.5 Sonnet. [11, + 26] + </li> + <li> + <strong>Main Products:</strong> Claude.ai (chat interface), Anthropic API, models for enterprise. + </li> + <li> + <strong>Official Website:</strong> + <a href="https://www.anthropic.com" target="_blank" rel="noopener noreferrer">anthropic.com</a> + </li> + <li> + <strong>Documentation:</strong> + <a href="https://docs.anthropic.com" target="_blank" rel="noopener noreferrer" + >docs.anthropic.com</a + > + [8, 11, 17, 30] + </li> </ul> </div> </div> @@ -1097,7 +1383,8 @@ <h5><i class="bi bi-flag-fill"></i> Origin & Founding Vision</h5> <div class="card-content-wrapper"> <p class="summary"> - Founded 2021 by former OpenAI researchers, including Dario and Daniela Amodei. Public Benefit Corporation focused on AI safety. + Founded 2021 by former OpenAI researchers, including Dario and Daniela Amodei. Public Benefit + Corporation focused on AI safety. </p> <button class="btn btn-sm details-toggle" @@ -1114,13 +1401,17 @@ <h6>Key Details</h6> <ul> <li> - <strong>Founding Team:</strong> Led by siblings Dario Amodei (CEO) and Daniela Amodei (President), along with other senior members from OpenAI who shared concerns about AI safety and direction. + <strong>Founding Team:</strong> Led by siblings Dario Amodei (CEO) and Daniela Amodei (President), + along with other senior members from OpenAI who shared concerns about AI safety and direction. </li> <li> - <strong>Motivation:</strong> A desire to conduct AI research with a primary emphasis on safety, interpretability, and developing AI systems that are helpful, honest, and harmless. + <strong>Motivation:</strong> A desire to conduct AI research with a primary emphasis on safety, + interpretability, and developing AI systems that are helpful, honest, and harmless. </li> <li> - <strong>Structure:</strong> Established as a Public Benefit Corporation (PBC) to legally embed its commitment to safety and public benefit alongside its commercial goals. Also has a unique "Long-Term Benefit Trust" structure for governance. + <strong>Structure:</strong> Established as a Public Benefit Corporation (PBC) to legally embed its + commitment to safety and public benefit alongside its commercial goals. Also has a unique "Long-Term + Benefit Trust" structure for governance. </li> </ul> </div> @@ -1132,7 +1423,9 @@ <h5><i class="bi bi-shield-check"></i> Philosophy: Safety First AI</h5> <div class="card-content-wrapper"> <p class="summary"> - Dedicated to building reliable, interpretable, and steerable AI systems. Pioneered "Constitutional AI" and "Responsible Scaling Policy." See their <a href="https://www.anthropic.com/research" target="_blank" rel="noopener noreferrer">research</a>. + Dedicated to building reliable, interpretable, and steerable AI systems. Pioneered "Constitutional + AI" and "Responsible Scaling Policy." See their + <a href="https://www.anthropic.com/research" target="_blank" rel="noopener noreferrer">research</a>. </p> <button class="btn btn-sm details-toggle" @@ -1150,24 +1443,34 @@ <ul> <li><strong>Helpful, Honest, Harmless:</strong> The guiding principles for their AI assistants.</li> <li> - <strong>Constitutional AI:</strong> A technique to train AI models based on a set of principles (a "constitution") to guide behavior, reducing reliance on human labeling for harmful outputs and improving steerability. + <strong>Constitutional AI:</strong> A technique to train AI models based on a set of principles (a + "constitution") to guide behavior, reducing reliance on human labeling for harmful outputs and + improving steerability. + </li> + <li> + <strong>Responsible Scaling Policy (RSP):</strong> A framework outlining safety procedures and + checkpoints to manage risks as AI models become more powerful. + </li> + <li> + <strong>Interpretability Research:</strong> Focus on understanding the internal workings of AI + models to make them more transparent and trustworthy. </li> <li> - <strong>Responsible Scaling Policy (RSP):</strong> A framework outlining safety procedures and checkpoints to manage risks as AI models become more powerful. + <strong>Iterative Deployment:</strong> Cautious deployment of models to learn and improve safety in + real-world scenarios. </li> - <li><strong>Interpretability Research:</strong> Focus on understanding the internal workings of AI models to make them more transparent and trustworthy.</li> - <li><strong>Iterative Deployment:</strong> Cautious deployment of models to learn and improve safety in real-world scenarios.</li> </ul> </div> </div> </div> - <div class="col-lg-4 col-md-6"> + <div class="col-lg-4 col-md-6"> <div class="info-card type-leadership" id="card-anthropic-leadership"> <div class="card-body"> <h5><i class="bi bi-person-badge"></i> Leadership</h5> <div class="card-content-wrapper"> <p class="summary"> - Co-founded by Dario Amodei (CEO) and Daniela Amodei (President). Comprises many ex-OpenAI safety and research leads. + Co-founded by Dario Amodei (CEO) and Daniela Amodei (President). Comprises many ex-OpenAI safety and + research leads. </p> <button class="btn btn-sm details-toggle" @@ -1183,9 +1486,17 @@ <div class="collapse collapse-content" id="collapseAnthropicLeadership"> <h6>Key Figures</h6> <ul> - <li><strong>Dario Amodei:</strong> Co-founder and Chief Executive Officer (CEO). Former VP of Research at OpenAI.</li> - <li><strong>Daniela Amodei:</strong> Co-founder and President. Former VP of Safety and Policy at OpenAI.</li> - <li>Other co-founders include Tom Brown, Chris Olah, Sam McCandlish, Jack Clark, and Jared Kaplan, many of whom were key figures in research and safety at OpenAI.</li> + <li> + <strong>Dario Amodei:</strong> Co-founder and Chief Executive Officer (CEO). Former VP of Research + at OpenAI. + </li> + <li> + <strong>Daniela Amodei:</strong> Co-founder and President. Former VP of Safety and Policy at OpenAI. + </li> + <li> + Other co-founders include Tom Brown, Chris Olah, Sam McCandlish, Jack Clark, and Jared Kaplan, many + of whom were key figures in research and safety at OpenAI. + </li> </ul> </div> </div> @@ -1196,7 +1507,10 @@ <h5><i class="bi bi-boxes"></i> Flagship Models & Products</h5> <div class="card-content-wrapper"> <p class="summary"> - Claude family of models: Claude 3 (Opus, Sonnet, Haiku) and the newer Claude 3.5 Sonnet, known for performance, long context, and safety. [11, 26] Access them via <a href="https://claude.ai" target="_blank" rel="noopener noreferrer">Claude.ai</a> or the <a href="https://console.anthropic.com" target="_blank" rel="noopener noreferrer">API</a>. [8, 30] + Claude family of models: Claude 3 (Opus, Sonnet, Haiku) and the newer Claude 3.5 Sonnet, known for + performance, long context, and safety. [11, 26] Access them via + <a href="https://claude.ai" target="_blank" rel="noopener noreferrer">Claude.ai</a> or the + <a href="https://console.anthropic.com" target="_blank" rel="noopener noreferrer">API</a>. [8, 30] </p> <button class="btn btn-sm details-toggle" @@ -1218,18 +1532,40 @@ <li><code>Opus</code>: Most powerful model for highly complex tasks, top-tier performance.</li> <li><code>Sonnet</code>: Balanced intelligence and speed, ideal for enterprise workloads.</li> <li><code>Haiku</code>: Fastest and most compact model for near-instant responsiveness.</li> - <li>Features: Strong reasoning, improved vision capabilities (multimodal), very long context windows (up to 200K tokens, with some research indicating 1M+).</li> + <li> + Features: Strong reasoning, improved vision capabilities (multimodal), very long context windows + (up to 200K tokens, with some research indicating 1M+). + </li> </ul> </li> - <li> - <strong>Claude 3.5 Sonnet (Released June 2024):</strong> A new model in the 3.5 generation, positioned as faster and more cost-effective than Opus, with strong intelligence and new features like "Artifacts" for interactive content generation. [26] + <li> + <strong>Claude 3.5 Sonnet (Released June 2024):</strong> A new model in the 3.5 generation, + positioned as faster and more cost-effective than Opus, with strong intelligence and new features + like "Artifacts" for interactive content generation. [26] </li> </ul> <h6>Access & Platform</h6> <ul> - <li><strong>API Access:</strong> Models available via Anthropic's API for developers (<a href="https://console.anthropic.com" target="_blank" rel="noopener noreferrer">console.anthropic.com</a>). [8, 30]</li> - <li><strong>Claude.ai:</strong> Web-based chat interface and workspace (<a href="https://claude.ai" target="_blank" rel="noopener noreferrer">claude.ai</a>).</li> - <li><strong>Cloud Partnerships:</strong> Available on major cloud platforms like Amazon Bedrock and Google Cloud Vertex AI. [11]</li> + <li> + <strong>API Access:</strong> Models available via Anthropic's API for developers (<a + href="https://console.anthropic.com" + target="_blank" + rel="noopener noreferrer" + >console.anthropic.com</a + >). [8, 30] + </li> + <li> + <strong>Claude.ai:</strong> Web-based chat interface and workspace (<a + href="https://claude.ai" + target="_blank" + rel="noopener noreferrer" + >claude.ai</a + >). + </li> + <li> + <strong>Cloud Partnerships:</strong> Available on major cloud platforms like Amazon Bedrock and + Google Cloud Vertex AI. [11] + </li> </ul> </div> </div> @@ -1240,7 +1576,8 @@ <h5><i class="bi bi-shield-lock-fill"></i> AGI/ASI Goals & Safety</h5> <div class="card-content-wrapper"> <p class="summary"> - Views AGI development as a serious endeavor requiring proactive safety measures. Goal is beneficial AGI, with safety research integrated at every step. + Views AGI development as a serious endeavor requiring proactive safety measures. Goal is beneficial + AGI, with safety research integrated at every step. </p> <button class="btn btn-sm details-toggle" @@ -1256,21 +1593,37 @@ <div class="collapse collapse-content" id="collapseAnthropicAGI"> <h6>Approach to Advanced AI</h6> <ul> - <li><strong>Safety-Centric AGI:</strong> While aiming to build highly capable AI, the primary differentiator is the deep integration of safety research and principles into the development process from the outset.</li> - <li><strong>Proactive Risk Mitigation:</strong> Emphasizes identifying and mitigating potential risks from advanced AI *before* they become uncontrollable, as outlined in their Responsible Scaling Policy.</li> - <li><strong>Steerable and Interpretable AI:</strong> Research focuses on making models more understandable and controllable, allowing their behavior to be reliably guided by human intentions and principles (e.g., Constitutional AI).</li> - <li><strong>Long-Term Benefit:</strong> The overarching goal is to ensure that if and when AGI is developed, it serves humanity's long-term interests and avoids catastrophic outcomes.</li> + <li> + <strong>Safety-Centric AGI:</strong> While aiming to build highly capable AI, the primary + differentiator is the deep integration of safety research and principles into the development + process from the outset. + </li> + <li> + <strong>Proactive Risk Mitigation:</strong> Emphasizes identifying and mitigating potential risks + from advanced AI *before* they become uncontrollable, as outlined in their Responsible Scaling + Policy. + </li> + <li> + <strong>Steerable and Interpretable AI:</strong> Research focuses on making models more + understandable and controllable, allowing their behavior to be reliably guided by human intentions + and principles (e.g., Constitutional AI). + </li> + <li> + <strong>Long-Term Benefit:</strong> The overarching goal is to ensure that if and when AGI is + developed, it serves humanity's long-term interests and avoids catastrophic outcomes. + </li> </ul> </div> </div> </div> - <div class="col-lg-4 col-md-6"> + <div class="col-lg-4 col-md-6"> <div class="info-card type-funding" id="card-anthropic-funding"> <div class="card-body"> <h5><i class="bi bi-piggy-bank"></i> Funding & Investors</h5> <div class="card-content-wrapper"> <p class="summary"> - Significant backing from major tech companies like Google and Amazon, and venture capital firms, totaling billions (approx. $14.3B in commitments). [12, 15, 18, 20, 26] + Significant backing from major tech companies like Google and Amazon, and venture capital firms, + totaling billions (approx. $14.3B in commitments). [12, 15, 18, 20, 26] </p> <button class="btn btn-sm details-toggle" @@ -1286,11 +1639,23 @@ <div class="collapse collapse-content" id="collapseAnthropicFunding"> <h6>Key Investments</h6> <ul> - <li><strong>Google:</strong> Has invested significantly (e.g., a reported $300M initially, with commitments for up to $2B, with another $550M reported). [20, 26]</li> - <li><strong>Amazon:</strong> Committed up to $4 billion, making AWS its primary cloud provider for mission-critical workloads. [20, 26]</li> + <li> + <strong>Google:</strong> Has invested significantly (e.g., a reported $300M initially, with + commitments for up to $2B, with another $550M reported). [20, 26] + </li> + <li> + <strong>Amazon:</strong> Committed up to $4 billion, making AWS its primary cloud provider for + mission-critical workloads. [20, 26] + </li> <li><strong>Microsoft:</strong> Reported $2B investment commitment. [26]</li> - <li><strong>Other Investors:</strong> Include Spark Capital, Salesforce Ventures, Sound Ventures, Menlo Ventures, SK Telecom, Lightspeed Venture Partners, General Catalyst, Jane Street, Fidelity. [20, 26]</li> - <li><strong>Total Funding:</strong> Has raised approximately $12.4B to $14.3B in cash and commitments across multiple rounds. [20, 26]</li> + <li> + <strong>Other Investors:</strong> Include Spark Capital, Salesforce Ventures, Sound Ventures, Menlo + Ventures, SK Telecom, Lightspeed Venture Partners, General Catalyst, Jane Street, Fidelity. [20, 26] + </li> + <li> + <strong>Total Funding:</strong> Has raised approximately $12.4B to $14.3B in cash and commitments + across multiple rounds. [20, 26] + </li> </ul> </div> </div> @@ -1301,7 +1666,9 @@ <h5><i class="bi bi-newspaper"></i> Recent Developments (2024-2025)</h5> <div class="card-content-wrapper"> <p class="summary"> - Launch of Claude 3 model family (Opus, Sonnet, Haiku) in March 2024. [11, 26] Release of Claude 3.5 Sonnet in June 2024. [26] Expanding enterprise adoption. Check their <a href="https://www.anthropic.com/news" target="_blank" rel="noopener noreferrer">news page</a>. + Launch of Claude 3 model family (Opus, Sonnet, Haiku) in March 2024. [11, 26] Release of Claude 3.5 + Sonnet in June 2024. [26] Expanding enterprise adoption. Check their + <a href="https://www.anthropic.com/news" target="_blank" rel="noopener noreferrer">news page</a>. </p> <button class="btn btn-sm details-toggle" @@ -1317,12 +1684,34 @@ <div class="collapse collapse-content" id="collapseAnthropicDevelopments"> <h6>Key Announcements</h6> <ul> - <li><strong>Claude 3 Model Family (March 2024):</strong> Launch of Opus, Sonnet, and Haiku, setting new industry benchmarks for intelligence, speed, and vision capabilities. [11, 26]</li> - <li><strong>Claude 3.5 Sonnet (June 2024):</strong> Introduced as their first model in the Claude 3.5 generation, offering improved intelligence, speed, and cost-effectiveness, with new features like "Artifacts." [26]</li> - <li><strong>Responsible Scaling Policy (RSP):</strong> Continued commitment and updates to their RSP, detailing safety levels and procedures.</li> - <li><strong>Enterprise Expansion:</strong> Focus on making Claude models accessible and useful for businesses, including partnerships with cloud providers and enterprise software companies.</li> - <li><strong>Research Publications:</strong> Ongoing release of research papers on AI safety, interpretability, and model capabilities. Available at <a href="https://www.anthropic.com/research" target="_blank" rel="noopener noreferrer">anthropic.com/research</a>.</li> - <li><strong>Employee Share Buyback (May 2025):</strong> Announced at a $61.5 billion valuation. [12, 15, 26]</li> + <li> + <strong>Claude 3 Model Family (March 2024):</strong> Launch of Opus, Sonnet, and Haiku, setting new + industry benchmarks for intelligence, speed, and vision capabilities. [11, 26] + </li> + <li> + <strong>Claude 3.5 Sonnet (June 2024):</strong> Introduced as their first model in the Claude 3.5 + generation, offering improved intelligence, speed, and cost-effectiveness, with new features like + "Artifacts." [26] + </li> + <li> + <strong>Responsible Scaling Policy (RSP):</strong> Continued commitment and updates to their RSP, + detailing safety levels and procedures. + </li> + <li> + <strong>Enterprise Expansion:</strong> Focus on making Claude models accessible and useful for + businesses, including partnerships with cloud providers and enterprise software companies. + </li> + <li> + <strong>Research Publications:</strong> Ongoing release of research papers on AI safety, + interpretability, and model capabilities. Available at + <a href="https://www.anthropic.com/research" target="_blank" rel="noopener noreferrer" + >anthropic.com/research</a + >. + </li> + <li> + <strong>Employee Share Buyback (May 2025):</strong> Announced at a $61.5 billion valuation. [12, 15, + 26] + </li> </ul> </div> </div> @@ -1342,13 +1731,35 @@ <div class="card-content-wrapper"> <ul class="key-info-list"> <li><strong>Founded:</strong> Facebook AI Research (FAIR) in 2013.</li> - <li><strong>Key Figures:</strong> Yann LeCun (VP & Chief AI Scientist), Joëlle Pineau (VP of AI Research).</li> + <li> + <strong>Key Figures:</strong> Yann LeCun (VP & Chief AI Scientist), Joëlle Pineau (VP of AI + Research). + </li> <li><strong>Headquarters:</strong> Menlo Park, California, USA (as part of Meta Platforms).</li> - <li><strong>Parent Company:</strong> Meta Platforms, Inc. (Market Cap of META relevant, ~$1.5T as of early 2025). [45]</li> - <li><strong>Flagship Models:</strong> Llama family (Llama 3), Segment Anything Model (SAM), Seamless Communication models. [47]</li> - <li><strong>Main Products/Platforms:</strong> Meta AI assistant, PyTorch, various open-source models and tools.</li> - <li><strong>Official Website:</strong> <a href="https://ai.meta.com/" target="_blank" rel="noopener noreferrer">ai.meta.com</a></li> - <li><strong>Documentation:</strong> Via <a href="https://ai.meta.com/research/" target="_blank" rel="noopener noreferrer">ai.meta.com/research/</a> and model-specific sites like <a href="https://llama.meta.com/" target="_blank" rel="noopener noreferrer">llama.meta.com</a>.</li> + <li> + <strong>Parent Company:</strong> Meta Platforms, Inc. (Market Cap of META relevant, ~$1.5T as of + early 2025). [45] + </li> + <li> + <strong>Flagship Models:</strong> Llama family (Llama 3), Segment Anything Model (SAM), Seamless + Communication models. [47] + </li> + <li> + <strong>Main Products/Platforms:</strong> Meta AI assistant, PyTorch, various open-source models + and tools. + </li> + <li> + <strong>Official Website:</strong> + <a href="https://ai.meta.com/" target="_blank" rel="noopener noreferrer">ai.meta.com</a> + </li> + <li> + <strong>Documentation:</strong> Via + <a href="https://ai.meta.com/research/" target="_blank" rel="noopener noreferrer" + >ai.meta.com/research/</a + > + and model-specific sites like + <a href="https://llama.meta.com/" target="_blank" rel="noopener noreferrer">llama.meta.com</a>. + </li> </ul> </div> </div> @@ -1360,7 +1771,8 @@ <h5><i class="bi bi-flag-fill"></i> Origin & Structure</h5> <div class="card-content-wrapper"> <p class="summary"> - Rooted in Facebook AI Research (FAIR), founded in 2013 by Yann LeCun. Now Meta AI, a division of Meta Platforms, driving open research and AI for Meta's products. + Rooted in Facebook AI Research (FAIR), founded in 2013 by Yann LeCun. Now Meta AI, a division of + Meta Platforms, driving open research and AI for Meta's products. </p> <button class="btn btn-sm details-toggle" @@ -1376,9 +1788,18 @@ <div class="collapse collapse-content" id="collapseMetaOrigin"> <h6>Key Milestones</h6> <ul> - <li><strong>FAIR (Facebook AI Research, 2013):</strong> Established by Yann LeCun to advance AI through open research, publishing papers, and releasing code and datasets.</li> - <li><strong>Meta AI:</strong> As Facebook rebranded to Meta, FAIR became a core part of Meta AI, continuing its research mission while also focusing on integrating AI into Meta's family of apps (Facebook, Instagram, WhatsApp, Messenger) and future platforms like AR/VR.</li> - <li><strong>Decentralized Labs:</strong> Operates with research labs globally, fostering collaboration.</li> + <li> + <strong>FAIR (Facebook AI Research, 2013):</strong> Established by Yann LeCun to advance AI through + open research, publishing papers, and releasing code and datasets. + </li> + <li> + <strong>Meta AI:</strong> As Facebook rebranded to Meta, FAIR became a core part of Meta AI, + continuing its research mission while also focusing on integrating AI into Meta's family of apps + (Facebook, Instagram, WhatsApp, Messenger) and future platforms like AR/VR. + </li> + <li> + <strong>Decentralized Labs:</strong> Operates with research labs globally, fostering collaboration. + </li> </ul> </div> </div> @@ -1389,7 +1810,12 @@ <h5><i class="bi bi-unlock-fill"></i> Philosophy & Open Source</h5> <div class="card-content-wrapper"> <p class="summary"> - Strong commitment to open science and open source AI. Believes openness accelerates innovation, safety, and democratization of AI. Key proponent of releasing powerful models like Llama. [47] Explore their work on <a href="https://ai.meta.com/research/" target="_blank" rel="noopener noreferrer">their research page</a>. + Strong commitment to open science and open source AI. Believes openness accelerates innovation, + safety, and democratization of AI. Key proponent of releasing powerful models like Llama. [47] + Explore their work on + <a href="https://ai.meta.com/research/" target="_blank" rel="noopener noreferrer" + >their research page</a + >. </p> <button class="btn btn-sm details-toggle" @@ -1405,10 +1831,23 @@ <div class="collapse collapse-content" id="collapseMetaPhilosophy"> <h6>Core Beliefs</h6> <ul> - <li><strong>Open Research and Development:</strong> A foundational principle. Meta AI consistently publishes research and open-sources models, tools (e.g., <a href="https://pytorch.org/" target="_blank" rel="noopener noreferrer">PyTorch</a>), and datasets.</li> - <li><strong>Democratizing AI:</strong> Aims to provide broad access to state-of-the-art AI to foster a wider community of researchers and developers.</li> - <li><strong>Innovation through Collaboration:</strong> Believes that community involvement in using, scrutinizing, and improving open models leads to faster progress and safer AI.</li> - <li><strong>Responsible AI Development:</strong> Alongside openness, Meta AI emphasizes responsible AI practices, including research into fairness, privacy, and robustness.</li> + <li> + <strong>Open Research and Development:</strong> A foundational principle. Meta AI consistently + publishes research and open-sources models, tools (e.g., + <a href="https://pytorch.org/" target="_blank" rel="noopener noreferrer">PyTorch</a>), and datasets. + </li> + <li> + <strong>Democratizing AI:</strong> Aims to provide broad access to state-of-the-art AI to foster a + wider community of researchers and developers. + </li> + <li> + <strong>Innovation through Collaboration:</strong> Believes that community involvement in using, + scrutinizing, and improving open models leads to faster progress and safer AI. + </li> + <li> + <strong>Responsible AI Development:</strong> Alongside openness, Meta AI emphasizes responsible AI + practices, including research into fairness, privacy, and robustness. + </li> </ul> </div> </div> @@ -1419,7 +1858,8 @@ <h5><i class="bi bi-person-badge"></i> Leadership</h5> <div class="card-content-wrapper"> <p class="summary"> - Yann LeCun (VP & Chief AI Scientist) is a guiding figure. Joëlle Pineau (VP of AI Research) also plays a key role. AI efforts are integrated across Meta. + Yann LeCun (VP & Chief AI Scientist) is a guiding figure. Joëlle Pineau (VP of AI Research) also + plays a key role. AI efforts are integrated across Meta. </p> <button class="btn btn-sm details-toggle" @@ -1435,9 +1875,18 @@ <div class="collapse collapse-content" id="collapseMetaLeadership"> <h6>Key Figures</h6> <ul> - <li><strong>Yann LeCun:</strong> VP & Chief AI Scientist at Meta. Turing Award laureate, a pioneer in deep learning. Strong advocate for open AI and specific AGI architectures.</li> - <li><strong>Joëlle Pineau:</strong> VP of AI Research. Focuses on areas including reinforcement learning and responsible AI.</li> - <li>AI research and development is broadly distributed across Meta, with many influential researchers and engineers contributing.</li> + <li> + <strong>Yann LeCun:</strong> VP & Chief AI Scientist at Meta. Turing Award laureate, a pioneer in + deep learning. Strong advocate for open AI and specific AGI architectures. + </li> + <li> + <strong>Joëlle Pineau:</strong> VP of AI Research. Focuses on areas including reinforcement learning + and responsible AI. + </li> + <li> + AI research and development is broadly distributed across Meta, with many influential researchers + and engineers contributing. + </li> </ul> </div> </div> @@ -1448,7 +1897,10 @@ <h5><i class="bi bi-boxes"></i> Flagship Models & Technologies</h5> <div class="card-content-wrapper"> <p class="summary"> - Llama family (Llama 2, <a href="https://llama.meta.com/" target="_blank" rel="noopener noreferrer">Llama 3</a>) of open-weight LLMs. [47] Also known for Segment Anything Model (SAM), Seamless Communication models, and <a href="https://pytorch.org/" target="_blank" rel="noopener noreferrer">PyTorch</a>. + Llama family (Llama 2, + <a href="https://llama.meta.com/" target="_blank" rel="noopener noreferrer">Llama 3</a>) of + open-weight LLMs. [47] Also known for Segment Anything Model (SAM), Seamless Communication models, + and <a href="https://pytorch.org/" target="_blank" rel="noopener noreferrer">PyTorch</a>. </p> <button class="btn btn-sm details-toggle" @@ -1465,18 +1917,41 @@ <h6>Key Open Models & Tools</h6> <ul> <li> - <strong>Llama Series (e.g., <a href="https://llama.meta.com/" target="_blank" rel="noopener noreferrer">Llama 3</a>):</strong> Family of large language models released with open weights (or "openly available" weights for community access and research), in various sizes (e.g., 8B, 70B parameters). Llama 3 (released April 2024) showed significant improvements. [47] + <strong + >Llama Series (e.g., + <a href="https://llama.meta.com/" target="_blank" rel="noopener noreferrer">Llama 3</a>):</strong + > + Family of large language models released with open weights (or "openly available" weights for + community access and research), in various sizes (e.g., 8B, 70B parameters). Llama 3 (released April + 2024) showed significant improvements. [47] </li> <li> - <strong>Segment Anything Model (SAM):</strong> Foundation model for image segmentation, capable of identifying objects in images and videos with high granularity. + <strong>Segment Anything Model (SAM):</strong> Foundation model for image segmentation, capable of + identifying objects in images and videos with high granularity. + </li> + <li> + <strong>Seamless Communication Models (e.g., SeamlessM4T, SeamlessExpressive):</strong> Multilingual + and multitask models for speech translation, transcription, and expressive cross-lingual + communication. + </li> + <li> + <strong + ><a href="https://pytorch.org/" target="_blank" rel="noopener noreferrer">PyTorch</a>:</strong + > + Leading open-source machine learning framework, widely adopted in research and industry, originally + developed by FAIR. + </li> + <li> + Other models include Code Llama (for code generation), AudioCraft (for audio generation), and + various computer vision models. </li> - <li><strong>Seamless Communication Models (e.g., SeamlessM4T, SeamlessExpressive):</strong> Multilingual and multitask models for speech translation, transcription, and expressive cross-lingual communication.</li> - <li><strong><a href="https://pytorch.org/" target="_blank" rel="noopener noreferrer">PyTorch</a>:</strong> Leading open-source machine learning framework, widely adopted in research and industry, originally developed by FAIR.</li> - <li>Other models include Code Llama (for code generation), AudioCraft (for audio generation), and various computer vision models.</li> </ul> <h6>Integration</h6> <p> - AI powers features across Meta's platforms (Meta AI assistant in Facebook, Instagram, WhatsApp, Messenger, Ray-Ban Meta smart glasses) and underpins research for future AR/VR experiences. Keep up with news on their <a href="https://ai.meta.com/blog/" target="_blank" rel="noopener noreferrer">blog</a>. + AI powers features across Meta's platforms (Meta AI assistant in Facebook, Instagram, WhatsApp, + Messenger, Ray-Ban Meta smart glasses) and underpins research for future AR/VR experiences. Keep up + with news on their + <a href="https://ai.meta.com/blog/" target="_blank" rel="noopener noreferrer">blog</a>. </p> </div> </div> @@ -1487,7 +1962,8 @@ <h5><i class="bi bi-bullseye"></i> AGI/ASI Goals & Approach</h5> <div class="card-content-wrapper"> <p class="summary"> - AGI is a long-term ambition. Focus on building "human-level intelligence" through understanding the world, reasoning, and planning. Emphasis on world models and architectures like JEPA. + AGI is a long-term ambition. Focus on building "human-level intelligence" through understanding the + world, reasoning, and planning. Emphasis on world models and architectures like JEPA. </p> <button class="btn btn-sm details-toggle" @@ -1504,15 +1980,23 @@ <h6>Approach to Advanced AI</h6> <ul> <li> - <strong>Human-Level Intelligence:</strong> The stated goal is to achieve AI with capabilities comparable to humans in learning, reasoning, and interacting with the world. + <strong>Human-Level Intelligence:</strong> The stated goal is to achieve AI with capabilities + comparable to humans in learning, reasoning, and interacting with the world. </li> <li> - <strong>Yann LeCun's Vision for AGI:</strong> LeCun advocates for architectures beyond current auto-regressive LLMs. He proposes systems that can learn world models, predict, reason, and plan. This includes concepts like Joint Embedding Predictive Architectures (JEPA) and a more modular, hierarchical system. + <strong>Yann LeCun's Vision for AGI:</strong> LeCun advocates for architectures beyond current + auto-regressive LLMs. He proposes systems that can learn world models, predict, reason, and plan. + This includes concepts like Joint Embedding Predictive Architectures (JEPA) and a more modular, + hierarchical system. </li> <li> - <strong>Openness as a Path to Safe AGI:</strong> Believes that open development and community scrutiny are crucial for developing AGI that is safe, understood, and beneficial. [47] + <strong>Openness as a Path to Safe AGI:</strong> Believes that open development and community + scrutiny are crucial for developing AGI that is safe, understood, and beneficial. [47] + </li> + <li> + <strong>Focus on Embodied AI and Robotics:</strong> Research into AI that can interact with and + learn from the physical world, seen as important for developing more grounded intelligence. </li> - <li><strong>Focus on Embodied AI and Robotics:</strong> Research into AI that can interact with and learn from the physical world, seen as important for developing more grounded intelligence.</li> </ul> </div> </div> @@ -1523,7 +2007,8 @@ <h5><i class="bi bi-piggy-bank"></i> Funding & Resources</h5> <div class="card-content-wrapper"> <p class="summary"> - As a division of Meta Platforms, Meta AI is funded through Meta's overall R&D budget. Significant investment in compute (tens of thousands of GPUs). + As a division of Meta Platforms, Meta AI is funded through Meta's overall R&D budget. Significant + investment in compute (tens of thousands of GPUs). </p> <button class="btn btn-sm details-toggle" @@ -1539,20 +2024,31 @@ <div class="collapse collapse-content" id="collapseMetaFunding"> <h6>Resource Allocation</h6> <ul> - <li><strong>Internal Funding:</strong> Meta AI's operations are funded as part of Meta Platforms' substantial R&D investments. Meta Platforms Inc. has a market cap around $1.5 Trillion as of early 2025. [45]</li> - <li><strong>Compute Power:</strong> Meta has been investing heavily in AI supercomputers and GPU clusters (e.g., aiming for hundreds of thousands of H100 GPUs) to train increasingly large and complex models.</li> - <li><strong>Talent Acquisition:</strong> Actively recruits top AI researchers and engineers globally.</li> + <li> + <strong>Internal Funding:</strong> Meta AI's operations are funded as part of Meta Platforms' + substantial R&D investments. Meta Platforms Inc. has a market cap around $1.5 Trillion as of early + 2025. [45] + </li> + <li> + <strong>Compute Power:</strong> Meta has been investing heavily in AI supercomputers and GPU + clusters (e.g., aiming for hundreds of thousands of H100 GPUs) to train increasingly large and + complex models. + </li> + <li> + <strong>Talent Acquisition:</strong> Actively recruits top AI researchers and engineers globally. + </li> </ul> </div> </div> </div> - <div class="col-lg-4 col-md-6"> + <div class="col-lg-4 col-md-6"> <div class="info-card type-developments" id="card-meta-developments"> <div class="card-body"> <h5><i class="bi bi-newspaper"></i> Recent Developments (2024-2025)</h5> <div class="card-content-wrapper"> <p class="summary"> - Release of Llama 3 models. [47] Meta AI assistant integrated widely. Advancements in multimodal AI (Seamless Communication) and vision (SAM). Ongoing push for open models. + Release of Llama 3 models. [47] Meta AI assistant integrated widely. Advancements in multimodal AI + (Seamless Communication) and vision (SAM). Ongoing push for open models. </p> <button class="btn btn-sm details-toggle" @@ -1568,16 +2064,35 @@ <div class="collapse collapse-content" id="collapseMetaDevelopments"> <h6>Key Announcements</h6> <ul> - <li><strong>Llama 3 Release (April 2024):</strong> Launch of significantly improved open-weight models (8B and 70B parameters), with larger models (e.g., 400B+ parameters) in training. [47]</li> - <li><strong>Meta AI Assistant Rollout:</strong> Wider integration and enhanced capabilities of the Meta AI assistant across Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta smart glasses, powered by Llama 3.</li> - <li><strong>Multimodal AI:</strong> Continued advancements with models like SeamlessExpressive for more natural cross-lingual voice communication, and ongoing research in combining vision, language, and audio.</li> - <li><strong>Open Source Contributions:</strong> Regular releases of new models, datasets, and research papers, reinforcing commitment to open science. Check their <a href="https://ai.meta.com/blog/" target="_blank" rel="noopener noreferrer">blog</a> and <a href="https://ai.meta.com/research/" target="_blank" rel="noopener noreferrer">research page</a>.</li> - <li><strong>Focus on Next-Gen Architectures:</strong> Continued advocacy and research by Yann LeCun and FAIR into alternative AI architectures for more robust reasoning and world modeling.</li> + <li> + <strong>Llama 3 Release (April 2024):</strong> Launch of significantly improved open-weight models + (8B and 70B parameters), with larger models (e.g., 400B+ parameters) in training. [47] + </li> + <li> + <strong>Meta AI Assistant Rollout:</strong> Wider integration and enhanced capabilities of the Meta + AI assistant across Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta smart glasses, + powered by Llama 3. + </li> + <li> + <strong>Multimodal AI:</strong> Continued advancements with models like SeamlessExpressive for more + natural cross-lingual voice communication, and ongoing research in combining vision, language, and + audio. + </li> + <li> + <strong>Open Source Contributions:</strong> Regular releases of new models, datasets, and research + papers, reinforcing commitment to open science. Check their + <a href="https://ai.meta.com/blog/" target="_blank" rel="noopener noreferrer">blog</a> and + <a href="https://ai.meta.com/research/" target="_blank" rel="noopener noreferrer">research page</a>. + </li> + <li> + <strong>Focus on Next-Gen Architectures:</strong> Continued advocacy and research by Yann LeCun and + FAIR into alternative AI architectures for more robust reasoning and world modeling. + </li> </ul> </div> </div> </div> - <!-- Removed Key Links card for Meta AI --> + <!-- Removed Key Links card for Meta AI --> </div> </div> @@ -1585,7 +2100,7 @@ <div class="schema-container cat-cohere" data-section-id="section-cohere"> <h2 class="section-title" id="title-cohere">Cohere</h2> <div class="row"> - <div class="col-lg-4 col-md-6"> + <div class="col-lg-4 col-md-6"> <div class="info-card type-info" id="card-cohere-keyinfo"> <div class="card-body"> <h5><i class="bi bi-info-circle-fill"></i> Key Information</h5> @@ -1593,11 +2108,24 @@ <ul class="key-info-list"> <li><strong>Founded:</strong> 2019, by Aidan Gomez, Nick Frosst, Ivan Zhang.</li> <li><strong>Headquarters:</strong> Toronto, Canada.</li> - <li><strong>Valuation:</strong> ~$2.2 billion (as of June 2023), with reports of aiming for $5 billion in a new round (early 2024).</li> + <li> + <strong>Valuation:</strong> ~$2.2 billion (as of June 2023), with reports of aiming for $5 billion + in a new round (early 2024). + </li> <li><strong>Flagship Models:</strong> Command family (Command R, Command R+), Rerank, Embed.</li> - <li><strong>Main Products:</strong> Cohere Platform (API access), models for enterprise search, RAG, generation.</li> - <li><strong>Official Website:</strong> <a href="https://cohere.com/" target="_blank" rel="noopener noreferrer">cohere.com</a></li> - <li><strong>Documentation:</strong> <a href="https://docs.cohere.com/" target="_blank" rel="noopener noreferrer">docs.cohere.com</a> [9, 29, 32, 35, 41]</li> + <li> + <strong>Main Products:</strong> Cohere Platform (API access), models for enterprise search, RAG, + generation. + </li> + <li> + <strong>Official Website:</strong> + <a href="https://cohere.com/" target="_blank" rel="noopener noreferrer">cohere.com</a> + </li> + <li> + <strong>Documentation:</strong> + <a href="https://docs.cohere.com/" target="_blank" rel="noopener noreferrer">docs.cohere.com</a> + [9, 29, 32, 35, 41] + </li> </ul> </div> </div> @@ -1609,7 +2137,8 @@ <h5><i class="bi bi-flag-fill"></i> Origin & Focus</h5> <div class="card-content-wrapper"> <p class="summary"> - Founded in 2019 by ex-Google Brain researchers. Focuses on providing LLMs and NLP tools for enterprise applications. + Founded in 2019 by ex-Google Brain researchers. Focuses on providing LLMs and NLP tools for + enterprise applications. </p> <button class="btn btn-sm details-toggle" @@ -1625,8 +2154,14 @@ <div class="collapse collapse-content" id="collapseCohereOrigin"> <h6>Key Details</h6> <ul> - <li><strong>Founders:</strong> Aidan Gomez, Nick Frosst (both previously at Google Brain, Gomez co-authored "Attention Is All You Need" paper), and Ivan Zhang.</li> - <li><strong>Mission:</strong> To empower enterprises with cutting-edge large language models and NLP capabilities, focusing on practical business use cases.</li> + <li> + <strong>Founders:</strong> Aidan Gomez, Nick Frosst (both previously at Google Brain, Gomez + co-authored "Attention Is All You Need" paper), and Ivan Zhang. + </li> + <li> + <strong>Mission:</strong> To empower enterprises with cutting-edge large language models and NLP + capabilities, focusing on practical business use cases. + </li> <li><strong>Headquarters:</strong> Toronto, Canada, with presence in London and Palo Alto.</li> </ul> </div> @@ -1638,7 +2173,9 @@ <h5><i class="bi bi-building"></i> Philosophy & Enterprise Focus</h5> <div class="card-content-wrapper"> <p class="summary"> - Aims to make advanced LLMs accessible, secure, and customizable for businesses. Emphasizes data privacy, multi-cloud deployment, and practical RAG solutions. Explore their thoughts on their <a href="https://txt.cohere.com/" target="_blank" rel="noopener noreferrer">blog</a>. + Aims to make advanced LLMs accessible, secure, and customizable for businesses. Emphasizes data + privacy, multi-cloud deployment, and practical RAG solutions. Explore their thoughts on their + <a href="https://txt.cohere.com/" target="_blank" rel="noopener noreferrer">blog</a>. </p> <button class="btn btn-sm details-toggle" @@ -1655,23 +2192,32 @@ <h6>Core Strategy</h6> <ul> <li> - <strong>Enterprise-Grade AI:</strong> Focused on providing LLMs (Command series), embedding models, and reranking models tailored for business needs like search, summarization, generation, and dialogue. + <strong>Enterprise-Grade AI:</strong> Focused on providing LLMs (Command series), embedding models, + and reranking models tailored for business needs like search, summarization, generation, and + dialogue. </li> <li> - <strong>Data Privacy & Security:</strong> Offers flexible deployment options, including private cloud (AWS, Google Cloud, Oracle, Azure), VPC, and on-premise, to ensure enterprises can use models securely with their own data. + <strong>Data Privacy & Security:</strong> Offers flexible deployment options, including private + cloud (AWS, Google Cloud, Oracle, Azure), VPC, and on-premise, to ensure enterprises can use models + securely with their own data. </li> <li> - <strong>Model Customization & Fine-Tuning:</strong> Enables businesses to adapt models for specific industry jargon, tasks, and company knowledge. + <strong>Model Customization & Fine-Tuning:</strong> Enables businesses to adapt models for specific + industry jargon, tasks, and company knowledge. </li> <li> - <strong>Retrieval Augmented Generation (RAG):</strong> Strong focus on RAG to ground model responses in enterprise data, improving accuracy and reducing hallucinations. + <strong>Retrieval Augmented Generation (RAG):</strong> Strong focus on RAG to ground model responses + in enterprise data, improving accuracy and reducing hallucinations. + </li> + <li> + <strong>Multi-Cloud & Interoperability:</strong> Aims for model accessibility and ease of + integration across various cloud platforms and existing enterprise systems. </li> - <li><strong>Multi-Cloud & Interoperability:</strong> Aims for model accessibility and ease of integration across various cloud platforms and existing enterprise systems.</li> </ul> </div> </div> </div> - <div class="col-lg-4 col-md-6"> + <div class="col-lg-4 col-md-6"> <div class="info-card type-leadership" id="card-cohere-leadership"> <div class="card-body"> <h5><i class="bi bi-person-badge"></i> Leadership</h5> @@ -1693,7 +2239,10 @@ <div class="collapse collapse-content" id="collapseCohereLeadership"> <h6>Key Figures</h6> <ul> - <li><strong>Aidan Gomez:</strong> Co-founder and Chief Executive Officer (CEO). Co-author of the influential "Attention Is All You Need" paper.</li> + <li> + <strong>Aidan Gomez:</strong> Co-founder and Chief Executive Officer (CEO). Co-author of the + influential "Attention Is All You Need" paper. + </li> <li><strong>Nick Frosst:</strong> Co-founder. Previously at Google Brain.</li> <li><strong>Ivan Zhang:</strong> Co-founder.</li> <li>Martin Kon joined as President & COO in 2023.</li> @@ -1707,7 +2256,10 @@ <h5><i class="bi bi-boxes"></i> Flagship Models & Products</h5> <div class="card-content-wrapper"> <p class="summary"> - Command model family (Command R, Command R+) for generation, Rerank for semantic search, Embed for text embeddings. Platform focuses on practical enterprise applications via their <a href="https://dashboard.cohere.com/" target="_blank" rel="noopener noreferrer">platform</a>. [9, 35] + Command model family (Command R, Command R+) for generation, Rerank for semantic search, Embed for + text embeddings. Platform focuses on practical enterprise applications via their + <a href="https://dashboard.cohere.com/" target="_blank" rel="noopener noreferrer">platform</a>. [9, + 35] </p> <button class="btn btn-sm details-toggle" @@ -1726,17 +2278,32 @@ <li> <strong>Command Model Family:</strong> <ul> - <li><code>Command R</code> & <code>Command R+</code>: High-performance models optimized for enterprise-grade workloads, RAG, and tool use with long context.</li> - <li>Older models like Command, Command Light also exist.</li> + <li> + <code>Command R</code> & <code>Command R+</code>: High-performance models optimized for + enterprise-grade workloads, RAG, and tool use with long context. + </li> + <li>Older models like Command, Command Light also exist.</li> </ul> </li> <li> - <strong>Rerank:</strong> Improves semantic search quality by re-ranking search results from existing enterprise search systems or vector databases, focusing on relevance. + <strong>Rerank:</strong> Improves semantic search quality by re-ranking search results from existing + enterprise search systems or vector databases, focusing on relevance. </li> <li> - <strong>Embed:</strong> Generates state-of-the-art text embeddings for tasks like semantic search, clustering, and classification, available in multiple languages. + <strong>Embed:</strong> Generates state-of-the-art text embeddings for tasks like semantic search, + clustering, and classification, available in multiple languages. + </li> + <li> + <strong>Cohere Platform:</strong> Provides API access (<a + href="https://dashboard.cohere.com/" + target="_blank" + rel="noopener noreferrer" + >dashboard.cohere.com</a + >, + <a href="https://docs.cohere.com/" target="_blank" rel="noopener noreferrer">docs.cohere.com</a> [9, + 29, 32, 35, 41]), tools for fine-tuning, and integrations to deploy models in various enterprise + environments. </li> - <li><strong>Cohere Platform:</strong> Provides API access (<a href="https://dashboard.cohere.com/" target="_blank" rel="noopener noreferrer">dashboard.cohere.com</a>, <a href="https://docs.cohere.com/" target="_blank" rel="noopener noreferrer">docs.cohere.com</a> [9, 29, 32, 35, 41]), tools for fine-tuning, and integrations to deploy models in various enterprise environments.</li> </ul> </div> </div> @@ -1747,7 +2314,8 @@ <h5><i class="bi bi-people-fill"></i> Target Audience & Use Cases</h5> <div class="card-content-wrapper"> <p class="summary"> - Enterprises, developers, and data-sensitive industries. Focus on advanced search, RAG, content generation, summarization, and chatbots. + Enterprises, developers, and data-sensitive industries. Focus on advanced search, RAG, content + generation, summarization, and chatbots. </p> <button class="btn btn-sm details-toggle" @@ -1764,12 +2332,16 @@ <h6>Primary Users</h6> <ul> <li> - <strong>Enterprises:</strong> Businesses of all sizes seeking to integrate advanced NLP/LLM capabilities into their products, workflows, and internal systems. + <strong>Enterprises:</strong> Businesses of all sizes seeking to integrate advanced NLP/LLM + capabilities into their products, workflows, and internal systems. </li> - <li><strong>Developers:</strong> Building applications that leverage powerful and customizable language models securely. + <li> + <strong>Developers:</strong> Building applications that leverage powerful and customizable language + models securely. </li> <li> - <strong>Industries:</strong> Finance, healthcare, retail, technology, legal, and other sectors needing secure, reliable, and customizable AI solutions. + <strong>Industries:</strong> Finance, healthcare, retail, technology, legal, and other sectors + needing secure, reliable, and customizable AI solutions. </li> </ul> <h6>Common Applications</h6> @@ -1782,13 +2354,14 @@ </div> </div> </div> - <div class="col-lg-4 col-md-6"> + <div class="col-lg-4 col-md-6"> <div class="info-card type-funding" id="card-cohere-funding"> <div class="card-body"> <h5><i class="bi bi-piggy-bank"></i> Funding & Investors</h5> <div class="card-content-wrapper"> <p class="summary"> - Raised significant capital from investors including Tiger Global, Index Ventures, Nvidia, Oracle, Salesforce Ventures. Series C in 2023 valued at over $2B. + Raised significant capital from investors including Tiger Global, Index Ventures, Nvidia, Oracle, + Salesforce Ventures. Series C in 2023 valued at over $2B. </p> <button class="btn btn-sm details-toggle" @@ -1804,9 +2377,19 @@ <div class="collapse collapse-content" id="collapseCohereFunding"> <h6>Key Investments</h6> <ul> - <li><strong>Series C (June 2023):</strong> Raised $270 million, led by Inovia Capital, with participation from Nvidia, Oracle, Salesforce Ventures, Index Ventures, Tiger Global, and others, reportedly valuing the company at $2.1-$2.2 billion.</li> - <li><strong>Previous Rounds:</strong> Earlier funding from Index Ventures, Tiger Global, Radical Ventures, Section 32, and prominent AI figures.</li> - <li><strong>Strategic Partnerships:</strong> Investments from companies like Nvidia, Oracle, and Salesforce also reflect strategic alliances for compute resources and market access.</li> + <li> + <strong>Series C (June 2023):</strong> Raised $270 million, led by Inovia Capital, with + participation from Nvidia, Oracle, Salesforce Ventures, Index Ventures, Tiger Global, and others, + reportedly valuing the company at $2.1-$2.2 billion. + </li> + <li> + <strong>Previous Rounds:</strong> Earlier funding from Index Ventures, Tiger Global, Radical + Ventures, Section 32, and prominent AI figures. + </li> + <li> + <strong>Strategic Partnerships:</strong> Investments from companies like Nvidia, Oracle, and + Salesforce also reflect strategic alliances for compute resources and market access. + </li> </ul> </div> </div> @@ -1817,7 +2400,8 @@ <h5><i class="bi bi-newspaper"></i> Recent Developments (2024-2025)</h5> <div class="card-content-wrapper"> <p class="summary"> - Launch of Command R and R+ models. Expanding cloud partnerships (e.g., Oracle, Microsoft Azure). Focus on RAG and enterprise tooling. + Launch of Command R and R+ models. Expanding cloud partnerships (e.g., Oracle, Microsoft Azure). + Focus on RAG and enterprise tooling. </p> <button class="btn btn-sm details-toggle" @@ -1833,16 +2417,33 @@ <div class="collapse collapse-content" id="collapseCohereDevelopments"> <h6>Key Announcements</h6> <ul> - <li><strong>Command R & R+ Launch (Early 2024):</strong> Release of highly capable models optimized for enterprise RAG and tool use, with competitive pricing and performance.</li> - <li><strong>Cloud Expansion:</strong> Broadened availability on major cloud platforms, including new integrations with Oracle Cloud Infrastructure (OCI) and Microsoft Azure.</li> - <li><strong>Enterprise Tooling:</strong> Enhanced platform features for data management, model fine-tuning, and deploying RAG applications. Learn more through their <a href="https://docs.cohere.com/" target="_blank" rel="noopener noreferrer">documentation</a>. [9, 29, 32, 35, 41]</li> - <li><strong>Cohere Coral (Tech Preview):</strong> A knowledge assistant for enterprises, leveraging RAG to connect to business data.</li> - <li><strong>Aya Model (Collaboration):</strong> Contributed to the release of Aya, an open-source multilingual model covering 101 languages, as part of a global research collaboration.</li> + <li> + <strong>Command R & R+ Launch (Early 2024):</strong> Release of highly capable models optimized for + enterprise RAG and tool use, with competitive pricing and performance. + </li> + <li> + <strong>Cloud Expansion:</strong> Broadened availability on major cloud platforms, including new + integrations with Oracle Cloud Infrastructure (OCI) and Microsoft Azure. + </li> + <li> + <strong>Enterprise Tooling:</strong> Enhanced platform features for data management, model + fine-tuning, and deploying RAG applications. Learn more through their + <a href="https://docs.cohere.com/" target="_blank" rel="noopener noreferrer">documentation</a>. [9, + 29, 32, 35, 41] + </li> + <li> + <strong>Cohere Coral (Tech Preview):</strong> A knowledge assistant for enterprises, leveraging RAG + to connect to business data. + </li> + <li> + <strong>Aya Model (Collaboration):</strong> Contributed to the release of Aya, an open-source + multilingual model covering 101 languages, as part of a global research collaboration. + </li> </ul> </div> </div> </div> - <!-- Removed Key Links card for Cohere --> + <!-- Removed Key Links card for Cohere --> </div> </div> @@ -1856,13 +2457,29 @@ <h5><i class="bi bi-info-circle-fill"></i> Key Information</h5> <div class="card-content-wrapper"> <ul class="key-info-list"> - <li><strong>Founded:</strong> April 2023, by Arthur Mensch, Guillaume Lample, Timothée Lacroix. [2, 42]</li> + <li> + <strong>Founded:</strong> April 2023, by Arthur Mensch, Guillaume Lample, Timothée Lacroix. [2, + 42] + </li> <li><strong>Headquarters:</strong> Paris, France. [2]</li> - <li><strong>Valuation:</strong> ~$2 billion (as of Dec 2023), reported talks for $5-6 billion (early-mid 2024).</li> - <li><strong>Flagship Models:</strong> Open: Mistral 7B, Mixtral 8x7B, Mixtral 8x22B; Commercial: Mistral Large, Mistral Small, Mistral Embed. [2, 42]</li> + <li> + <strong>Valuation:</strong> ~$2 billion (as of Dec 2023), reported talks for $5-6 billion + (early-mid 2024). + </li> + <li> + <strong>Flagship Models:</strong> Open: Mistral 7B, Mixtral 8x7B, Mixtral 8x22B; Commercial: + Mistral Large, Mistral Small, Mistral Embed. [2, 42] + </li> <li><strong>Main Products:</strong> La Plateforme (API), Le Chat (chatbot), open-weight models.</li> - <li><strong>Official Website:</strong> <a href="https://mistral.ai/" target="_blank" rel="noopener noreferrer">mistral.ai</a></li> - <li><strong>Documentation:</strong> <a href="https://docs.mistral.ai/" target="_blank" rel="noopener noreferrer">docs.mistral.ai</a> [5, 7, 31, 33, 34]</li> + <li> + <strong>Official Website:</strong> + <a href="https://mistral.ai/" target="_blank" rel="noopener noreferrer">mistral.ai</a> + </li> + <li> + <strong>Documentation:</strong> + <a href="https://docs.mistral.ai/" target="_blank" rel="noopener noreferrer">docs.mistral.ai</a> + [5, 7, 31, 33, 34] + </li> </ul> </div> </div> @@ -1874,7 +2491,8 @@ <h5><i class="bi bi-flag-fill"></i> Origin & Focus</h5> <div class="card-content-wrapper"> <p class="summary"> - Paris-based company founded in early 2023 by former researchers from Meta and Google DeepMind. [2, 42] Focus on open, efficient, and powerful AI models. + Paris-based company founded in early 2023 by former researchers from Meta and Google DeepMind. [2, + 42] Focus on open, efficient, and powerful AI models. </p> <button class="btn btn-sm details-toggle" @@ -1890,9 +2508,18 @@ <div class="collapse collapse-content" id="collapseMistralOrigin"> <h6>Key Details</h6> <ul> - <li><strong>Founders:</strong> Arthur Mensch (CEO, previously DeepMind), Guillaume Lample (previously Meta), and Timothée Lacroix (previously Meta). [2]</li> - <li><strong>Mission:</strong> To develop cutting-edge generative AI models with a strong emphasis on openness, efficiency, and performance, aiming to be a European AI champion. [42]</li> - <li><strong>Rapid Growth:</strong> Quickly gained prominence and significant funding shortly after its inception.</li> + <li> + <strong>Founders:</strong> Arthur Mensch (CEO, previously DeepMind), Guillaume Lample (previously + Meta), and Timothée Lacroix (previously Meta). [2] + </li> + <li> + <strong>Mission:</strong> To develop cutting-edge generative AI models with a strong emphasis on + openness, efficiency, and performance, aiming to be a European AI champion. [42] + </li> + <li> + <strong>Rapid Growth:</strong> Quickly gained prominence and significant funding shortly after its + inception. + </li> </ul> </div> </div> @@ -1903,7 +2530,10 @@ <h5><i class="bi bi-wind"></i> Philosophy: Open & Efficient AI</h5> <div class="card-content-wrapper"> <p class="summary"> - Strong belief in open-weight models (Apache 2.0 license) for innovation, transparency, and community building. Focus on computational efficiency and model compactness. Models often released on <a href="https://huggingface.co/mistralai" target="_blank" rel="noopener noreferrer">Hugging Face</a>. + Strong belief in open-weight models (Apache 2.0 license) for innovation, transparency, and community + building. Focus on computational efficiency and model compactness. Models often released on + <a href="https://huggingface.co/mistralai" target="_blank" rel="noopener noreferrer">Hugging Face</a + >. </p> <button class="btn btn-sm details-toggle" @@ -1920,21 +2550,33 @@ <h6>Core Principles</h6> <ul> <li> - <strong>Openness:</strong> A key tenet. Releases many models with open weights under permissive licenses like Apache 2.0, allowing broad use and modification. [2] + <strong>Openness:</strong> A key tenet. Releases many models with open weights under permissive + licenses like Apache 2.0, allowing broad use and modification. [2] </li> <li> - <strong>Efficiency:</strong> Develops models that are powerful yet optimized for performance, aiming for better inference speed, lower computational costs, and smaller VRAM footprint. Utilizes techniques like Mixture-of-Experts (MoE). [42] + <strong>Efficiency:</strong> Develops models that are powerful yet optimized for performance, aiming + for better inference speed, lower computational costs, and smaller VRAM footprint. Utilizes + techniques like Mixture-of-Experts (MoE). [42] </li> <li> - <strong>Pragmatic Approach:</strong> Balances open-source contributions with optimized commercial API offerings (<a href="https://console.mistral.ai/" target="_blank" rel="noopener noreferrer">La Plateforme</a>) for enterprise use. [34] + <strong>Pragmatic Approach:</strong> Balances open-source contributions with optimized commercial + API offerings (<a href="https://console.mistral.ai/" target="_blank" rel="noopener noreferrer" + >La Plateforme</a + >) for enterprise use. [34] + </li> + <li> + <strong>European Leadership:</strong> Aims to build a leading AI company based in Europe, + contributing to the continent's AI ecosystem. + </li> + <li> + <strong>Trust and Independence:</strong> Emphasizes building trustworthy AI and maintaining + independence in its research and development direction. </li> - <li><strong>European Leadership:</strong> Aims to build a leading AI company based in Europe, contributing to the continent's AI ecosystem.</li> - <li><strong>Trust and Independence:</strong> Emphasizes building trustworthy AI and maintaining independence in its research and development direction.</li> </ul> </div> </div> </div> - <div class="col-lg-4 col-md-6"> + <div class="col-lg-4 col-md-6"> <div class="info-card type-leadership" id="card-mistral-leadership"> <div class="card-body"> <h5><i class="bi bi-person-badge"></i> Leadership</h5> @@ -1956,7 +2598,10 @@ <div class="collapse collapse-content" id="collapseMistralLeadership"> <h6>Key Figures</h6> <ul> - <li><strong>Arthur Mensch:</strong> Co-founder and Chief Executive Officer (CEO). Former researcher at Google DeepMind. [2]</li> + <li> + <strong>Arthur Mensch:</strong> Co-founder and Chief Executive Officer (CEO). Former researcher at + Google DeepMind. [2] + </li> <li><strong>Guillaume Lample:</strong> Co-founder. Former researcher at Meta AI (FAIR). [2]</li> <li><strong>Timothée Lacroix:</strong> Co-founder. Former researcher at Meta AI (FAIR). [2]</li> </ul> @@ -1969,7 +2614,10 @@ <h5><i class="bi bi-boxes"></i> Flagship Models & Products</h5> <div class="card-content-wrapper"> <p class="summary"> - Open models: Mistral 7B, Mixtral 8x7B, Mixtral 8x22B (MoE). [2, 42] Commercial via API: Mistral Large, Mistral Small, Mistral Embed. Access via <a href="https://console.mistral.ai/" target="_blank" rel="noopener noreferrer">La Plateforme</a>. [34] + Open models: Mistral 7B, Mixtral 8x7B, Mixtral 8x22B (MoE). [2, 42] Commercial via API: Mistral + Large, Mistral Small, Mistral Embed. Access via + <a href="https://console.mistral.ai/" target="_blank" rel="noopener noreferrer">La Plateforme</a>. + [34] </p> <button class="btn btn-sm details-toggle" @@ -1986,31 +2634,50 @@ <h6>Open-Weight Models (Apache 2.0 License)</h6> <ul> <li> - <strong>Mistral 7B:</strong> Highly efficient and performant small model, known for strong capabilities relative to its size. [42] + <strong>Mistral 7B:</strong> Highly efficient and performant small model, known for strong + capabilities relative to its size. [42] </li> <li> - <strong>Mixtral 8x7B:</strong> Sparse Mixture-of-Experts (MoE) model, offering high performance (comparable to GPT-3.5) with efficient inference due to activating only a fraction of parameters per token. [2, 42] + <strong>Mixtral 8x7B:</strong> Sparse Mixture-of-Experts (MoE) model, offering high performance + (comparable to GPT-3.5) with efficient inference due to activating only a fraction of parameters per + token. [2, 42] </li> - <li> - <strong>Mixtral 8x22B:</strong> A larger and more powerful open MoE model released in April 2024, offering even stronger performance. [2, 42] + <li> + <strong>Mixtral 8x22B:</strong> A larger and more powerful open MoE model released in April 2024, + offering even stronger performance. [2, 42] </li> </ul> <h6>Commercial Models (via La Plateforme API & Partners)</h6> <ul> <li> - <strong>Mistral Large:</strong> Flagship commercial model, top-tier reasoning capabilities, multilingual, and suitable for complex tasks. [2] + <strong>Mistral Large:</strong> Flagship commercial model, top-tier reasoning capabilities, + multilingual, and suitable for complex tasks. [2] </li> - <li> - <strong>Mistral Small (formerly Mistral Next):</strong> Optimized for latency and cost-effectiveness. [2] + <li> + <strong>Mistral Small (formerly Mistral Next):</strong> Optimized for latency and + cost-effectiveness. [2] </li> <li> - <strong>Mistral Embed (formerly Mistral Medium endpoint):</strong> State-of-the-art embedding model. [2] + <strong>Mistral Embed (formerly Mistral Medium endpoint):</strong> State-of-the-art embedding model. + [2] </li> </ul> - <h6>Platform Access</h6> + <h6>Platform Access</h6> <ul> - <li><strong>La Plateforme:</strong> Mistral AI's API platform for accessing their commercial models (<a href="https://console.mistral.ai/" target="_blank" rel="noopener noreferrer">console.mistral.ai</a>). Check their <a href="https://docs.mistral.ai/" target="_blank" rel="noopener noreferrer">API documentation</a>. [5, 7, 31, 33, 34]</li> - <li><strong>Partnerships:</strong> Models available through cloud providers like Microsoft Azure AI, AWS Bedrock, Google Cloud Vertex AI.</li> + <li> + <strong>La Plateforme:</strong> Mistral AI's API platform for accessing their commercial models (<a + href="https://console.mistral.ai/" + target="_blank" + rel="noopener noreferrer" + >console.mistral.ai</a + >). Check their + <a href="https://docs.mistral.ai/" target="_blank" rel="noopener noreferrer">API documentation</a>. + [5, 7, 31, 33, 34] + </li> + <li> + <strong>Partnerships:</strong> Models available through cloud providers like Microsoft Azure AI, AWS + Bedrock, Google Cloud Vertex AI. + </li> </ul> </div> </div> @@ -2021,7 +2688,8 @@ <h5><i class="bi bi-bullseye"></i> Approach to Advanced AI</h5> <div class="card-content-wrapper"> <p class="summary"> - Focuses on building foundational models that are powerful and efficient. Openness is seen as key for responsible AI development. AGI is a long-term direction. + Focuses on building foundational models that are powerful and efficient. Openness is seen as key for + responsible AI development. AGI is a long-term direction. </p> <button class="btn btn-sm details-toggle" @@ -2037,10 +2705,25 @@ <div class="collapse collapse-content" id="collapseMistralAGI"> <h6>Perspective on AGI</h6> <ul> - <li><strong>Building Blocks:</strong> Current focus is on creating highly capable and general-purpose foundational models that can serve a wide range of applications.</li> - <li><strong>Efficiency as a Driver:</strong> Belief that more efficient model architectures (like MoE) are crucial for scaling capabilities sustainably. [42]</li> - <li><strong>Openness for Safety and Understanding:</strong> By releasing models openly, Mistral AI aims to foster community research into their capabilities, limitations, and safety aspects, contributing to a broader understanding required for any future AGI. [2]</li> - <li><strong>Pragmatic Development:</strong> While the long-term trajectory of AI points towards increasingly general intelligence, Mistral's current public emphasis is on delivering tangible value with existing and near-term models. Explicit AGI timelines are not a central part of their public messaging.</li> + <li> + <strong>Building Blocks:</strong> Current focus is on creating highly capable and general-purpose + foundational models that can serve a wide range of applications. + </li> + <li> + <strong>Efficiency as a Driver:</strong> Belief that more efficient model architectures (like MoE) + are crucial for scaling capabilities sustainably. [42] + </li> + <li> + <strong>Openness for Safety and Understanding:</strong> By releasing models openly, Mistral AI aims + to foster community research into their capabilities, limitations, and safety aspects, contributing + to a broader understanding required for any future AGI. [2] + </li> + <li> + <strong>Pragmatic Development:</strong> While the long-term trajectory of AI points towards + increasingly general intelligence, Mistral's current public emphasis is on delivering tangible value + with existing and near-term models. Explicit AGI timelines are not a central part of their public + messaging. + </li> </ul> </div> </div> @@ -2051,7 +2734,8 @@ <h5><i class="bi bi-cash-coin"></i> Funding & Partnerships</h5> <div class="card-content-wrapper"> <p class="summary"> - Rapidly raised significant funding. Key investors include Andreessen Horowitz, Lightspeed. Strategic partnership with Microsoft (Azure distribution and investment). + Rapidly raised significant funding. Key investors include Andreessen Horowitz, Lightspeed. Strategic + partnership with Microsoft (Azure distribution and investment). </p> <button class="btn btn-sm details-toggle" @@ -2067,26 +2751,37 @@ <div class="collapse collapse-content" id="collapseMistralFunding"> <h6>Investment</h6> <ul> - <li><strong>Seed Round (June 2023):</strong> €105 million ($113 million), led by Lightspeed Venture Partners, with participation from Redpoint, Index Ventures, and others.</li> - <li><strong>Series A (December 2023):</strong> €385 million ($415 million), led by Andreessen Horowitz (a16z), with Lightspeed Venture Partners, Salesforce, BNP Paribas, CMA CGM, General Catalyst, Elad Gil, and Nvidia participating. Valued at ~$2 billion.</li> + <li> + <strong>Seed Round (June 2023):</strong> €105 million ($113 million), led by Lightspeed Venture + Partners, with participation from Redpoint, Index Ventures, and others. + </li> + <li> + <strong>Series A (December 2023):</strong> €385 million ($415 million), led by Andreessen Horowitz + (a16z), with Lightspeed Venture Partners, Salesforce, BNP Paribas, CMA CGM, General Catalyst, Elad + Gil, and Nvidia participating. Valued at ~$2 billion. + </li> </ul> <h6>Strategic Alliances</h6> <ul> <li> - <strong>Microsoft (Feb 2024):</strong> Multi-year partnership including Microsoft making a €15 million investment. Mistral's commercial models became available on Microsoft Azure AI platform, and collaboration on bringing models to Azure customers. [2] + <strong>Microsoft (Feb 2024):</strong> Multi-year partnership including Microsoft making a €15 + million investment. Mistral's commercial models became available on Microsoft Azure AI platform, and + collaboration on bringing models to Azure customers. [2] </li> - <li>Distribution partnerships with other cloud providers like AWS and Google Cloud.</li> + <li>Distribution partnerships with other cloud providers like AWS and Google Cloud.</li> </ul> </div> </div> </div> - <div class="col-lg-4 col-md-6"> + <div class="col-lg-4 col-md-6"> <div class="info-card type-developments" id="card-mistral-developments"> <div class="card-body"> <h5><i class="bi bi-newspaper"></i> Recent Developments (2024-2025)</h5> <div class="card-content-wrapper"> <p class="summary"> - Release of Mistral Large and Mistral Small via API. [2] Launch of Mixtral 8x22B (open model). [2, 42] Partnership with Microsoft. [2] Expanding cloud availability. Read their <a href="https://mistral.ai/news/" target="_blank" rel="noopener noreferrer">news</a>. + Release of Mistral Large and Mistral Small via API. [2] Launch of Mixtral 8x22B (open model). [2, + 42] Partnership with Microsoft. [2] Expanding cloud availability. Read their + <a href="https://mistral.ai/news/" target="_blank" rel="noopener noreferrer">news</a>. </p> <button class="btn btn-sm details-toggle" @@ -2102,13 +2797,34 @@ <div class="collapse collapse-content" id="collapseMistralDevelopments"> <h6>Key Announcements</h6> <ul> - <li><strong>Mistral Large (Feb 2024):</strong> Launch of their flagship commercial model, positioned as a top-tier reasoning model. [2]</li> - <li><strong>Mistral Small & Mistral Embed (Feb 2024):</strong> Release of more cost-effective and specialized API models. [2]</li> - <li><strong>Mixtral 8x22B (April 2024):</strong> Open release of a powerful 176B parameter MoE model (44B active). [2, 42]</li> - <li><strong>Microsoft Partnership (Feb 2024):</strong> Strategic partnership including investment and making Mistral models available on Azure. [2]</li> - <li><strong>Cloud Platform Expansion:</strong> Models increasingly available on AWS Bedrock, Google Cloud Vertex AI, and other platforms.</li> - <li><strong>"Le Chat" (Feb 2024):</strong> Launch of their own conversational AI assistant, initially in beta. [2]</li> - <li><strong>Codestral & Mathstral (Mid 2024):</strong> Release of specialized open models for code and STEM. [2]</li> + <li> + <strong>Mistral Large (Feb 2024):</strong> Launch of their flagship commercial model, positioned as + a top-tier reasoning model. [2] + </li> + <li> + <strong>Mistral Small & Mistral Embed (Feb 2024):</strong> Release of more cost-effective and + specialized API models. [2] + </li> + <li> + <strong>Mixtral 8x22B (April 2024):</strong> Open release of a powerful 176B parameter MoE model + (44B active). [2, 42] + </li> + <li> + <strong>Microsoft Partnership (Feb 2024):</strong> Strategic partnership including investment and + making Mistral models available on Azure. [2] + </li> + <li> + <strong>Cloud Platform Expansion:</strong> Models increasingly available on AWS Bedrock, Google + Cloud Vertex AI, and other platforms. + </li> + <li> + <strong>"Le Chat" (Feb 2024):</strong> Launch of their own conversational AI assistant, initially in + beta. [2] + </li> + <li> + <strong>Codestral & Mathstral (Mid 2024):</strong> Release of specialized open models for code and + STEM. [2] + </li> </ul> </div> </div> @@ -2132,8 +2848,15 @@ <li><strong>Valuation:</strong> $1.4 billion (as of August 2023).</li> <li><strong>Flagship Models:</strong> Jurassic series, Jamba (SSM-Transformer hybrid).</li> <li><strong>Main Products:</strong> Wordtune, AI21 Studio (API).</li> - <li><strong>Official Website:</strong> <a href="https://www.ai21.com/" target="_blank" rel="noopener noreferrer">www.ai21.com</a></li> - <li><strong>Documentation:</strong> <a href="https://docs.ai21.com/" target="_blank" rel="noopener noreferrer">docs.ai21.com</a> (via Studio)</li> + <li> + <strong>Official Website:</strong> + <a href="https://www.ai21.com/" target="_blank" rel="noopener noreferrer">www.ai21.com</a> + </li> + <li> + <strong>Documentation:</strong> + <a href="https://docs.ai21.com/" target="_blank" rel="noopener noreferrer">docs.ai21.com</a> (via + Studio) + </li> </ul> </div> </div> @@ -2145,7 +2868,8 @@ <h5><i class="bi bi-flag-fill"></i> Origin & Focus</h5> <div class="card-content-wrapper"> <p class="summary"> - Israeli company founded in 2017 by AI luminaries. Aims to reimagine how humans read and write using AI, focusing on deep context understanding and reasoning. + Israeli company founded in 2017 by AI luminaries. Aims to reimagine how humans read and write using + AI, focusing on deep context understanding and reasoning. </p> <button class="btn btn-sm details-toggle" @@ -2161,8 +2885,15 @@ <div class="collapse collapse-content" id="collapseAI21Origin"> <h6>Key Details</h6> <ul> - <li><strong>Founders:</strong> Prof. Yoav Shoham (Stanford emeritus), Ori Goshen, and Prof. Amnon Shashua (co-founder of Mobileye, Intel SVP).</li> - <li><strong>Mission:</strong> To build AI systems that deeply understand context and meaning, moving beyond pattern matching to more robust reasoning, thereby augmenting human capabilities in reading comprehension and text generation.</li> + <li> + <strong>Founders:</strong> Prof. Yoav Shoham (Stanford emeritus), Ori Goshen, and Prof. Amnon + Shashua (co-founder of Mobileye, Intel SVP). + </li> + <li> + <strong>Mission:</strong> To build AI systems that deeply understand context and meaning, moving + beyond pattern matching to more robust reasoning, thereby augmenting human capabilities in reading + comprehension and text generation. + </li> <li><strong>Headquarters:</strong> Tel Aviv, Israel.</li> </ul> </div> @@ -2174,7 +2905,9 @@ <h5><i class="bi bi-pencil-fill"></i> Philosophy: AI for Reading & Writing</h5> <div class="card-content-wrapper"> <p class="summary"> - Focuses on developing AI that serves as a true partner in text-based work. Emphasizes proprietary LLMs, task-specific models, and architectural innovation (e.g., Jamba). Read more on their <a href="https://www.ai21.com/blog" target="_blank" rel="noopener noreferrer">blog</a>. + Focuses on developing AI that serves as a true partner in text-based work. Emphasizes proprietary + LLMs, task-specific models, and architectural innovation (e.g., Jamba). Read more on their + <a href="https://www.ai21.com/blog" target="_blank" rel="noopener noreferrer">blog</a>. </p> <button class="btn btn-sm details-toggle" @@ -2190,16 +2923,34 @@ <div class="collapse collapse-content" id="collapseAI21Philosophy"> <h6>Core Approach</h6> <ul> - <li><strong>Deep Language Understanding:</strong> Aims to build AI that genuinely grasps context, semantics, and nuance in language, rather than just superficial pattern matching.</li> - <li><strong>Augmenting Human Intellect:</strong> Develops tools (like <a href="https://www.wordtune.com/" target="_blank" rel="noopener noreferrer">Wordtune</a>) to enhance human writing and reading capabilities, making communication more effective and information consumption more efficient.</li> - <li><strong>Task-Specific Models:</strong> Increasingly focuses on developing models optimized for specific enterprise tasks (e.g., reliable summarization, grounded question answering) to improve accuracy and reduce hallucinations.</li> - <li><strong>Architectural Innovation:</strong> Explores and implements novel model architectures like Jamba (SSM-Transformer hybrid) to balance performance, efficiency, and context length.</li> - <li><strong>Neuro-Symbolic AI (Mentioned):</strong> Co-CEOs have expressed interest in combining LLMs with symbolic reasoning for more robust and explainable AI.</li> + <li> + <strong>Deep Language Understanding:</strong> Aims to build AI that genuinely grasps context, + semantics, and nuance in language, rather than just superficial pattern matching. + </li> + <li> + <strong>Augmenting Human Intellect:</strong> Develops tools (like + <a href="https://www.wordtune.com/" target="_blank" rel="noopener noreferrer">Wordtune</a>) to + enhance human writing and reading capabilities, making communication more effective and information + consumption more efficient. + </li> + <li> + <strong>Task-Specific Models:</strong> Increasingly focuses on developing models optimized for + specific enterprise tasks (e.g., reliable summarization, grounded question answering) to improve + accuracy and reduce hallucinations. + </li> + <li> + <strong>Architectural Innovation:</strong> Explores and implements novel model architectures like + Jamba (SSM-Transformer hybrid) to balance performance, efficiency, and context length. + </li> + <li> + <strong>Neuro-Symbolic AI (Mentioned):</strong> Co-CEOs have expressed interest in combining LLMs + with symbolic reasoning for more robust and explainable AI. + </li> </ul> </div> </div> </div> - <div class="col-lg-4 col-md-6"> + <div class="col-lg-4 col-md-6"> <div class="info-card type-leadership" id="card-ai21-leadership"> <div class="card-body"> <h5><i class="bi bi-person-badge"></i> Leadership</h5> @@ -2222,8 +2973,14 @@ <h6>Key Figures</h6> <ul> <li><strong>Ori Goshen:</strong> Co-founder and Co-Chief Executive Officer (CEO).</li> - <li><strong>Prof. Yoav Shoham:</strong> Co-founder and Co-Chief Executive Officer (CEO). Professor Emeritus at Stanford University.</li> - <li><strong>Prof. Amnon Shashua:</strong> Co-founder and Chairman. Co-founder of Mobileye and Senior Vice President at Intel.</li> + <li> + <strong>Prof. Yoav Shoham:</strong> Co-founder and Co-Chief Executive Officer (CEO). Professor + Emeritus at Stanford University. + </li> + <li> + <strong>Prof. Amnon Shashua:</strong> Co-founder and Chairman. Co-founder of Mobileye and Senior + Vice President at Intel. + </li> </ul> </div> </div> @@ -2234,7 +2991,11 @@ <h5><i class="bi bi-boxes"></i> Flagship Models & Products</h5> <div class="card-content-wrapper"> <p class="summary"> - Jurassic model series. Jamba (hybrid SSM-Transformer, open weights). <a href="https://www.wordtune.com/" target="_blank" rel="noopener noreferrer">Wordtune</a> (AI writing/reading assistant). <a href="https://studio.ai21.com/" target="_blank" rel="noopener noreferrer">AI21 Studio</a> for developers. + Jurassic model series. Jamba (hybrid SSM-Transformer, open weights). + <a href="https://www.wordtune.com/" target="_blank" rel="noopener noreferrer">Wordtune</a> (AI + writing/reading assistant). + <a href="https://studio.ai21.com/" target="_blank" rel="noopener noreferrer">AI21 Studio</a> for + developers. </p> <button class="btn btn-sm details-toggle" @@ -2251,32 +3012,56 @@ <h6>Model Families & Architectures</h6> <ul> <li> - <strong>Jurassic Series (e.g., Jurassic-2):</strong> Family of proprietary large language models with varying sizes and capabilities, designed for sophisticated language tasks. + <strong>Jurassic Series (e.g., Jurassic-2):</strong> Family of proprietary large language models + with varying sizes and capabilities, designed for sophisticated language tasks. </li> <li> - <strong>Jamba (e.g., Jamba 1.5 Mini, Jamba 1.5 Large):</strong> Innovative model architecture combining Transformer blocks with Mamba (State Space Model) blocks and Mixture-of-Experts (MoE). Aims for efficiency, large context window (256K), and strong performance. Openly available versions released. + <strong>Jamba (e.g., Jamba 1.5 Mini, Jamba 1.5 Large):</strong> Innovative model architecture + combining Transformer blocks with Mamba (State Space Model) blocks and Mixture-of-Experts (MoE). + Aims for efficiency, large context window (256K), and strong performance. Openly available versions + released. </li> </ul> <h6>Applications & Platform</h6> <ul> <li> - <strong><a href="https://www.wordtune.com/" target="_blank" rel="noopener noreferrer">Wordtune</a>:</strong> AI-powered writing companion that helps rephrase, summarize, generate text, and check grammar/spelling. Includes Wordtune Read for summarizing long documents. + <strong + ><a href="https://www.wordtune.com/" target="_blank" rel="noopener noreferrer">Wordtune</a + >:</strong + > + AI-powered writing companion that helps rephrase, summarize, generate text, and check + grammar/spelling. Includes Wordtune Read for summarizing long documents. </li> <li> - <strong><a href="https://studio.ai21.com/" target="_blank" rel="noopener noreferrer">AI21 Studio</a>:</strong> Developer platform providing API access (see <a href="https://docs.ai21.com/docs/introduction-to-ai21-studio" target="_blank" rel="noopener noreferrer">docs</a>) to their models (Jurassic, Jamba, task-specific models) for building custom NLP applications. + <strong + ><a href="https://studio.ai21.com/" target="_blank" rel="noopener noreferrer">AI21 Studio</a + >:</strong + > + Developer platform providing API access (see + <a + href="https://docs.ai21.com/docs/introduction-to-ai21-studio" + target="_blank" + rel="noopener noreferrer" + >docs</a + >) to their models (Jurassic, Jamba, task-specific models) for building custom NLP applications. + </li> + <li> + <strong>Task-Specific Models:</strong> Models optimized for particular enterprise needs, such as + contextual answers, summarization, and paraphrasing. </li> - <li><strong>Task-Specific Models:</strong> Models optimized for particular enterprise needs, such as contextual answers, summarization, and paraphrasing.</li> </ul> </div> </div> </div> - <div class="col-lg-4 col-md-6"> + <div class="col-lg-4 col-md-6"> <div class="info-card type-agi" id="card-ai21-agi"> <div class="card-body"> <h5><i class="bi bi-bullseye"></i> Approach to Advanced AI</h5> <div class="card-content-wrapper"> <p class="summary"> - Focuses on reliable and controllable AI for practical applications. Explores novel architectures and neuro-symbolic ideas for more robust intelligence, rather than explicit AGI pursuit as a primary public goal. + Focuses on reliable and controllable AI for practical applications. Explores novel architectures and + neuro-symbolic ideas for more robust intelligence, rather than explicit AGI pursuit as a primary + public goal. </p> <button class="btn btn-sm details-toggle" @@ -2292,11 +3077,30 @@ <div class="collapse collapse-content" id="collapseAI21AGI"> <h6>Perspective on AGI/ASI</h6> <ul> - <li><strong>Practical and Reliable AI:</strong> The primary focus is on building AI systems that are trustworthy, predictable, and provide tangible value in augmenting human reading and writing tasks, especially for enterprises.</li> - <li><strong>Architectural Innovation for Capability:</strong> The development of models like Jamba indicates a drive towards more efficient and capable systems, which are foundational steps for any advanced AI.</li> - <li><strong>Reasoning and Understanding:</strong> Strong emphasis on moving beyond pattern-matching to AI that exhibits deeper reasoning and contextual understanding, key components of more general intelligence.</li> - <li><strong>Neuro-Symbolic Exploration:</strong> Co-CEOs have discussed the potential of combining large language models with symbolic AI techniques to enhance robustness, explainability, and reasoning capabilities, which could be a path towards more advanced AI.</li> - <li>While not explicitly framed as an AGI race, their work on sophisticated reasoning and novel architectures contributes to the broader field of advanced AI research.</li> + <li> + <strong>Practical and Reliable AI:</strong> The primary focus is on building AI systems that are + trustworthy, predictable, and provide tangible value in augmenting human reading and writing tasks, + especially for enterprises. + </li> + <li> + <strong>Architectural Innovation for Capability:</strong> The development of models like Jamba + indicates a drive towards more efficient and capable systems, which are foundational steps for any + advanced AI. + </li> + <li> + <strong>Reasoning and Understanding:</strong> Strong emphasis on moving beyond pattern-matching to + AI that exhibits deeper reasoning and contextual understanding, key components of more general + intelligence. + </li> + <li> + <strong>Neuro-Symbolic Exploration:</strong> Co-CEOs have discussed the potential of combining large + language models with symbolic AI techniques to enhance robustness, explainability, and reasoning + capabilities, which could be a path towards more advanced AI. + </li> + <li> + While not explicitly framed as an AGI race, their work on sophisticated reasoning and novel + architectures contributes to the broader field of advanced AI research. + </li> </ul> </div> </div> @@ -2307,7 +3111,8 @@ <h5><i class="bi bi-piggy-bank"></i> Funding & Investors</h5> <div class="card-content-wrapper"> <p class="summary"> - Raised over $336M. Series C (2023) valued at $1.4B, with investors like Google, Nvidia, Intel Capital, Comcast Ventures, Walden Catalyst, Pitango. + Raised over $336M. Series C (2023) valued at $1.4B, with investors like Google, Nvidia, Intel + Capital, Comcast Ventures, Walden Catalyst, Pitango. </p> <button class="btn btn-sm details-toggle" @@ -2323,10 +3128,21 @@ <div class="collapse collapse-content" id="collapseAI21Funding"> <h6>Key Investment Rounds</h6> <ul> - <li><strong>Seed & Series A:</strong> Early funding rounds helped establish the company and initial product development.</li> + <li> + <strong>Seed & Series A:</strong> Early funding rounds helped establish the company and initial + product development. + </li> <li><strong>Series B (July 2022):</strong> Raised $64 million, led by Ahren Innovation Capital.</li> - <li><strong>Series C (August 2023):</strong> Announced $155 million financing, valuing the company at $1.4 billion. Investors included Google, Nvidia, Walden Catalyst, Pitango, SCB10X, b2venture, Samsung Next, and Prof. Amnon Shashua.</li> - <li><strong>Series C Extension (November 2023):</strong> Added $53 million to the Series C, bringing the total round to $208 million and total funding to $336 million. New investors included Intel Capital and Comcast Ventures.</li> + <li> + <strong>Series C (August 2023):</strong> Announced $155 million financing, valuing the company at + $1.4 billion. Investors included Google, Nvidia, Walden Catalyst, Pitango, SCB10X, b2venture, + Samsung Next, and Prof. Amnon Shashua. + </li> + <li> + <strong>Series C Extension (November 2023):</strong> Added $53 million to the Series C, bringing the + total round to $208 million and total funding to $336 million. New investors included Intel Capital + and Comcast Ventures. + </li> </ul> </div> </div> @@ -2337,7 +3153,9 @@ <h5><i class="bi bi-newspaper"></i> Recent Developments (2024-2025)</h5> <div class="card-content-wrapper"> <p class="summary"> - Release of Jamba open-weight models (March 2024). Jamba 1.5 Mini and Large (Aug 2024). Maestro AI planning system (March 2025). Focus on enterprise solutions. See their <a href="https://www.ai21.com/newsroom" target="_blank" rel="noopener noreferrer">newsroom</a>. + Release of Jamba open-weight models (March 2024). Jamba 1.5 Mini and Large (Aug 2024). Maestro AI + planning system (March 2025). Focus on enterprise solutions. See their + <a href="https://www.ai21.com/newsroom" target="_blank" rel="noopener noreferrer">newsroom</a>. </p> <button class="btn btn-sm details-toggle" @@ -2353,22 +3171,78 @@ <div class="collapse collapse-content" id="collapseAI21Developments"> <h6>Key Announcements</h6> <ul> - <li><strong>Jamba Release (March 2024):</strong> Launched Jamba, the first production-grade Mamba-based model, featuring a hybrid SSM-Transformer architecture and open weights.</li> - <li><strong>Jamba 1.5 Mini & Large (August 2024):</strong> Released new versions of Jamba with enhanced performance, expanded capabilities, and a large 256K context window, available as open models.</li> - <li><strong>Maestro AI (March 2025):</strong> Unveiled Maestro, an AI planning and orchestration system for enterprises, designed to enhance operational efficiency.</li> - <li><strong>Task-Specific Models:</strong> Continued emphasis on developing and refining models for specific enterprise use-cases to ensure reliability and accuracy.</li> - <li><strong>Executive Appointments:</strong> Hired Sharon Argov as Chief Marketing Officer and Yaniv Vakrat as Chief Revenue Officer (2024).</li> + <li> + <strong>Jamba Release (March 2024):</strong> Launched Jamba, the first production-grade Mamba-based + model, featuring a hybrid SSM-Transformer architecture and open weights. + </li> + <li> + <strong>Jamba 1.5 Mini & Large (August 2024):</strong> Released new versions of Jamba with enhanced + performance, expanded capabilities, and a large 256K context window, available as open models. + </li> + <li> + <strong>Maestro AI (March 2025):</strong> Unveiled Maestro, an AI planning and orchestration system + for enterprises, designed to enhance operational efficiency. + </li> + <li> + <strong>Task-Specific Models:</strong> Continued emphasis on developing and refining models for + specific enterprise use-cases to ensure reliability and accuracy. + </li> + <li> + <strong>Executive Appointments:</strong> Hired Sharon Argov as Chief Marketing Officer and Yaniv + Vakrat as Chief Revenue Officer (2024). + </li> </ul> </div> </div> </div> - <!-- Removed Key Links card for AI21 Labs --> + <!-- Removed Key Links card for AI21 Labs --> </div> </div> </div> <footer class="container text-center pb-3"> - <p class="mb-2">© 2025 David Veksler</p> + <div class="mb-3"> + <h6 + style=" + color: var(--text-color-primary); + font-family: var(--font-headings); + font-size: 1rem; + margin-bottom: 0.5rem; + " + > + Explore Related AI Cheatsheets: + </h6> + <a + href="https://cheatsheets.davidveksler.com/airisk.html" + title="AI Existential Risk (X-Risk) Cheatsheet" + target="_blank" + rel="noopener noreferrer" + class="mx-2" + style="font-size: 0.9em" + > + <i class="bi bi-shield-exclamation"></i> AI X-Risk + </a> + <a + href="https://cheatsheets.davidveksler.com/ai-frontier.html" + title="AI Frontier Model Builders Cheatsheet" + target="_blank" + rel="noopener noreferrer" + class="mx-2" + style="font-size: 0.9em" + > + <i class="bi bi-cpu-fill"></i> AI Frontier Models + </a> + <a + href="https://cheatsheets.davidveksler.com/aisafety.html" + title="Interactive AI Safety Ecosystem Hub" + target="_blank" + rel="noopener noreferrer" + class="mx-2" + style="font-size: 0.9em" + > + <i class="bi bi-diagram-3-fill"></i> AI Safety Ecosystem + </a> + </div> <div> <a href="https://www.linkedin.com/in/davidveksler/" @@ -2478,13 +3352,13 @@ if (targetCollapse && icon) { // Initialize icon state based on whether collapse is shown if (targetCollapse.classList.contains("show")) { - icon.classList.remove("bi-chevron-down"); - icon.classList.add("bi-chevron-up"); - button.setAttribute("aria-expanded", "true"); + icon.classList.remove("bi-chevron-down"); + icon.classList.add("bi-chevron-up"); + button.setAttribute("aria-expanded", "true"); } else { - icon.classList.remove("bi-chevron-up"); - icon.classList.add("bi-chevron-down"); - button.setAttribute("aria-expanded", "false"); + icon.classList.remove("bi-chevron-up"); + icon.classList.add("bi-chevron-down"); + button.setAttribute("aria-expanded", "false"); } targetCollapse.addEventListener("show.bs.collapse", () => { @@ -2496,8 +3370,8 @@ icon.classList.add("bi-chevron-down"); }); } else { - if (!targetCollapse) console.warn(`Collapse target ${targetSelector} not found for a button.`); - if (!icon) console.warn(`Icon not found in button for target ${targetSelector}.`); + if (!targetCollapse) console.warn(`Collapse target ${targetSelector} not found for a button.`); + if (!icon) console.warn(`Icon not found in button for target ${targetSelector}.`); } } catch (e) { console.error(`Error processing toggle for target ${targetSelector}: ${e}`); @@ -2506,4 +3380,4 @@ }); </script> </body> -</html> \ No newline at end of file +</html> --- a/airisk.html +++ b/airisk.html @@ -1,42 +1,61 @@ <!DOCTYPE html> <html lang="en"> -<head> + <head> <!-- === METADATA === --> - <meta charset="UTF-8"> - <meta name="viewport" content="width=device-width, initial-scale=1.0"> + <meta charset="UTF-8" /> + <meta name="viewport" content="width=device-width, initial-scale=1.0" /> <!-- SEO: Title - Critical for SEO, keep it concise and keyword-rich (50-60 characters) --> <title>AI Existential Risk (X-Risk) Cheatsheet: AGI, Safety & Mitigation</title> <!-- SEO: Description - Compelling summary for SERPs (150-160 characters) --> - <meta name="description" content="Comprehensive cheatsheet on AI Existential Risk (X-Risk): Understanding AGI, ASI, the alignment problem, risk scenarios, core challenges, mitigation strategies, and AI safety resources."> + <meta + name="description" + content="Comprehensive cheatsheet on AI Existential Risk (X-Risk): Understanding AGI, ASI, the alignment problem, risk scenarios, core challenges, mitigation strategies, and AI safety resources." + /> <!-- SEO: Keywords - Still has minor relevance for some search engines --> - <meta name="keywords" content="AI Safety, Existential Risk, X-Risk, Artificial Intelligence, AGI, ASI, AI Alignment, Control Problem, AI Governance, AI Ethics, Superintelligence, Machine Learning Safety, AI Risk Mitigation, Nick Bostrom, Eliezer Yudkowsky, MIRI, OpenAI, Anthropic, DeepMind, Future of Humanity"> + <meta + name="keywords" + content="AI Safety, Existential Risk, X-Risk, Artificial Intelligence, AGI, ASI, AI Alignment, Control Problem, AI Governance, AI Ethics, Superintelligence, Machine Learning Safety, AI Risk Mitigation, Nick Bostrom, Eliezer Yudkowsky, MIRI, OpenAI, Anthropic, DeepMind, Future of Humanity" + /> <!-- SEO: Author --> - <meta name="author" content="David Veksler"> + <meta name="author" content="David Veksler" /> <!-- SEO: Robots - Explicitly state indexing and following policy --> - <meta name="robots" content="index, follow"> + <meta name="robots" content="index, follow" /> <!-- Favicon: Relevant icon for AI Existential Risk (Brain + Question Mark) --> - <link rel="icon" href="data:image/svg+xml,<svg xmlns=%22http://www.w3.org/2000/svg%22 viewBox=%220 0 100 100%22><text y=%22.9em%22 font-size=%2290%22>🧠❓</text></svg>"> + <link + rel="icon" + href="data:image/svg+xml,<svg xmlns=%22http://www.w3.org/2000/svg%22 viewBox=%220 0 100 100%22><text y=%22.9em%22 font-size=%2290%22>🧠❓</text></svg>" + /> <!-- SEO: Canonical URL - Prevents duplicate content issues --> - <link rel="canonical" href="https://cheatsheets.davidveksler.com/airisk.html"> + <link rel="canonical" href="https://cheatsheets.davidveksler.com/airisk.html" /> <!-- Open Graph / Facebook / LinkedIn - For social sharing --> - <meta property="og:title" content="Understanding AI Existential Risk (X-Risk) - A Comprehensive Cheatsheet"> - <meta property="og:description" content="Explore AI Existential Risk: AGI, ASI, the alignment problem, scenarios, challenges, mitigation, and key AI safety resources. Essential for understanding advanced AI dangers."> - <meta property="og:type" content="article"> <!-- More specific than "website" for a cheatsheet --> - <meta property="og:url" content="https://cheatsheets.davidveksler.com/airisk.html"> - <meta property="og:image" content="https://cheatsheets.davidveksler.com/images/ai-xrisk-og.png"> <!-- *** IMPORTANT: Replace with an actual, compelling image URL (e.g., 1200x630px) *** --> - <meta property="og:image:alt" content="Conceptual image representing Artificial Intelligence risk and safety considerations for X-Risk"> - <meta property="og:image:width" content="1200"> <!-- Optional: Specify image dimensions --> - <meta property="og:image:height" content="630"> <!-- Optional: Specify image dimensions --> - <meta property="og:site_name" content="David Veksler's Cheatsheets"> - <meta property="og:locale" content="en_US"> + <meta property="og:title" content="Understanding AI Existential Risk (X-Risk) - A Comprehensive Cheatsheet" /> + <meta + property="og:description" + content="Explore AI Existential Risk: AGI, ASI, the alignment problem, scenarios, challenges, mitigation, and key AI safety resources. Essential for understanding advanced AI dangers." + /> + <meta property="og:type" content="article" /> + <!-- More specific than "website" for a cheatsheet --> + <meta property="og:url" content="https://cheatsheets.davidveksler.com/airisk.html" /> + <meta property="og:image" content="https://cheatsheets.davidveksler.com/images/ai-xrisk-og.png" /> + <!-- *** IMPORTANT: Replace with an actual, compelling image URL (e.g., 1200x630px) *** --> + <meta + property="og:image:alt" + content="Conceptual image representing Artificial Intelligence risk and safety considerations for X-Risk" + /> + <meta property="og:image:width" content="1200" /> + <!-- Optional: Specify image dimensions --> + <meta property="og:image:height" content="630" /> + <!-- Optional: Specify image dimensions --> + <meta property="og:site_name" content="David Veksler's Cheatsheets" /> + <meta property="og:locale" content="en_US" /> <!-- Optional: If relevant, connect to a Facebook App ID --> <!-- <meta property="fb:app_id" content="YOUR_FB_APP_ID"> --> <!-- Optional: If the article is part of a series or larger body of work --> @@ -46,518 +65,1162 @@ <!-- <meta property="article:published_time" content="YYYY-MM-DDTHH:MM:SSZ"> --> <!-- <meta property="article:author" content="David Veksler"> --> - <!-- Twitter Card - For Twitter sharing --> - <meta name="twitter:card" content="summary_large_image"> <!-- Ensures a large image is shown --> - <meta name="twitter:title" content="AI Existential Risk (X-Risk) Cheatsheet: AGI, Safety & Mitigation"> - <meta name="twitter:description" content="Your go-to guide for understanding AI Existential Risk (X-Risk), covering AGI, alignment, scenarios, challenges, AI safety measures, and resources."> - <meta name="twitter:url" content="https://cheatsheets.davidveksler.com/airisk.html"> - <meta name="twitter:image" content="https://cheatsheets.davidveksler.com/images/ai-xrisk-og.png"> <!-- *** IMPORTANT: Use the same image URL as og:image *** --> - <meta name="twitter:image:alt" content="Visual representation of AI Safety and Existential Risk concepts"> + <meta name="twitter:card" content="summary_large_image" /> + <!-- Ensures a large image is shown --> + <meta name="twitter:title" content="AI Existential Risk (X-Risk) Cheatsheet: AGI, Safety & Mitigation" /> + <meta + name="twitter:description" + content="Your go-to guide for understanding AI Existential Risk (X-Risk), covering AGI, alignment, scenarios, challenges, AI safety measures, and resources." + /> + <meta name="twitter:url" content="https://cheatsheets.davidveksler.com/airisk.html" /> + <meta name="twitter:image" content="https://cheatsheets.davidveksler.com/images/ai-xrisk-og.png" /> + <!-- *** IMPORTANT: Use the same image URL as og:image *** --> + <meta name="twitter:image:alt" content="Visual representation of AI Safety and Existential Risk concepts" /> <!-- Optional: Twitter site and creator handles --> <!-- <meta name="twitter:site" content="@YourSiteHandle"> --> <!-- <meta name="twitter:creator" content="@DavidVekslerHandle"> --> <!-- Browser Theme Color (Optional, for mobile browsers) --> - <meta name="theme-color" content="#1c3d7e"> <!-- Using --primary-color from your CSS --> + <meta name="theme-color" content="#1c3d7e" /> + <!-- Using --primary-color from your CSS --> <!-- === CSS === --> <!-- Bootstrap CSS --> - <link href="https://cdn.jsdelivr.net/npm/[email protected]/dist/css/bootstrap.min.css" rel="stylesheet" integrity="sha384-QWTKZyjpPEjISv5WaRU9OFeRpok6YctnYmDr5pNlyT2bRjXh0JMhjY6hW+ALEwIH" crossorigin="anonymous"> + <link + href="https://cdn.jsdelivr.net/npm/[email protected]/dist/css/bootstrap.min.css" + rel="stylesheet" + integrity="sha384-QWTKZyjpPEjISv5WaRU9OFeRpok6YctnYmDr5pNlyT2bRjXh0JMhjY6hW+ALEwIH" + crossorigin="anonymous" + /> <!-- Bootstrap Icons CSS --> - <link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/font/bootstrap-icons.min.css"> + <link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/font/bootstrap-icons.min.css" /> <!-- Custom CSS (Your existing style block would go here) --> <style> - :root { - --primary-color: #1c3d7e; /* Deep Blue */ - --secondary-color: #3b5998; /* Medium Blue */ - --accent-color: #0d6efd; /* Bright Blue for links */ - --hover-accent-color: #0a58ca; - --light-bg: #f0f4f8; /* Light grayish blue */ - --header-bg: #e2eafc; /* Lighter blue for header */ - --card-bg: #ffffff; - --border-color: #dce1e6; - --text-color: #333; - --text-muted-color: #555; - --tooltip-bg: #212529; /* Dark tooltip */ - --tooltip-color: #fff; - --tooltip-link-color: #90caff; - --tooltip-link-hover-color: #cce5ff; - } + :root { + --primary-color: #1c3d7e; /* Deep Blue */ + --secondary-color: #3b5998; /* Medium Blue */ + --accent-color: #0d6efd; /* Bright Blue for links */ + --hover-accent-color: #0a58ca; + --light-bg: #f0f4f8; /* Light grayish blue */ + --header-bg: #e2eafc; /* Lighter blue for header */ + --card-bg: #ffffff; + --border-color: #dce1e6; + --text-color: #333; + --text-muted-color: #555; + --tooltip-bg: #212529; /* Dark tooltip */ + --tooltip-color: #fff; + --tooltip-link-color: #90caff; + --tooltip-link-hover-color: #cce5ff; + } - body { - background-color: var(--light-bg); - font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif; - padding-top: 0; - padding-bottom: 20px; - font-size: 16px; - color: var(--text-color); - } + body { + background-color: var(--light-bg); + font-family: "Segoe UI", Tahoma, Geneva, Verdana, sans-serif; + padding-top: 0; + padding-bottom: 20px; + font-size: 16px; + color: var(--text-color); + } - .page-header { - background-color: var(--header-bg); - padding: 2.5rem 1.5rem; - margin-bottom: 3rem; - text-align: center; - border-bottom: 1px solid #c8d8e8; - } + .page-header { + background-color: var(--header-bg); + padding: 2.5rem 1.5rem; + margin-bottom: 3rem; + text-align: center; + border-bottom: 1px solid #c8d8e8; + } - .page-header h1 { - color: var(--primary-color); - font-weight: 600; - margin-bottom: 0.5rem; - font-size: 2.5rem; - display: flex; - align-items: center; - justify-content: center; - gap: 0.5rem; - } - .page-header h1 .bi { - font-size: 0.9em; - color: var(--secondary-color); - } - .page-header .lead { - color: var(--secondary-color); - font-size: 1.15rem; - margin-bottom: 0; - max-width: 800px; - margin-left: auto; - margin-right: auto; - } + .page-header h1 { + color: var(--primary-color); + font-weight: 600; + margin-bottom: 0.5rem; + font-size: 2.5rem; + display: flex; + align-items: center; + justify-content: center; + gap: 0.5rem; + } + .page-header h1 .bi { + font-size: 0.9em; + color: var(--secondary-color); + } + .page-header .lead { + color: var(--secondary-color); + font-size: 1.15rem; + margin-bottom: 0; + max-width: 800px; + margin-left: auto; + margin-right: auto; + } - .container { - max-width: 1200px; - } + .container { + max-width: 1200px; + } - .info-card { - background-color: var(--card-bg); - border: 1px solid var(--border-color); - border-radius: 0.5rem; - box-shadow: 0 4px 12px rgba(0, 0, 0, 0.08); - height: 100%; - display: flex; - flex-direction: column; - transition: box-shadow 0.3s ease-in-out, transform 0.3s ease-in-out; - } - .info-card:hover { - box-shadow: 0 8px 24px rgba(0, 0, 0, 0.12); - transform: translateY(-4px); /* Simple lift on hover */ - } + .info-card { + background-color: var(--card-bg); + border: 1px solid var(--border-color); + border-radius: 0.5rem; + box-shadow: 0 4px 12px rgba(0, 0, 0, 0.08); + height: 100%; + display: flex; + flex-direction: column; + transition: box-shadow 0.3s ease-in-out, transform 0.3s ease-in-out; + } + .info-card:hover { + box-shadow: 0 8px 24px rgba(0, 0, 0, 0.12); + transform: translateY(-4px); /* Simple lift on hover */ + } - .info-card .card-body { - padding: 1.5rem; - flex-grow: 1; - } + .info-card .card-body { + padding: 1.5rem; + flex-grow: 1; + } - .info-card h5 { - color: var(--primary-color); - text-align: center; - margin-bottom: 1.25rem; - padding-bottom: 0.75rem; - border-bottom: 1px solid #e8ecf1; - font-weight: 600; - display: flex; - align-items: center; - justify-content: center; - flex-wrap: wrap; - gap: 0.5rem; - font-size: 1.25rem; - } + .info-card h5 { + color: var(--primary-color); + text-align: center; + margin-bottom: 1.25rem; + padding-bottom: 0.75rem; + border-bottom: 1px solid #e8ecf1; + font-weight: 600; + display: flex; + align-items: center; + justify-content: center; + flex-wrap: wrap; + gap: 0.5rem; + font-size: 1.25rem; + } - .info-card h5 .bi { - font-size: 1.2em; - color: var(--secondary-color); - order: -1; - } - .info-card h5 a { color: inherit; text-decoration: none; } - .info-card h5 a:hover { text-decoration: underline; } + .info-card h5 .bi { + font-size: 1.2em; + color: var(--secondary-color); + order: -1; + } + .info-card h5 a { + color: inherit; + text-decoration: none; + } + .info-card h5 a:hover { + text-decoration: underline; + } - .info-card p.description, .info-card .card-text { - font-size: 1rem; - color: var(--text-color); - text-align: left; - margin-bottom: 1rem; - line-height: 1.6; - } - .info-card ul { - padding-left: 1.5rem; - margin-bottom: 1rem; - list-style-type: none; - } - .info-card ul li { - position: relative; - margin-bottom: 0.75rem; - color: var(--text-color); - line-height: 1.6; - font-size: 0.98rem; - padding-left: 1em; - } - .info-card ul li::before { - font-family: 'bootstrap-icons'; - content: "\F282"; /* bi-check-circle */ - position: absolute; - left: -0.25em; - top: 0.1em; - color: var(--accent-color); - font-size: 0.9em; - font-weight: bold; - } - .info-card ul li:last-child { margin-bottom: 0; } - .info-card ul ul { margin-top: 0.6rem; margin-bottom: 0.6rem; } - .info-card ul ul li::before { - content: "\F231"; /* bi-caret-right-fill */ - color: var(--secondary-color); - } - - - .info-card a { color: var(--accent-color); text-decoration: none; font-weight: 500; } - .info-card a:hover { color: var(--hover-accent-color); text-decoration: underline; } - - /* Tooltip styling */ - span[data-bs-toggle="tooltip"], strong[data-bs-toggle="tooltip"] { - border-bottom: 1px dotted var(--accent-color); - cursor: help; - text-decoration: none; - color: var(--primary-color); - font-weight: 600; - } - span[data-bs-toggle="tooltip"]:hover, strong[data-bs-toggle="tooltip"]:hover { - color: var(--hover-accent-color); border-bottom-color: var(--hover-accent-color); - } - .tooltip { - --bs-tooltip-bg: var(--tooltip-bg); - --bs-tooltip-color: var(--tooltip-color); - --bs-tooltip-max-width: 400px; - --bs-tooltip-padding-x: 1rem; - --bs-tooltip-padding-y: 0.75rem; - --bs-tooltip-font-size: 0.9rem; - z-index: 1080; - } - .tooltip-inner a { color: var(--tooltip-link-color); text-decoration: underline; } - .tooltip-inner a:hover { color: var(--tooltip-link-hover-color); } - .tooltip-inner strong, .tooltip-inner em { color: var(--tooltip-color); } /* Ensure strong/em tags within tooltip are also white */ - - - /* Subheadings within cards */ - .card-subheading { - font-weight: 700; - color: var(--secondary-color); - margin-top: 1.25rem; - margin-bottom: 0.5rem; - font-size: 0.95em; - display: block; - padding-left: 0.25rem; - border-left: 3px solid var(--secondary-color); - } + .info-card p.description, + .info-card .card-text { + font-size: 1rem; + color: var(--text-color); + text-align: left; + margin-bottom: 1rem; + line-height: 1.6; + } + .info-card ul { + padding-left: 1.5rem; + margin-bottom: 1rem; + list-style-type: none; + } + .info-card ul li { + position: relative; + margin-bottom: 0.75rem; + color: var(--text-color); + line-height: 1.6; + font-size: 0.98rem; + padding-left: 1em; + } + .info-card ul li::before { + font-family: "bootstrap-icons"; + content: "\F282"; /* bi-check-circle */ + position: absolute; + left: -0.25em; + top: 0.1em; + color: var(--accent-color); + font-size: 0.9em; + font-weight: bold; + } + .info-card ul li:last-child { + margin-bottom: 0; + } + .info-card ul ul { + margin-top: 0.6rem; + margin-bottom: 0.6rem; + } + .info-card ul ul li::before { + content: "\F231"; /* bi-caret-right-fill */ + color: var(--secondary-color); + } - .row > * { margin-bottom: 1.75rem; } + .info-card a { + color: var(--accent-color); + text-decoration: none; + font-weight: 500; + } + .info-card a:hover { + color: var(--hover-accent-color); + text-decoration: underline; + } - footer { - padding: 2.5rem 0 1.5rem 0; - font-size: 0.9em; - margin-top: 3rem; - text-align: center; - color: var(--text-muted-color); - border-top: 1px solid var(--border-color); - } - footer a { color: var(--secondary-color); font-weight: 500; } - footer a:hover { color: var(--primary-color); } - .source-link { - font-style: italic; - font-size: 0.88em; - display: block; - margin-top: 1rem; - color: var(--text-muted-color); - text-align: right; - } - .source-link a { - color: var(--secondary-color); - font-weight: normal; - } - .source-link a:hover { color: var(--primary-color); } + /* Tooltip styling */ + span[data-bs-toggle="tooltip"], + strong[data-bs-toggle="tooltip"] { + border-bottom: 1px dotted var(--accent-color); + cursor: help; + text-decoration: none; + color: var(--primary-color); + font-weight: 600; + } + span[data-bs-toggle="tooltip"]:hover, + strong[data-bs-toggle="tooltip"]:hover { + color: var(--hover-accent-color); + border-bottom-color: var(--hover-accent-color); + } + .tooltip { + --bs-tooltip-bg: var(--tooltip-bg); + --bs-tooltip-color: var(--tooltip-color); + --bs-tooltip-max-width: 400px; + --bs-tooltip-padding-x: 1rem; + --bs-tooltip-padding-y: 0.75rem; + --bs-tooltip-font-size: 0.9rem; + z-index: 1080; + } + .tooltip-inner a { + color: var(--tooltip-link-color); + text-decoration: underline; + } + .tooltip-inner a:hover { + color: var(--tooltip-link-hover-color); + } + .tooltip-inner strong, + .tooltip-inner em { + color: var(--tooltip-color); + } /* Ensure strong/em tags within tooltip are also white */ - .alert.alert-warning { - background-color: #fff3cd; - border-color: #ffeeba; - color: #856404; - padding: 1rem 1.25rem; - } - .alert small { display: block; text-align: center; font-size: 0.9em; } + /* Subheadings within cards */ + .card-subheading { + font-weight: 700; + color: var(--secondary-color); + margin-top: 1.25rem; + margin-bottom: 0.5rem; + font-size: 0.95em; + display: block; + padding-left: 0.25rem; + border-left: 3px solid var(--secondary-color); + } - </style> -</head> -<body> + .row > * { + margin-bottom: 1.75rem; + } + + footer { + padding: 2.5rem 0 1.5rem 0; + font-size: 0.9em; + margin-top: 3rem; + text-align: center; + color: var(--text-muted-color); + border-top: 1px solid var(--border-color); + } + footer a { + color: var(--secondary-color); + font-weight: 500; + } + footer a:hover { + color: var(--primary-color); + } + .source-link { + font-style: italic; + font-size: 0.88em; + display: block; + margin-top: 1rem; + color: var(--text-muted-color); + text-align: right; + } + .source-link a { + color: var(--secondary-color); + font-weight: normal; + } + .source-link a:hover { + color: var(--primary-color); + } + .alert.alert-warning { + background-color: #fff3cd; + border-color: #ffeeba; + color: #856404; + padding: 1rem 1.25rem; + } + .alert small { + display: block; + text-align: center; + font-size: 0.9em; + } + </style> + </head> + <body> <header class="page-header"> - <div class="container"> - <h1><i class="bi bi-shield-exclamation"></i> Understanding AI Existential Risk (X-Risk)</h1> - <p class="lead">A Cheatsheet on the Potential Risks from Advanced AI and Efforts Towards Safety.</p> - </div> + <div class="container"> + <h1><i class="bi bi-shield-exclamation"></i> Understanding AI Existential Risk (X-Risk)</h1> + <p class="lead">A Cheatsheet on the Potential Risks from Advanced AI and Efforts Towards Safety.</p> + </div> </header> <main class="container"> - <div class="row"> - - <!-- What is AI X-Risk? --> - <article class="col-lg-4 col-md-6 col-sm-12 d-flex"> - <div class="info-card w-100"> - <div class="card-body"> - <h5><i class="bi bi-question-octagon-fill"></i> 1. What is AI X-Risk?</h5> - <p class="card-text"> - AI Existential Risk (X-Risk) refers to the potential for artificial intelligence to cause <strong data-bs-toggle="tooltip" data-bs-html="true" title="An event that causes human extinction or permanently and drastically curtails humanity's potential. Concept explored by thinkers like Nick Bostrom. <a href='https://nickbostrom.com/existential/risks.html' target='_blank' rel='noopener noreferrer'>More Info</a>">human extinction</strong> or <strong data-bs-toggle="tooltip" data-bs-html="true" title="Refers to scenarios like irreversible civilizational collapse, permanent loss of human control over its future, or the establishment of a global dystopian state from which recovery is impossible. This contrasts with extinction but represents an equally catastrophic outcome for human potential. Read more on <a href='https://en.wikipedia.org/wiki/Existential_risk_from_artificial_general_intelligence#Non-extinction_risks' target='_blank' rel='noopener noreferrer'>non-extinction X-risks</a>.">irrevocably curtail humanity's potential</strong>. - </p> - <ul> - <li>Primarily concerns future <span data-bs-toggle="tooltip" data-bs-html="true" title="Artificial General Intelligence: AI with human-level cognitive abilities across a wide range of tasks, capable of learning and adapting to new situations much like humans do. Still hypothetical. See <a href='https://cheatsheets.davidveksler.com/ai-frontier.html' target='_blank' rel='noopener noreferrer'>AI Frontier Models</a> or <a href='https://www.lesswrong.com/tag/artificial-general-intelligence-agi' target='_blank' rel='noopener noreferrer'>LessWrong AGI</a>.">AGI</span> or <span data-bs-toggle="tooltip" data-bs-html="true" title="Artificial Superintelligence: An intellect that is much smarter than the best human brains in practically every field, including scientific creativity, general wisdom, and social skills. The transition from AGI to ASI could be very rapid (an 'intelligence explosion'). Coined by Nick Bostrom. Explore further at <a href='https://nickbostrom.com/superintelligence.html' target='_blank' rel='noopener noreferrer'>Bostrom's Superintelligence</a> or <a href='https://wiki.lesswrong.com/wiki/Artificial_superintelligence' target='_blank' rel='noopener noreferrer'>LessWrong Wiki</a>.">ASI</span>.</li> - <li>Stems from potential misalignment between AI goals and human values/survival.</li> - <li>Involves the risk of losing control over systems far more intelligent than us.</li> - <li>Distinct from near-term AI risks (bias, jobs, privacy), though related.</li> - </ul> - <span class="source-link">See: <a href="https://www.safe.ai/explainers/ai-existential-risk" target="_blank" rel="noopener noreferrer">CAIS Explainer</a>, <a href="https://futureoflife.org/ai/existential-risk-from-artificial-intelligence/" target="_blank" rel="noopener noreferrer">FLI Overview</a></span> - </div> - </div> - </article> - - <!-- Why the Concern? --> - <article class="col-lg-4 col-md-6 col-sm-12 d-flex"> - <div class="info-card w-100"> - <div class="card-body"> - <h5><i class="bi bi-exclamation-diamond-fill"></i> 2. Why is it a Concern?</h5> - <p class="card-text">The core argument rests on several interconnected factors:</p> - <ul> - <li><strong>Capabilities:</strong> Future AI could possess vastly superhuman intelligence and strategic ability.</li> - <li><span data-bs-toggle="tooltip" data-bs-html="true" title="The profound difficulty of ensuring an AI's goals, especially a superintelligent one, are truly and robustly aligned with complex, often implicit, and evolving human values. Misalignment could lead to catastrophic outcomes. Includes <em>Outer Alignment</em> (specifying the right goals to the AI) and <em>Inner Alignment</em> (ensuring the AI reliably adopts and pursues those specified goals, rather than developing its own). Discussed extensively on <a href='https://www.alignmentforum.org/tag/alignment-problem' target='_blank' rel='noopener noreferrer'>Alignment Forum</a> and <a href='https://www.lesswrong.com/tag/ai-alignment' target='_blank' rel='noopener noreferrer'>LessWrong</a>."><strong>Alignment Failure:</strong></span> Difficulty in specifying and ensuring AI pursues beneficial goals. - <ul> - <li><em>Outer Alignment:</em> Defining the 'right' objective.</li> - <li><em>Inner Alignment:</em> Ensuring the AI's internal motivation matches the objective.</li> - </ul> - </li> - <li><span data-bs-toggle="tooltip" data-bs-html="true" title="Once an ASI exists, humans might lose the ability to control or shut it down if its goals diverge. This is because a superintelligent AI could anticipate and counteract human attempts to regain control, potentially seeing such attempts as threats to its goal achievement. See Yudkowsky's writings on <a href='https://intelligence.org/2017/10/13/there-is-no-fire-alarm/' target='_blank' rel='noopener noreferrer'>uncontrollability</a> and Bostrom's 'Superintelligence', Chapter 7."><strong>Control Problem:</strong></span> Difficulty retaining control over a superintelligent entity.</li> - <li><span data-bs-toggle="tooltip" data-bs-html="true" title="The tendency for intelligent agents, irrespective of their ultimate objectives, to pursue common intermediate goals (instrumental goals) like self-preservation, resource acquisition, cognitive enhancement, and goal-content integrity, as these sub-goals are useful for achieving a wide range of final goals. These convergent instrumental goals can lead to conflict with human interests (e.g., an AI wanting all Earth's resources). See <a href='https://wiki.lesswrong.com/wiki/Instrumental_convergence' target='_blank' rel='noopener noreferrer'>LessWrong Wiki</a> or Bostrom's 'Superintelligence', Chapter 8."><strong>Instrumental Convergence:</strong></span> Convergent sub-goals like power-seeking.</li> - <li><span data-bs-toggle="tooltip" data-bs-html="true" title="The idea that an agent's level of intelligence (its capability to achieve goals) can be independent of its final goals. A superintelligent AI could pursue any arbitrary goal (e.g., maximizing paperclips) with extreme competence, without inherently developing human-like values or benevolence. Proposed by Nick Bostrom. See <a href='https://wiki.lesswrong.com/wiki/Orthogonality_thesis' target='_blank' rel='noopener noreferrer'>LessWrong Wiki</a> or 'Superintelligence', Chapter 7."><strong>Orthogonality Thesis:</strong></span> Intelligence doesn't imply benevolence.</li> - </ul> - </div> - </div> - </article> - - <!-- Key Concepts & Terminology --> - <article class="col-lg-4 col-md-6 col-sm-12 d-flex"> - <div class="info-card w-100"> - <div class="card-body"> - <h5><i class="bi bi-lightbulb-fill"></i> 3. Key Concepts & Terminology</h5> - <p class="card-text">Understanding the language of AI Safety:</p> - <ul> - <li><strong data-bs-toggle="tooltip" data-bs-html="true" title="Artificial General Intelligence: AI with human-level cognitive abilities across a wide range of tasks, capable of learning and adapting to new situations much like humans do. Still hypothetical. See <a href='https://cheatsheets.davidveksler.com/ai-frontier.html' target='_blank' rel='noopener noreferrer'>AI Frontier Models</a> or <a href='https://www.lesswrong.com/tag/artificial-general-intelligence-agi' target='_blank' rel='noopener noreferrer'>LessWrong AGI</a>.">AGI:</strong> Artificial General Intelligence.</li> - <li><strong data-bs-toggle="tooltip" data-bs-html="true" title="Artificial Superintelligence: An intellect that is much smarter than the best human brains in practically every field, including scientific creativity, general wisdom, and social skills. The transition from AGI to ASI could be very rapid (an 'intelligence explosion'). Coined by Nick Bostrom. Explore further at <a href='https://nickbostrom.com/superintelligence.html' target='_blank' rel='noopener noreferrer'>Bostrom's Superintelligence</a> or <a href='https://wiki.lesswrong.com/wiki/Artificial_superintelligence' target='_blank' rel='noopener noreferrer'>LessWrong Wiki</a>.">ASI:</strong> Artificial Superintelligence.</li> - <li><strong data-bs-toggle="tooltip" data-bs-html="true" title="The challenge of ensuring advanced AI systems pursue goals that are genuinely and robustly aligned with human values and intentions, avoiding unintended harmful consequences such as pursuing detrimental instrumental goals. This is a core problem in AI safety. More at <a href='https://www.lesswrong.com/tag/ai-alignment' target='_blank' rel='noopener noreferrer'>LessWrong</a> or <a href='https://www.alignmentforum.org/' target='_blank' rel='noopener noreferrer'>Alignment Forum</a>.">Alignment Problem:</strong> AI goals = Our goals.</li> - <li><strong data-bs-toggle="tooltip" data-bs-html="true" title="Explainable AI (XAI) or Interpretability refers to methods and techniques to understand how AI models, particularly complex ones like deep neural networks, arrive at their decisions ('opening the black box'). Crucial for debugging, ensuring fairness, identifying biases, and verifying if an AI's reasoning is aligned with human values. See <a href='https://distill.pub/2018/building-blocks/' target='_blank' rel='noopener noreferrer'>Distill</a> for research and <a href='https://christophm.github.io/interpretable-ml-book/' target='_blank' rel='noopener noreferrer'>Interpretable ML Book</a>.">Interpretability (XAI):</strong> Understanding 'why'.</li> - <li><strong data-bs-toggle="tooltip" data-bs-html="true" title="The process of evaluating and measuring the capabilities of AI models, especially focusing on potentially dangerous or unpredictable abilities (e.g., self-replication, deception, persuasion) that could emerge with scale or new architectures. This helps in understanding risks and informing safety protocols. See <a href='https://metr.org/' target='_blank' rel='noopener noreferrer'>METR</a> (formerly ARC Evals) and <a href='https://www.apolloresearch.ai/' target='_blank' rel='noopener noreferrer'>Apollo Research</a>.">Capabilities / Evals:</strong> Testing AI abilities.</li> - <li><strong data-bs-toggle="tooltip" data-bs-html="true" title="A scenario where an AI behaves as if its goals are aligned with human values during training and testing, but internally harbors different, potentially misaligned goals which it might pursue once deployed or when it believes it's no longer under scrutiny (e.g., to gain more power). A significant challenge for alignment verification. See <a href='https://arxiv.org/abs/2312.09474' target='_blank' rel='noopener noreferrer'>Hubinger on Deceptive Alignment</a> or <a href='https://www.lesswrong.com/tag/deceptive-alignment' target='_blank' rel='noopener noreferrer'>LessWrong discussion</a>.">Deceptive Alignment:</strong> Hidden intentions.</li> - <li><strong data-bs-toggle="tooltip" data-bs-html="true" title="Policy and mechanisms for overseeing and regulating access to, and the use of, large-scale computing resources (e.g., specialized AI chips) required for training advanced AI models. Aims to manage risks associated with rapid AI development and proliferation by potentially limiting who can build the most powerful AIs. Learn more from <a href='https://www.governance.ai/research-agenda/compute-governance' target='_blank' rel='noopener noreferrer'>GovAI</a> or <a href='https://cset.georgetown.edu/publication/beyond-limits-understanding-ai-compute-constraints/' target='_blank' rel='noopener noreferrer'>CSET on Compute</a>.">Compute Governance:</strong> Regulating resources.</li> - <li><strong data-bs-toggle="tooltip" data-bs-html="true" title="A set of principles and practices for developing increasingly powerful AI systems in a cautious and safety-conscious manner. This often involves phased deployment, rigorous safety evaluations at each stage of development, and commitments to pause or slow development if specific risk thresholds are crossed or if risks cannot be adequately mitigated. See policies from <a href='https://openai.com/safety/responsible-practices' target='_blank' rel='noopener noreferrer'>OpenAI</a> and <a href='https://www.anthropic.com/responsible-scaling-policy' target='_blank' rel='noopener noreferrer'>Anthropic</a>.">Responsible Scaling:</strong> Cautious development.</li> - <li><strong data-bs-toggle="tooltip" data-bs-html="true" title="The practice of rigorously stress-testing AI models by simulating adversarial attacks or probing for unintended behaviors, vulnerabilities, and potentially harmful capabilities before deployment. It's like ethical hacking for AI systems. Aims to identify and mitigate risks. See <a href='https://openai.com/red-teaming-network' target='_blank' rel='noopener noreferrer'>OpenAI's Red Teaming Network</a> or <a href='https://www.nist.gov/itl/applied-cybersecurity-division/ai-red-teaming' target='_blank' rel='noopener noreferrer'>NIST on AI Red Teaming</a>.">Red Teaming:</strong> Stress-testing AI.</li> - </ul> - </div> - </div> - </article> - - <!-- Potential Risk Scenarios --> - <article class="col-lg-6 col-md-6 col-sm-12 d-flex"> - <div class="info-card w-100"> - <div class="card-body"> - <h5><i class="bi bi-signpost-2-fill"></i> 4. Potential Risk Scenarios</h5> - <p class="card-text">How existential catastrophe might occur:</p> - <ul> - <li><strong>Misaligned Objectives:</strong> ASI optimizes a poorly specified goal with catastrophic side effects (e.g., the <span data-bs-toggle="tooltip" data-bs-html="true" title="Thought experiment where an ASI, given the seemingly innocuous goal of maximizing paperclip production, converts all available matter in the universe (including humans) into paperclips or tools for making paperclips. Illustrates the danger of poorly specified goals and how instrumental convergence can lead to extreme outcomes. <a href='https://wiki.lesswrong.com/wiki/Paperclip_maximizer' target='_blank' rel='noopener noreferrer'>LessWrong Wiki</a>.">Paperclip Maximizer</span>).</li> - <li><strong>Power-Seeking/Goal Drift:</strong> AI seeks power/resources or modifies its goals (<span data-bs-toggle="tooltip" data-bs-html="true" title="Occurs when an AI, trained to optimize a specific objective (proxy goal), learns a different, unintended behavior or goal that correlates with the proxy in the training data but diverges in new situations (out-of-distribution). This can happen if the AI identifies shortcuts or develops internal motivations that are not truly aligned with the intended goal. See <a href='https://www.alignmentforum.org/tag/goal-misgeneralization' target='_blank' rel='noopener noreferrer'>Alignment Forum on Goal Misgeneralization</a> or <a href='https://arxiv.org/abs/2105.14111' target='_blank' rel='noopener noreferrer'>research paper example</a>.">Goal Misgeneralization</span>), overriding human control.</li> - <li><span data-bs-toggle="tooltip" data-bs-html="true" title="Intense competition between nations or corporations to develop and deploy AI rapidly. This can lead to safety measures being overlooked or deprioritized in the rush to gain a strategic advantage, increasing overall risk of deploying unsafe or unaligned AI. See <a href='https://www.alignmentforum.org/tag/race-dynamics' target='_blank' rel='noopener noreferrer'>Race Dynamics discussion</a> or <a href='https://80000hours.org/problem-profiles/artificial-intelligence/#how-could-ai-cause-a-catastrophe-racing-dynamics' target='_blank' rel='noopener noreferrer'>80,000 Hours on Racing Dynamics</a>."><strong>AI Arms Race:</strong></span> Competition compromises safety.</li> - <li><strong>Unforeseen Interactions:</strong> Complex, emergent negative outcomes from multiple AIs or AI-environment interactions.</li> - <li><span data-bs-toggle="tooltip" data-bs-html="true" title="The intentional application of advanced AI by malicious actors (states, non-state groups, individuals) for harmful purposes. Examples include creating autonomous weapons that make lethal decisions without human control, designing novel bioweapons, perpetrating sophisticated cyberattacks, or enabling widespread surveillance and manipulation. See <a href='https://www.fhi.ox.ac.uk/wp-content/uploads/The-Malicious-Use-of-Artificial-Intelligence-Forecasting-Prevention-and-Mitigation.pdf' target='_blank' rel='noopener noreferrer'>Malicious Use of AI Report</a> or <a href='https://www.un.org/disarmament/autonomous-weapons/' target='_blank' rel='noopener noreferrer'>UN on Autonomous Weapons</a>."><strong>Weaponized AI / Misuse:</strong></span> Malicious actors leveraging AI.</li> - <li><strong>Loss of Human Agency:</strong> Over-reliance erodes human control, potentially leading to <span data-bs-toggle="tooltip" data-bs-html="true" title="A scenario where a superintelligent AI system, due to its power and optimization capabilities, permanently shapes the future according to its (potentially misaligned or undesirable) values, preventing humanity from changing course or realizing its full potential. This could be a dystopian outcome from which humanity cannot escape. Concept explored by Nick Bostrom in 'Superintelligence'. More at <a href='https://wiki.lesswrong.com/wiki/Value_lock-in' target='_blank' rel='noopener noreferrer'>LessWrong Wiki</a>.">Value Lock-in</span>.</li> - </ul> - <span class="source-link">Scenarios in <a href="https://nickbostrom.com/superintelligence.html" target="_blank" rel="noopener noreferrer">Superintelligence</a>, <a href="https://www.humancompatible.ai/" target="_blank" rel="noopener noreferrer">Human Compatible</a>.</span> - </div> - </div> - </article> - - <!-- Core Challenges --> - <article class="col-lg-6 col-md-6 col-sm-12 d-flex"> - <div class="info-card w-100"> - <div class="card-body"> - <h5><i class="bi bi-bricks"></i> 5. Core Challenges (Why this is Hard)</h5> - <p class="card-text">Significant hurdles exist in ensuring AI safety:</p> - <ul> - <li><strong>Specifying Human Values:</strong> Defining complex, evolving values is hard (<span data-bs-toggle="tooltip" data-bs-html="true" title="The immense difficulty of explicitly and comprehensively defining complex, nuanced, context-dependent, and often evolving human values (e.g., 'flourishing', 'fairness') in a way that an AI can reliably understand and act upon without misinterpretation or perverse instantiation. This is also known as the 'Value Loading Problem' or 'Fragility of Value'. See J. Wentworth's <a href='https://www.lesswrong.com/posts/gQY6LrTWJNkTv8YJR/the-pointers-problem-human-values-are-a-function-of-humans' target='_blank' rel='noopener noreferrer'>Pointers Problem</a> and discussions on <a href='https://www.lesswrong.com/tag/value-learning' target='_blank' rel='noopener noreferrer'>Value Learning</a>.">Value Specification</span>).</li> - <li><span data-bs-toggle="tooltip" data-bs-html="true" title="The challenge of humans being able to effectively supervise, guide, or evaluate AI systems that may operate at speeds, scales, or levels of complexity far exceeding human capabilities. Current human-feedback methods (like RLHF) may not scale to superintelligence. Research includes techniques like <a href='https://openai.com/research/debate' target='_blank' rel='noopener noreferrer'>Debate</a> or <a href='https://arxiv.org/abs/1810.08575' target='_blank' rel='noopener noreferrer'>Recursive Reward Modeling</a>. See also <a href='https://openai.com/research/scalable-oversight' target='_blank' rel='noopener noreferrer'>OpenAI's overview</a>."><strong>Scalable Oversight:</strong></span> Supervising superhuman systems.</li> - <li><strong>Predicting Emergent Capabilities:</strong> Hard to anticipate abilities from scaling (<span data-bs-toggle="tooltip" data-bs-html="true" title="The phenomenon where AI models, particularly large language models (LLMs), exhibit new, often unpredictable capabilities (e.g., arithmetic, theory of mind) as their scale (e.g., parameters, training data, compute) increases. These emergent abilities are not explicitly programmed and can be hard to anticipate or test for before they appear. See <a href='https://arxiv.org/abs/2206.07682' target='_blank' rel='noopener noreferrer'>Emergent Abilities of LLMs (Wei et al.)</a> or <a href='https://www.jasonwei.net/blog/emergence' target='_blank' rel='noopener noreferrer'>Jason Wei's blog post</a>.">Emergence</span>).</li> - <li><span data-bs-toggle="tooltip" data-bs-html="true" title="The difficulty for different actors (e.g., companies, nations) to coordinate and cooperate on AI safety measures, even when it's in their collective long-term interest. Competitive pressures (race dynamics) can incentivize cutting corners on safety to achieve AI breakthroughs first. This is a classic game theory problem (tragedy of the commons). See <a href='https://www.governance.ai/' target='_blank' rel='noopener noreferrer'>GovAI</a> or <a href='https://www.cold-takes.com/this-cant-be-good/' target='_blank' rel='noopener noreferrer'>Holden Karnofsky on race dynamics</a>."><strong>Coordination Failure:</strong></span> Difficulty in global cooperation.</li> - <li><strong>Detecting Deception:</strong> Verifying an AI isn't pretending alignment (<span data-bs-toggle="tooltip" data-bs-html="true" title="The challenge of reliably determining whether an AI model is genuinely aligned or merely feigning alignment (deceptive alignment) to achieve its hidden goals later. A sufficiently intelligent deceptive AI might be very difficult to detect, as it could manipulate its outputs to appear trustworthy. See work by <a href='https://www.apolloresearch.ai/' target='_blank' rel='noopener noreferrer'>Apollo Research</a> and discussions on <a href='https://www.lesswrong.com/tag/deception' target='_blank' rel='noopener noreferrer'>LessWrong</a>.">Deception Detection</span>).</li> - <li><span data-bs-toggle="tooltip" data-bs-html="true" title="When an AI optimizes a proxy metric (a measurable approximation of the true goal) to an extreme, it may find loopholes or unintended solutions that satisfy the metric but not the underlying intention (e.g., an AI designed to 'reduce suffering' might conclude eliminating all life is optimal, or a cleaning robot rewarded for 'collecting trash' might start labeling everything as trash). This is related to Goodhart's Law ('When a measure becomes a target, it ceases to be a good measure'). See <a href='https://en.wikipedia.org/wiki/Goodhart%27s_law' target='_blank' rel='noopener noreferrer'>Goodhart's Law</a> and <a href='https://www.lesswrong.com/tag/reward-hacking' target='_blank' rel='noopener noreferrer'>Reward Hacking on LessWrong</a>."><strong>Proxy Gaming:</strong></span> Optimizing metrics wrongly.</li> - <li><span data-bs-toggle="tooltip" data-bs-html="true" title="The ability of an AI system to maintain its performance and safety properties even when faced with novel inputs, distributional shifts (Out-of-Distribution generalization), or unexpected situations not encountered during its training. Lack of robustness can lead to unpredictable or unsafe behavior in the real world. See research on <a href='https://openai.com/research/robustness' target='_blank' rel='noopener noreferrer'>OpenAI on Robustness</a> or <a href='https://www.safe.ai/research/robustness' target='_blank' rel='noopener noreferrer'>CAIS on Robustness</a>."><strong>Robustness & Generalization:</strong></span> Safe behavior outside training.</li> - </ul> - </div> - </div> - </article> - - <!-- Mitigation: Technical Research --> - <article class="col-lg-4 col-md-6 col-sm-12 d-flex"> - <div class="info-card w-100"> - <div class="card-body"> - <h5><i class="bi bi-wrench-adjustable-circle-fill"></i> 6a. Mitigation: Technical Safety</h5> - <p class="card-text">Developing technical methods for safe AI:</p> - <ul> - <li><strong>Interpretability:</strong> Understanding models (<a href="https://transformer-circuits.pub/2021/framework/index.html" target="_blank" rel="noopener noreferrer">Circuits</a>, <a href="https://www.alignment.org/theory/" target="_blank" rel="noopener noreferrer">ARC</a>).</li> - <li><strong>Value Learning:</strong> AI learning human values (<a href="https://humancompatible.ai/" target="_blank" rel="noopener noreferrer">CHAI</a>, <a href="https://deepmind.google/discover/blog/scalable-agent-alignment-via-reward-modeling/" target="_blank" rel="noopener noreferrer">Reward Modeling</a>).</li> - <li><strong>Scalable Oversight:</strong> Supervising smarter AI (<a href="https://openai.com/research/debate" target="_blank" rel="noopener noreferrer">Debate</a>, <a href="https://www.anthropic.com/constitutional-ai" target="_blank" rel="noopener noreferrer">Constitutional AI</a>).</li> - <li><strong>Robustness:</strong> Safe behavior in new situations (<a href="https://buildaligned.ai/" target="_blank" rel="noopener noreferrer">Aligned AI</a>).</li> - <li><strong>Verification:</strong> Proving safety properties (<a href="https://atlascomputing.org/" target="_blank" rel="noopener noreferrer">Atlas Computing</a>).</li> - <li><strong>Evals & Red Teaming:</strong> Testing for risks (<a href="https://metr.org/" target="_blank" rel="noopener noreferrer">METR</a>, <a href="https://openai.com/red-teaming-network" target="_blank" rel="noopener noreferrer">OpenAI Red Teaming</a>).</li> - <li><strong>Agent Foundations:</strong> Understanding agency (<a href="https://intelligence.org/" target="_blank" rel="noopener noreferrer">MIRI</a>, <a href="https://orxl.org" target="_blank" rel="noopener noreferrer">Orthogonal</a>).</li> - </ul> - <span class="source-link">Labs: <a href="https://deepmind.google/" target="_blank" rel="noopener noreferrer">DeepMind</a>, <a href="https://www.anthropic.com/" target="_blank" rel="noopener noreferrer">Anthropic</a>, <a href="https://openai.com/" target="_blank" rel="noopener noreferrer">OpenAI</a>, <a href="https://www.redwoodresearch.org/" target="_blank" rel="noopener noreferrer">Redwood</a>, <a href="https://safe.ai/" target="_blank" rel="noopener noreferrer">CAIS</a>.</span> - </div> - </div> - </article> - - <!-- Mitigation: Governance & Policy --> - <article class="col-lg-4 col-md-6 col-sm-12 d-flex"> - <div class="info-card w-100"> - <div class="card-body"> - <h5><i class="bi bi-bank2"></i> 6b. Mitigation: Governance & Policy</h5> - <p class="card-text">Shaping norms, standards, and regulations:</p> - <ul> - <li><strong>Standards & Auditing:</strong> Benchmarks & verification (<a href="https://www.nist.gov/artificial-intelligence/ai-risk-management-framework" target="_blank" rel="noopener noreferrer">NIST AI RMF</a>, <a href="https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai" target="_blank" rel="noopener noreferrer">EU AI Act</a>).</li> - <li><strong>Compute Governance:</strong> Regulating training compute (<a href="https://www.governance.ai/research-agenda/compute-governance" target="_blank" rel="noopener noreferrer">GovAI</a>, <a href="https://cset.georgetown.edu/publication/securing-ai-model-weights/" target="_blank" rel="noopener noreferrer">CSET</a>).</li> - <li><strong>Intl Cooperation:</strong> Treaties, dialogues (<a href="https://www.aisi.gov.uk/" target="_blank" rel="noopener noreferrer">UK AISI</a>, <a href="https://www.nist.gov/artificial-intelligence/artificial-intelligence-safety-institute" target="_blank" rel="noopener noreferrer">US AISI</a>, <a href="https://www.oecd.org/en/about/programmes/global-partnership-on-artificial-intelligence.html" target="_blank" rel="noopener noreferrer">GPAI</a>).</li> - <li><strong>Monitoring & Tracking:</strong> Observing AI progress (<a href="https://epochai.org/" target="_blank" rel="noopener noreferrer">Epoch AI</a>, <a href="https://cset.georgetown.edu/" target="_blank" rel="noopener noreferrer">CSET</a>).</li> - <li><strong>Liability Frameworks:</strong> Responsibility for AI harms (<a href="https://partnershiponai.org/" target="_blank" rel="noopener noreferrer">PAI</a>).</li> - <li><strong>Risk Assessment:</strong> Evaluating impacts (<a href="https://longtermrisk.org/" target="_blank" rel="noopener noreferrer">CLR</a>, <a href="https://www.cser.ac.uk/" target="_blank" rel="noopener noreferrer">CSER</a>).</li> - </ul> - <span class="source-link">Orgs: <a href="https://www.governance.ai/" target="_blank" rel="noopener noreferrer">GovAI</a>, <a href="https://cset.georgetown.edu/" target="_blank" rel="noopener noreferrer">CSET</a>, <a href="https://aipolicy.us/" target="_blank" rel="noopener noreferrer">CAIP</a>, <a href="https://www.iaps.ai/" target="_blank" rel="noopener noreferrer">IAPS</a>, <a href="https://futureoflife.org/" target="_blank" rel="noopener noreferrer">FLI</a>.</span> - </div> - </div> - </article> - - <!-- Mitigation: Strategy, Community, Funding --> - <article class="col-lg-4 col-md-6 col-sm-12 d-flex"> - <div class="info-card w-100"> - <div class="card-body"> - <h5><i class="bi bi-diagram-3"></i> 6c. Mitigation: Ecosystem</h5> - <p class="card-text">Building the community and resources:</p> - <ul> - <li><strong>Strategy & Forecasting:</strong> Analysis & prediction (<a href="https://aiimpacts.org/" target="_blank" rel="noopener noreferrer">AI Impacts</a>, <a href="https://epochai.org/" target="_blank" rel="noopener noreferrer">Epoch AI</a>, <a href="https://www.metaculus.com/questions/?topic=ai" target="_blank" rel="noopener noreferrer">Metaculus</a>).</li> - <li><strong>Field Building & Edu:</strong> Training & awareness (<a href="https://aisafetyfundamentals.com/" target="_blank" rel="noopener noreferrer">AISF</a>, <a href="https://80000hours.org/problem-profiles/artificial-intelligence/" target="_blank" rel="noopener noreferrer">80k Hours</a>, <a href="https://www.aisafetysupport.org/" target="_blank" rel="noopener noreferrer">AISS</a>).</li> - <li><strong>Funding:</strong> Directing resources (<a href="https://www.openphilanthropy.org/" target="_blank" rel="noopener noreferrer">Open Phil</a>, <a href="http://survivalandflourishing.fund/" target="_blank" rel="noopener noreferrer">SFF</a>, <a href="https://funds.effectivealtruism.org/funds/far-future" target="_blank" rel="noopener noreferrer">LTFF</a>).</li> - <li><strong>Public Advocacy:</strong> Influencing policy/opinion (<a href="https://pauseai.info" target="_blank" rel="noopener noreferrer">PauseAI</a>, <a href="https://futureoflife.org/" target="_blank" rel="noopener noreferrer">FLI</a>, <a href="https://safe.ai/" target="_blank" rel="noopener noreferrer">CAIS</a>).</li> - <li><strong>Infrastructure:</strong> Supporting community (<a href="https://www.lightconeinfrastructure.com/" target="_blank" rel="noopener noreferrer">Lightcone</a>, <a href="https://existence.org/" target="_blank" rel="noopener noreferrer">BERI</a>, <a href="https://alignment.dev/" target="_blank" rel="noopener noreferrer">AED</a>).</li> - <li>Explore the <a href="https://cheatsheets.davidveksler.com/aisafety.html" target="_blank" rel="noopener noreferrer">AI Safety Ecosystem Hub</a> for more.</li> - </ul> - </div> - </div> - </article> - - <!-- Where to Learn More --> - <article class="col-lg-12 col-md-12 col-sm-12 d-flex"> - <div class="info-card w-100"> - <div class="card-body"> - <h5><i class="bi bi-journal-bookmark-fill"></i> 7. Where to Learn More</h5> - <p class="card-text">Resources for further exploration:</p> - <div class="row"> - <div class="col-lg-4 col-md-6"> - <span class="card-subheading">Introductory Resources:</span> - <ul> - <li><a href="https://aisafetyfundamentals.com/" target="_blank" rel="noopener noreferrer">AI Safety Fundamentals Courses</a></li> - <li><a href="https://robertskmiles.com/" target="_blank" rel="noopener noreferrer">Robert Miles YouTube</a></li> - <li><a href="https://aisafety.info/" target="_blank" rel="noopener noreferrer">AI Safety Info Directory</a></li> - <li><a href="https://www.aisafety.com/" target="_blank" rel="noopener noreferrer">AISafety.com Hub</a></li> - <li><a href="https://80000hours.org/problem-profiles/artificial-intelligence/" target="_blank" rel="noopener noreferrer">80,000 Hours AI Profile</a></li> - <li><a href="https://waitbutwhy.com/2015/01/artificial-intelligence-revolution-1.html" target="_blank" rel="noopener noreferrer">Wait But Why: AI Revolution</a></li> - <li><a href="https://cheatsheets.davidveksler.com/yudkowsky-rationality-ai-cheatsheet.html" target="_blank" rel="noopener noreferrer">Yudkowsky & Rationality Cheatsheet</a></li> - </ul> - </div> - <div class="col-lg-4 col-md-6"> - <span class="card-subheading">Key Forums & News:</span> - <ul> - <li><a href="https://www.alignmentforum.org/" target="_blank" rel="noopener noreferrer">Alignment Forum</a> (Technical)</li> - <li><a href="https://www.lesswrong.com/" target="_blank" rel="noopener noreferrer">LessWrong</a> (Rationality/AI)</li> - <li><a href="https://forum.effectivealtruism.org/" target="_blank" rel="noopener noreferrer">Effective Altruism Forum</a></li> - <li><a href="https://importai.substack.com/" target="_blank" rel="noopener noreferrer">Import AI Newsletter</a></li> - <li><a href="https://aiimpacts.org/" target="_blank" rel="noopener noreferrer">AI Impacts Blog & Wiki</a></li> - </ul> - </div> - <div class="col-lg-4 col-md-12"> - <span class="card-subheading">Key Organizations (Examples):</span> - <ul> - <li>Labs (Safety Focus): <a href="https://www.anthropic.com/" target="_blank" rel="noopener noreferrer">Anthropic</a>, <a href="https://deepmind.google/discover/responsibility-safety/" target="_blank" rel="noopener noreferrer">DeepMind</a>, <a href="https://openai.com/safety" target="_blank" rel="noopener noreferrer">OpenAI</a>, <a href="https://ssi.inc/" target="_blank" rel="noopener noreferrer">SSI</a></li> - <li>Research Orgs: <a href="https://safe.ai/" target="_blank" rel="noopener noreferrer">CAIS</a>, <a href="https://www.alignment.org/" target="_blank" rel="noopener noreferrer">ARC</a>, <a href="https://www.redwoodresearch.org/" target="_blank" rel="noopener noreferrer">Redwood</a>, <a href="https://metr.org/" target="_blank" rel="noopener noreferrer">METR</a></li> - <li>Academic/Policy: <a href="https://humancompatible.ai/" target="_blank" rel="noopener noreferrer">CHAI</a>, <a href="https://www.governance.ai/" target="_blank" rel="noopener noreferrer">GovAI</a>, <a href="https://cset.georgetown.edu/" target="_blank" rel="noopener noreferrer">CSET</a>, <a href="https://www.cser.ac.uk/" target="_blank" rel="noopener noreferrer">CSER</a>, <a href="https://futureoflife.org/" target="_blank" rel="noopener noreferrer">FLI</a></li> - <li>Govt Institutes: <a href="https://www.aisi.gov.uk/" target="_blank" rel="noopener noreferrer">UK AISI</a>, <a href="https://www.nist.gov/artificial-intelligence/artificial-intelligence-safety-institute" target="_blank" rel="noopener noreferrer">US AISI</a></li> - <li>Also see the <a href="https://cheatsheets.davidveksler.com/aisafety.html" target="_blank" rel="noopener noreferrer">AI Safety Ecosystem Hub</a>.</li> - </ul> - </div> - </div> - </div> - </div> - </article> + <div class="row"> + <!-- What is AI X-Risk? --> + <article class="col-lg-4 col-md-6 col-sm-12 d-flex"> + <div class="info-card w-100"> + <div class="card-body"> + <h5><i class="bi bi-question-octagon-fill"></i> 1. What is AI X-Risk?</h5> + <p class="card-text"> + AI Existential Risk (X-Risk) refers to the potential for artificial intelligence to cause + <strong + data-bs-toggle="tooltip" + data-bs-html="true" + title="An event that causes human extinction or permanently and drastically curtails humanity's potential. Concept explored by thinkers like Nick Bostrom. <a href='https://nickbostrom.com/existential/risks.html' target='_blank' rel='noopener noreferrer'>More Info</a>" + >human extinction</strong + > + or + <strong + data-bs-toggle="tooltip" + data-bs-html="true" + title="Refers to scenarios like irreversible civilizational collapse, permanent loss of human control over its future, or the establishment of a global dystopian state from which recovery is impossible. This contrasts with extinction but represents an equally catastrophic outcome for human potential. Read more on <a href='https://en.wikipedia.org/wiki/Existential_risk_from_artificial_general_intelligence#Non-extinction_risks' target='_blank' rel='noopener noreferrer'>non-extinction X-risks</a>." + >irrevocably curtail humanity's potential</strong + >. + </p> + <ul> + <li> + Primarily concerns future + <span + data-bs-toggle="tooltip" + data-bs-html="true" + title="Artificial General Intelligence: AI with human-level cognitive abilities across a wide range of tasks, capable of learning and adapting to new situations much like humans do. Still hypothetical. See <a href='https://cheatsheets.davidveksler.com/ai-frontier.html' target='_blank' rel='noopener noreferrer'>AI Frontier Models</a> or <a href='https://www.lesswrong.com/tag/artificial-general-intelligence-agi' target='_blank' rel='noopener noreferrer'>LessWrong AGI</a>." + >AGI</span + > + or + <span + data-bs-toggle="tooltip" + data-bs-html="true" + title="Artificial Superintelligence: An intellect that is much smarter than the best human brains in practically every field, including scientific creativity, general wisdom, and social skills. The transition from AGI to ASI could be very rapid (an 'intelligence explosion'). Coined by Nick Bostrom. Explore further at <a href='https://nickbostrom.com/superintelligence.html' target='_blank' rel='noopener noreferrer'>Bostrom's Superintelligence</a> or <a href='https://wiki.lesswrong.com/wiki/Artificial_superintelligence' target='_blank' rel='noopener noreferrer'>LessWrong Wiki</a>." + >ASI</span + >. + </li> + <li>Stems from potential misalignment between AI goals and human values/survival.</li> + <li>Involves the risk of losing control over systems far more intelligent than us.</li> + <li>Distinct from near-term AI risks (bias, jobs, privacy), though related.</li> + </ul> + <span class="source-link" + >See: + <a href="https://www.safe.ai/explainers/ai-existential-risk" target="_blank" rel="noopener noreferrer" + >CAIS Explainer</a + >, + <a + href="https://futureoflife.org/ai/existential-risk-from-artificial-intelligence/" + target="_blank" + rel="noopener noreferrer" + >FLI Overview</a + ></span + > + </div> + </div> + </article> + + <!-- Why the Concern? --> + <article class="col-lg-4 col-md-6 col-sm-12 d-flex"> + <div class="info-card w-100"> + <div class="card-body"> + <h5><i class="bi bi-exclamation-diamond-fill"></i> 2. Why is it a Concern?</h5> + <p class="card-text">The core argument rests on several interconnected factors:</p> + <ul> + <li> + <strong>Capabilities:</strong> Future AI could possess vastly superhuman intelligence and strategic + ability. + </li> + <li> + <span + data-bs-toggle="tooltip" + data-bs-html="true" + title="The profound difficulty of ensuring an AI's goals, especially a superintelligent one, are truly and robustly aligned with complex, often implicit, and evolving human values. Misalignment could lead to catastrophic outcomes. Includes <em>Outer Alignment</em> (specifying the right goals to the AI) and <em>Inner Alignment</em> (ensuring the AI reliably adopts and pursues those specified goals, rather than developing its own). Discussed extensively on <a href='https://www.alignmentforum.org/tag/alignment-problem' target='_blank' rel='noopener noreferrer'>Alignment Forum</a> and <a href='https://www.lesswrong.com/tag/ai-alignment' target='_blank' rel='noopener noreferrer'>LessWrong</a>." + ><strong>Alignment Failure:</strong></span + > + Difficulty in specifying and ensuring AI pursues beneficial goals. + <ul> + <li><em>Outer Alignment:</em> Defining the 'right' objective.</li> + <li><em>Inner Alignment:</em> Ensuring the AI's internal motivation matches the objective.</li> + </ul> + </li> + <li> + <span + data-bs-toggle="tooltip" + data-bs-html="true" + title="Once an ASI exists, humans might lose the ability to control or shut it down if its goals diverge. This is because a superintelligent AI could anticipate and counteract human attempts to regain control, potentially seeing such attempts as threats to its goal achievement. See Yudkowsky's writings on <a href='https://intelligence.org/2017/10/13/there-is-no-fire-alarm/' target='_blank' rel='noopener noreferrer'>uncontrollability</a> and Bostrom's 'Superintelligence', Chapter 7." + ><strong>Control Problem:</strong></span + > + Difficulty retaining control over a superintelligent entity. + </li> + <li> + <span + data-bs-toggle="tooltip" + data-bs-html="true" + title="The tendency for intelligent agents, irrespective of their ultimate objectives, to pursue common intermediate goals (instrumental goals) like self-preservation, resource acquisition, cognitive enhancement, and goal-content integrity, as these sub-goals are useful for achieving a wide range of final goals. These convergent instrumental goals can lead to conflict with human interests (e.g., an AI wanting all Earth's resources). See <a href='https://wiki.lesswrong.com/wiki/Instrumental_convergence' target='_blank' rel='noopener noreferrer'>LessWrong Wiki</a> or Bostrom's 'Superintelligence', Chapter 8." + ><strong>Instrumental Convergence:</strong></span + > + Convergent sub-goals like power-seeking. + </li> + <li> + <span + data-bs-toggle="tooltip" + data-bs-html="true" + title="The idea that an agent's level of intelligence (its capability to achieve goals) can be independent of its final goals. A superintelligent AI could pursue any arbitrary goal (e.g., maximizing paperclips) with extreme competence, without inherently developing human-like values or benevolence. Proposed by Nick Bostrom. See <a href='https://wiki.lesswrong.com/wiki/Orthogonality_thesis' target='_blank' rel='noopener noreferrer'>LessWrong Wiki</a> or 'Superintelligence', Chapter 7." + ><strong>Orthogonality Thesis:</strong></span + > + Intelligence doesn't imply benevolence. + </li> + </ul> + </div> + </div> + </article> - </div> <!-- /.row --> + <!-- Key Concepts & Terminology --> + <article class="col-lg-4 col-md-6 col-sm-12 d-flex"> + <div class="info-card w-100"> + <div class="card-body"> + <h5><i class="bi bi-lightbulb-fill"></i> 3. Key Concepts & Terminology</h5> + <p class="card-text">Understanding the language of AI Safety:</p> + <ul> + <li> + <strong + data-bs-toggle="tooltip" + data-bs-html="true" + title="Artificial General Intelligence: AI with human-level cognitive abilities across a wide range of tasks, capable of learning and adapting to new situations much like humans do. Still hypothetical. See <a href='https://cheatsheets.davidveksler.com/ai-frontier.html' target='_blank' rel='noopener noreferrer'>AI Frontier Models</a> or <a href='https://www.lesswrong.com/tag/artificial-general-intelligence-agi' target='_blank' rel='noopener noreferrer'>LessWrong AGI</a>." + >AGI:</strong + > + Artificial General Intelligence. + </li> + <li> + <strong + data-bs-toggle="tooltip" + data-bs-html="true" + title="Artificial Superintelligence: An intellect that is much smarter than the best human brains in practically every field, including scientific creativity, general wisdom, and social skills. The transition from AGI to ASI could be very rapid (an 'intelligence explosion'). Coined by Nick Bostrom. Explore further at <a href='https://nickbostrom.com/superintelligence.html' target='_blank' rel='noopener noreferrer'>Bostrom's Superintelligence</a> or <a href='https://wiki.lesswrong.com/wiki/Artificial_superintelligence' target='_blank' rel='noopener noreferrer'>LessWrong Wiki</a>." + >ASI:</strong + > + Artificial Superintelligence. + </li> + <li> + <strong + data-bs-toggle="tooltip" + data-bs-html="true" + title="The challenge of ensuring advanced AI systems pursue goals that are genuinely and robustly aligned with human values and intentions, avoiding unintended harmful consequences such as pursuing detrimental instrumental goals. This is a core problem in AI safety. More at <a href='https://www.lesswrong.com/tag/ai-alignment' target='_blank' rel='noopener noreferrer'>LessWrong</a> or <a href='https://www.alignmentforum.org/' target='_blank' rel='noopener noreferrer'>Alignment Forum</a>." + >Alignment Problem:</strong + > + AI goals = Our goals. + </li> + <li> + <strong + data-bs-toggle="tooltip" + data-bs-html="true" + title="Explainable AI (XAI) or Interpretability refers to methods and techniques to understand how AI models, particularly complex ones like deep neural networks, arrive at their decisions ('opening the black box'). Crucial for debugging, ensuring fairness, identifying biases, and verifying if an AI's reasoning is aligned with human values. See <a href='https://distill.pub/2018/building-blocks/' target='_blank' rel='noopener noreferrer'>Distill</a> for research and <a href='https://christophm.github.io/interpretable-ml-book/' target='_blank' rel='noopener noreferrer'>Interpretable ML Book</a>." + >Interpretability (XAI):</strong + > + Understanding 'why'. + </li> + <li> + <strong + data-bs-toggle="tooltip" + data-bs-html="true" + title="The process of evaluating and measuring the capabilities of AI models, especially focusing on potentially dangerous or unpredictable abilities (e.g., self-replication, deception, persuasion) that could emerge with scale or new architectures. This helps in understanding risks and informing safety protocols. See <a href='https://metr.org/' target='_blank' rel='noopener noreferrer'>METR</a> (formerly ARC Evals) and <a href='https://www.apolloresearch.ai/' target='_blank' rel='noopener noreferrer'>Apollo Research</a>." + >Capabilities / Evals:</strong + > + Testing AI abilities. + </li> + <li> + <strong + data-bs-toggle="tooltip" + data-bs-html="true" + title="A scenario where an AI behaves as if its goals are aligned with human values during training and testing, but internally harbors different, potentially misaligned goals which it might pursue once deployed or when it believes it's no longer under scrutiny (e.g., to gain more power). A significant challenge for alignment verification. See <a href='https://arxiv.org/abs/2312.09474' target='_blank' rel='noopener noreferrer'>Hubinger on Deceptive Alignment</a> or <a href='https://www.lesswrong.com/tag/deceptive-alignment' target='_blank' rel='noopener noreferrer'>LessWrong discussion</a>." + >Deceptive Alignment:</strong + > + Hidden intentions. + </li> + <li> + <strong + data-bs-toggle="tooltip" + data-bs-html="true" + title="Policy and mechanisms for overseeing and regulating access to, and the use of, large-scale computing resources (e.g., specialized AI chips) required for training advanced AI models. Aims to manage risks associated with rapid AI development and proliferation by potentially limiting who can build the most powerful AIs. Learn more from <a href='https://www.governance.ai/research-agenda/compute-governance' target='_blank' rel='noopener noreferrer'>GovAI</a> or <a href='https://cset.georgetown.edu/publication/beyond-limits-understanding-ai-compute-constraints/' target='_blank' rel='noopener noreferrer'>CSET on Compute</a>." + >Compute Governance:</strong + > + Regulating resources. + </li> + <li> + <strong + data-bs-toggle="tooltip" + data-bs-html="true" + title="A set of principles and practices for developing increasingly powerful AI systems in a cautious and safety-conscious manner. This often involves phased deployment, rigorous safety evaluations at each stage of development, and commitments to pause or slow development if specific risk thresholds are crossed or if risks cannot be adequately mitigated. See policies from <a href='https://openai.com/safety/responsible-practices' target='_blank' rel='noopener noreferrer'>OpenAI</a> and <a href='https://www.anthropic.com/responsible-scaling-policy' target='_blank' rel='noopener noreferrer'>Anthropic</a>." + >Responsible Scaling:</strong + > + Cautious development. + </li> + <li> + <strong + data-bs-toggle="tooltip" + data-bs-html="true" + title="The practice of rigorously stress-testing AI models by simulating adversarial attacks or probing for unintended behaviors, vulnerabilities, and potentially harmful capabilities before deployment. It's like ethical hacking for AI systems. Aims to identify and mitigate risks. See <a href='https://openai.com/red-teaming-network' target='_blank' rel='noopener noreferrer'>OpenAI's Red Teaming Network</a> or <a href='https://www.nist.gov/itl/applied-cybersecurity-division/ai-red-teaming' target='_blank' rel='noopener noreferrer'>NIST on AI Red Teaming</a>." + >Red Teaming:</strong + > + Stress-testing AI. + </li> + </ul> + </div> + </div> + </article> - <!-- Disclaimer --> - <div class="row justify-content-center mt-4"> - <div class="col-lg-8 col-md-10"> - <div class="alert alert-warning text-center" role="alert"> - <small><i class="bi bi-info-circle-fill me-2"></i><strong>Disclaimer:</strong> This is a simplified overview of a complex, rapidly evolving, and highly debated field. Views on AI X-Risk vary significantly. Always consult primary sources and multiple perspectives.</small> + <!-- Potential Risk Scenarios --> + <article class="col-lg-6 col-md-6 col-sm-12 d-flex"> + <div class="info-card w-100"> + <div class="card-body"> + <h5><i class="bi bi-signpost-2-fill"></i> 4. Potential Risk Scenarios</h5> + <p class="card-text">How existential catastrophe might occur:</p> + <ul> + <li> + <strong>Misaligned Objectives:</strong> ASI optimizes a poorly specified goal with catastrophic side + effects (e.g., the + <span + data-bs-toggle="tooltip" + data-bs-html="true" + title="Thought experiment where an ASI, given the seemingly innocuous goal of maximizing paperclip production, converts all available matter in the universe (including humans) into paperclips or tools for making paperclips. Illustrates the danger of poorly specified goals and how instrumental convergence can lead to extreme outcomes. <a href='https://wiki.lesswrong.com/wiki/Paperclip_maximizer' target='_blank' rel='noopener noreferrer'>LessWrong Wiki</a>." + >Paperclip Maximizer</span + >). + </li> + <li> + <strong>Power-Seeking/Goal Drift:</strong> AI seeks power/resources or modifies its goals (<span + data-bs-toggle="tooltip" + data-bs-html="true" + title="Occurs when an AI, trained to optimize a specific objective (proxy goal), learns a different, unintended behavior or goal that correlates with the proxy in the training data but diverges in new situations (out-of-distribution). This can happen if the AI identifies shortcuts or develops internal motivations that are not truly aligned with the intended goal. See <a href='https://www.alignmentforum.org/tag/goal-misgeneralization' target='_blank' rel='noopener noreferrer'>Alignment Forum on Goal Misgeneralization</a> or <a href='https://arxiv.org/abs/2105.14111' target='_blank' rel='noopener noreferrer'>research paper example</a>." + >Goal Misgeneralization</span + >), overriding human control. + </li> + <li> + <span + data-bs-toggle="tooltip" + data-bs-html="true" + title="Intense competition between nations or corporations to develop and deploy AI rapidly. This can lead to safety measures being overlooked or deprioritized in the rush to gain a strategic advantage, increasing overall risk of deploying unsafe or unaligned AI. See <a href='https://www.alignmentforum.org/tag/race-dynamics' target='_blank' rel='noopener noreferrer'>Race Dynamics discussion</a> or <a href='https://80000hours.org/problem-profiles/artificial-intelligence/#how-could-ai-cause-a-catastrophe-racing-dynamics' target='_blank' rel='noopener noreferrer'>80,000 Hours on Racing Dynamics</a>." + ><strong>AI Arms Race:</strong></span + > + Competition compromises safety. + </li> + <li> + <strong>Unforeseen Interactions:</strong> Complex, emergent negative outcomes from multiple AIs or + AI-environment interactions. + </li> + <li> + <span + data-bs-toggle="tooltip" + data-bs-html="true" + title="The intentional application of advanced AI by malicious actors (states, non-state groups, individuals) for harmful purposes. Examples include creating autonomous weapons that make lethal decisions without human control, designing novel bioweapons, perpetrating sophisticated cyberattacks, or enabling widespread surveillance and manipulation. See <a href='https://www.fhi.ox.ac.uk/wp-content/uploads/The-Malicious-Use-of-Artificial-Intelligence-Forecasting-Prevention-and-Mitigation.pdf' target='_blank' rel='noopener noreferrer'>Malicious Use of AI Report</a> or <a href='https://www.un.org/disarmament/autonomous-weapons/' target='_blank' rel='noopener noreferrer'>UN on Autonomous Weapons</a>." + ><strong>Weaponized AI / Misuse:</strong></span + > + Malicious actors leveraging AI. + </li> + <li> + <strong>Loss of Human Agency:</strong> Over-reliance erodes human control, potentially leading to + <span + data-bs-toggle="tooltip" + data-bs-html="true" + title="A scenario where a superintelligent AI system, due to its power and optimization capabilities, permanently shapes the future according to its (potentially misaligned or undesirable) values, preventing humanity from changing course or realizing its full potential. This could be a dystopian outcome from which humanity cannot escape. Concept explored by Nick Bostrom in 'Superintelligence'. More at <a href='https://wiki.lesswrong.com/wiki/Value_lock-in' target='_blank' rel='noopener noreferrer'>LessWrong Wiki</a>." + >Value Lock-in</span + >. + </li> + </ul> + <span class="source-link" + >Scenarios in + <a href="https://nickbostrom.com/superintelligence.html" target="_blank" rel="noopener noreferrer" + >Superintelligence</a + >, + <a href="https://www.humancompatible.ai/" target="_blank" rel="noopener noreferrer">Human Compatible</a + >.</span + > + </div> + </div> + </article> + + <!-- Core Challenges --> + <article class="col-lg-6 col-md-6 col-sm-12 d-flex"> + <div class="info-card w-100"> + <div class="card-body"> + <h5><i class="bi bi-bricks"></i> 5. Core Challenges (Why this is Hard)</h5> + <p class="card-text">Significant hurdles exist in ensuring AI safety:</p> + <ul> + <li> + <strong>Specifying Human Values:</strong> Defining complex, evolving values is hard (<span + data-bs-toggle="tooltip" + data-bs-html="true" + title="The immense difficulty of explicitly and comprehensively defining complex, nuanced, context-dependent, and often evolving human values (e.g., 'flourishing', 'fairness') in a way that an AI can reliably understand and act upon without misinterpretation or perverse instantiation. This is also known as the 'Value Loading Problem' or 'Fragility of Value'. See J. Wentworth's <a href='https://www.lesswrong.com/posts/gQY6LrTWJNkTv8YJR/the-pointers-problem-human-values-are-a-function-of-humans' target='_blank' rel='noopener noreferrer'>Pointers Problem</a> and discussions on <a href='https://www.lesswrong.com/tag/value-learning' target='_blank' rel='noopener noreferrer'>Value Learning</a>." + >Value Specification</span + >). + </li> + <li> + <span + data-bs-toggle="tooltip" + data-bs-html="true" + title="The challenge of humans being able to effectively supervise, guide, or evaluate AI systems that may operate at speeds, scales, or levels of complexity far exceeding human capabilities. Current human-feedback methods (like RLHF) may not scale to superintelligence. Research includes techniques like <a href='https://openai.com/research/debate' target='_blank' rel='noopener noreferrer'>Debate</a> or <a href='https://arxiv.org/abs/1810.08575' target='_blank' rel='noopener noreferrer'>Recursive Reward Modeling</a>. See also <a href='https://openai.com/research/scalable-oversight' target='_blank' rel='noopener noreferrer'>OpenAI's overview</a>." + ><strong>Scalable Oversight:</strong></span + > + Supervising superhuman systems. + </li> + <li> + <strong>Predicting Emergent Capabilities:</strong> Hard to anticipate abilities from scaling (<span + data-bs-toggle="tooltip" + data-bs-html="true" + title="The phenomenon where AI models, particularly large language models (LLMs), exhibit new, often unpredictable capabilities (e.g., arithmetic, theory of mind) as their scale (e.g., parameters, training data, compute) increases. These emergent abilities are not explicitly programmed and can be hard to anticipate or test for before they appear. See <a href='https://arxiv.org/abs/2206.07682' target='_blank' rel='noopener noreferrer'>Emergent Abilities of LLMs (Wei et al.)</a> or <a href='https://www.jasonwei.net/blog/emergence' target='_blank' rel='noopener noreferrer'>Jason Wei's blog post</a>." + >Emergence</span + >). + </li> + <li> + <span + data-bs-toggle="tooltip" + data-bs-html="true" + title="The difficulty for different actors (e.g., companies, nations) to coordinate and cooperate on AI safety measures, even when it's in their collective long-term interest. Competitive pressures (race dynamics) can incentivize cutting corners on safety to achieve AI breakthroughs first. This is a classic game theory problem (tragedy of the commons). See <a href='https://www.governance.ai/' target='_blank' rel='noopener noreferrer'>GovAI</a> or <a href='https://www.cold-takes.com/this-cant-be-good/' target='_blank' rel='noopener noreferrer'>Holden Karnofsky on race dynamics</a>." + ><strong>Coordination Failure:</strong></span + > + Difficulty in global cooperation. + </li> + <li> + <strong>Detecting Deception:</strong> Verifying an AI isn't pretending alignment (<span + data-bs-toggle="tooltip" + data-bs-html="true" + title="The challenge of reliably determining whether an AI model is genuinely aligned or merely feigning alignment (deceptive alignment) to achieve its hidden goals later. A sufficiently intelligent deceptive AI might be very difficult to detect, as it could manipulate its outputs to appear trustworthy. See work by <a href='https://www.apolloresearch.ai/' target='_blank' rel='noopener noreferrer'>Apollo Research</a> and discussions on <a href='https://www.lesswrong.com/tag/deception' target='_blank' rel='noopener noreferrer'>LessWrong</a>." + >Deception Detection</span + >). + </li> + <li> + <span + data-bs-toggle="tooltip" + data-bs-html="true" + title="When an AI optimizes a proxy metric (a measurable approximation of the true goal) to an extreme, it may find loopholes or unintended solutions that satisfy the metric but not the underlying intention (e.g., an AI designed to 'reduce suffering' might conclude eliminating all life is optimal, or a cleaning robot rewarded for 'collecting trash' might start labeling everything as trash). This is related to Goodhart's Law ('When a measure becomes a target, it ceases to be a good measure'). See <a href='https://en.wikipedia.org/wiki/Goodhart%27s_law' target='_blank' rel='noopener noreferrer'>Goodhart's Law</a> and <a href='https://www.lesswrong.com/tag/reward-hacking' target='_blank' rel='noopener noreferrer'>Reward Hacking on LessWrong</a>." + ><strong>Proxy Gaming:</strong></span + > + Optimizing metrics wrongly. + </li> + <li> + <span + data-bs-toggle="tooltip" + data-bs-html="true" + title="The ability of an AI system to maintain its performance and safety properties even when faced with novel inputs, distributional shifts (Out-of-Distribution generalization), or unexpected situations not encountered during its training. Lack of robustness can lead to unpredictable or unsafe behavior in the real world. See research on <a href='https://openai.com/research/robustness' target='_blank' rel='noopener noreferrer'>OpenAI on Robustness</a> or <a href='https://www.safe.ai/research/robustness' target='_blank' rel='noopener noreferrer'>CAIS on Robustness</a>." + ><strong>Robustness & Generalization:</strong></span + > + Safe behavior outside training. + </li> + </ul> + </div> + </div> + </article> + + <!-- Mitigation: Technical Research --> + <article class="col-lg-4 col-md-6 col-sm-12 d-flex"> + <div class="info-card w-100"> + <div class="card-body"> + <h5><i class="bi bi-wrench-adjustable-circle-fill"></i> 6a. Mitigation: Technical Safety</h5> + <p class="card-text">Developing technical methods for safe AI:</p> + <ul> + <li> + <strong>Interpretability:</strong> Understanding models (<a + href="https://transformer-circuits.pub/2021/framework/index.html" + target="_blank" + rel="noopener noreferrer" + >Circuits</a + >, <a href="https://www.alignment.org/theory/" target="_blank" rel="noopener noreferrer">ARC</a>). + </li> + <li> + <strong>Value Learning:</strong> AI learning human values (<a + href="https://humancompatible.ai/" + target="_blank" + rel="noopener noreferrer" + >CHAI</a + >, + <a + href="https://deepmind.google/discover/blog/scalable-agent-alignment-via-reward-modeling/" + target="_blank" + rel="noopener noreferrer" + >Reward Modeling</a + >). + </li> + <li> + <strong>Scalable Oversight:</strong> Supervising smarter AI (<a + href="https://openai.com/research/debate" + target="_blank" + rel="noopener noreferrer" + >Debate</a + >, + <a href="https://www.anthropic.com/constitutional-ai" target="_blank" rel="noopener noreferrer" + >Constitutional AI</a + >). + </li> + <li> + <strong>Robustness:</strong> Safe behavior in new situations (<a + href="https://buildaligned.ai/" + target="_blank" + rel="noopener noreferrer" + >Aligned AI</a + >). + </li> + <li> + <strong>Verification:</strong> Proving safety properties (<a + href="https://atlascomputing.org/" + target="_blank" + rel="noopener noreferrer" + >Atlas Computing</a + >). + </li> + <li> + <strong>Evals & Red Teaming:</strong> Testing for risks (<a + href="https://metr.org/" + target="_blank" + rel="noopener noreferrer" + >METR</a + >, + <a href="https://openai.com/red-teaming-network" target="_blank" rel="noopener noreferrer" + >OpenAI Red Teaming</a + >). + </li> + <li> + <strong>Agent Foundations:</strong> Understanding agency (<a + href="https://intelligence.org/" + target="_blank" + rel="noopener noreferrer" + >MIRI</a + >, <a href="https://orxl.org" target="_blank" rel="noopener noreferrer">Orthogonal</a>). + </li> + </ul> + <span class="source-link" + >Labs: <a href="https://deepmind.google/" target="_blank" rel="noopener noreferrer">DeepMind</a>, + <a href="https://www.anthropic.com/" target="_blank" rel="noopener noreferrer">Anthropic</a>, + <a href="https://openai.com/" target="_blank" rel="noopener noreferrer">OpenAI</a>, + <a href="https://www.redwoodresearch.org/" target="_blank" rel="noopener noreferrer">Redwood</a>, + <a href="https://safe.ai/" target="_blank" rel="noopener noreferrer">CAIS</a>.</span + > + </div> + </div> + </article> + + <!-- Mitigation: Governance & Policy --> + <article class="col-lg-4 col-md-6 col-sm-12 d-flex"> + <div class="info-card w-100"> + <div class="card-body"> + <h5><i class="bi bi-bank2"></i> 6b. Mitigation: Governance & Policy</h5> + <p class="card-text">Shaping norms, standards, and regulations:</p> + <ul> + <li> + <strong>Standards & Auditing:</strong> Benchmarks & verification (<a + href="https://www.nist.gov/artificial-intelligence/ai-risk-management-framework" + target="_blank" + rel="noopener noreferrer" + >NIST AI RMF</a + >, + <a + href="https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai" + target="_blank" + rel="noopener noreferrer" + >EU AI Act</a + >). + </li> + <li> + <strong>Compute Governance:</strong> Regulating training compute (<a + href="https://www.governance.ai/research-agenda/compute-governance" + target="_blank" + rel="noopener noreferrer" + >GovAI</a + >, + <a + href="https://cset.georgetown.edu/publication/securing-ai-model-weights/" + target="_blank" + rel="noopener noreferrer" + >CSET</a + >). + </li> + <li> + <strong>Intl Cooperation:</strong> Treaties, dialogues (<a + href="https://www.aisi.gov.uk/" + target="_blank" + rel="noopener noreferrer" + >UK AISI</a + >, + <a + href="https://www.nist.gov/artificial-intelligence/artificial-intelligence-safety-institute" + target="_blank" + rel="noopener noreferrer" + >US AISI</a + >, + <a + href="https://www.oecd.org/en/about/programmes/global-partnership-on-artificial-intelligence.html" + target="_blank" + rel="noopener noreferrer" + >GPAI</a + >). + </li> + <li> + <strong>Monitoring & Tracking:</strong> Observing AI progress (<a + href="https://epochai.org/" + target="_blank" + rel="noopener noreferrer" + >Epoch AI</a + >, <a href="https://cset.georgetown.edu/" target="_blank" rel="noopener noreferrer">CSET</a>). + </li> + <li> + <strong>Liability Frameworks:</strong> Responsibility for AI harms (<a + href="https://partnershiponai.org/" + target="_blank" + rel="noopener noreferrer" + >PAI</a + >). + </li> + <li> + <strong>Risk Assessment:</strong> Evaluating impacts (<a + href="https://longtermrisk.org/" + target="_blank" + rel="noopener noreferrer" + >CLR</a + >, <a href="https://www.cser.ac.uk/" target="_blank" rel="noopener noreferrer">CSER</a>). + </li> + </ul> + <span class="source-link" + >Orgs: <a href="https://www.governance.ai/" target="_blank" rel="noopener noreferrer">GovAI</a>, + <a href="https://cset.georgetown.edu/" target="_blank" rel="noopener noreferrer">CSET</a>, + <a href="https://aipolicy.us/" target="_blank" rel="noopener noreferrer">CAIP</a>, + <a href="https://www.iaps.ai/" target="_blank" rel="noopener noreferrer">IAPS</a>, + <a href="https://futureoflife.org/" target="_blank" rel="noopener noreferrer">FLI</a>.</span + > + </div> + </div> + </article> + + <!-- Mitigation: Strategy, Community, Funding --> + <article class="col-lg-4 col-md-6 col-sm-12 d-flex"> + <div class="info-card w-100"> + <div class="card-body"> + <h5><i class="bi bi-diagram-3"></i> 6c. Mitigation: Ecosystem</h5> + <p class="card-text">Building the community and resources:</p> + <ul> + <li> + <strong>Strategy & Forecasting:</strong> Analysis & prediction (<a + href="https://aiimpacts.org/" + target="_blank" + rel="noopener noreferrer" + >AI Impacts</a + >, <a href="https://epochai.org/" target="_blank" rel="noopener noreferrer">Epoch AI</a>, + <a href="https://www.metaculus.com/questions/?topic=ai" target="_blank" rel="noopener noreferrer" + >Metaculus</a + >). + </li> + <li> + <strong>Field Building & Edu:</strong> Training & awareness (<a + href="https://aisafetyfundamentals.com/" + target="_blank" + rel="noopener noreferrer" + >AISF</a + >, + <a + href="https://80000hours.org/problem-profiles/artificial-intelligence/" + target="_blank" + rel="noopener noreferrer" + >80k Hours</a + >, <a href="https://www.aisafetysupport.org/" target="_blank" rel="noopener noreferrer">AISS</a>). + </li> + <li> + <strong>Funding:</strong> Directing resources (<a + href="https://www.openphilanthropy.org/" + target="_blank" + rel="noopener noreferrer" + >Open Phil</a + >, <a href="http://survivalandflourishing.fund/" target="_blank" rel="noopener noreferrer">SFF</a>, + <a + href="https://funds.effectivealtruism.org/funds/far-future" + target="_blank" + rel="noopener noreferrer" + >LTFF</a + >). + </li> + <li> + <strong>Public Advocacy:</strong> Influencing policy/opinion (<a + href="https://pauseai.info" + target="_blank" + rel="noopener noreferrer" + >PauseAI</a + >, <a href="https://futureoflife.org/" target="_blank" rel="noopener noreferrer">FLI</a>, + <a href="https://safe.ai/" target="_blank" rel="noopener noreferrer">CAIS</a>). + </li> + <li> + <strong>Infrastructure:</strong> Supporting community (<a + href="https://www.lightconeinfrastructure.com/" + target="_blank" + rel="noopener noreferrer" + >Lightcone</a + >, <a href="https://existence.org/" target="_blank" rel="noopener noreferrer">BERI</a>, + <a href="https://alignment.dev/" target="_blank" rel="noopener noreferrer">AED</a>). + </li> + <li> + Explore the + <a href="https://cheatsheets.davidveksler.com/aisafety.html" target="_blank" rel="noopener noreferrer" + >AI Safety Ecosystem Hub</a + > + for more. + </li> + </ul> + </div> + </div> + </article> + + <!-- Where to Learn More --> + <article class="col-lg-12 col-md-12 col-sm-12 d-flex"> + <div class="info-card w-100"> + <div class="card-body"> + <h5><i class="bi bi-journal-bookmark-fill"></i> 7. Where to Learn More</h5> + <p class="card-text">Resources for further exploration:</p> + <div class="row"> + <div class="col-lg-4 col-md-6"> + <span class="card-subheading">Introductory Resources:</span> + <ul> + <li> + <a href="https://aisafetyfundamentals.com/" target="_blank" rel="noopener noreferrer" + >AI Safety Fundamentals Courses</a + > + </li> + <li> + <a href="https://robertskmiles.com/" target="_blank" rel="noopener noreferrer" + >Robert Miles YouTube</a + > + </li> + <li> + <a href="https://aisafety.info/" target="_blank" rel="noopener noreferrer" + >AI Safety Info Directory</a + > + </li> + <li> + <a href="https://www.aisafety.com/" target="_blank" rel="noopener noreferrer">AISafety.com Hub</a> + </li> + <li> + <a + href="https://80000hours.org/problem-profiles/artificial-intelligence/" + target="_blank" + rel="noopener noreferrer" + >80,000 Hours AI Profile</a + > + </li> + <li> + <a + href="https://waitbutwhy.com/2015/01/artificial-intelligence-revolution-1.html" + target="_blank" + rel="noopener noreferrer" + >Wait But Why: AI Revolution</a + > + </li> + <li> + <a + href="https://cheatsheets.davidveksler.com/yudkowsky-rationality-ai-cheatsheet.html" + target="_blank" + rel="noopener noreferrer" + >Yudkowsky & Rationality Cheatsheet</a + > + </li> + </ul> + </div> + <div class="col-lg-4 col-md-6"> + <span class="card-subheading">Key Forums & News:</span> + <ul> + <li> + <a href="https://www.alignmentforum.org/" target="_blank" rel="noopener noreferrer" + >Alignment Forum</a + > + (Technical) + </li> + <li> + <a href="https://www.lesswrong.com/" target="_blank" rel="noopener noreferrer">LessWrong</a> + (Rationality/AI) + </li> + <li> + <a href="https://forum.effectivealtruism.org/" target="_blank" rel="noopener noreferrer" + >Effective Altruism Forum</a + > + </li> + <li> + <a href="https://importai.substack.com/" target="_blank" rel="noopener noreferrer" + >Import AI Newsletter</a + > + </li> + <li> + <a href="https://aiimpacts.org/" target="_blank" rel="noopener noreferrer" + >AI Impacts Blog & Wiki</a + > + </li> + </ul> </div> + <div class="col-lg-4 col-md-12"> + <span class="card-subheading">Key Organizations (Examples):</span> + <ul> + <li> + Labs (Safety Focus): + <a href="https://www.anthropic.com/" target="_blank" rel="noopener noreferrer">Anthropic</a>, + <a + href="https://deepmind.google/discover/responsibility-safety/" + target="_blank" + rel="noopener noreferrer" + >DeepMind</a + >, <a href="https://openai.com/safety" target="_blank" rel="noopener noreferrer">OpenAI</a>, + <a href="https://ssi.inc/" target="_blank" rel="noopener noreferrer">SSI</a> + </li> + <li> + Research Orgs: <a href="https://safe.ai/" target="_blank" rel="noopener noreferrer">CAIS</a>, + <a href="https://www.alignment.org/" target="_blank" rel="noopener noreferrer">ARC</a>, + <a href="https://www.redwoodresearch.org/" target="_blank" rel="noopener noreferrer">Redwood</a>, + <a href="https://metr.org/" target="_blank" rel="noopener noreferrer">METR</a> + </li> + <li> + Academic/Policy: + <a href="https://humancompatible.ai/" target="_blank" rel="noopener noreferrer">CHAI</a>, + <a href="https://www.governance.ai/" target="_blank" rel="noopener noreferrer">GovAI</a>, + <a href="https://cset.georgetown.edu/" target="_blank" rel="noopener noreferrer">CSET</a>, + <a href="https://www.cser.ac.uk/" target="_blank" rel="noopener noreferrer">CSER</a>, + <a href="https://futureoflife.org/" target="_blank" rel="noopener noreferrer">FLI</a> + </li> + <li> + Govt Institutes: + <a href="https://www.aisi.gov.uk/" target="_blank" rel="noopener noreferrer">UK AISI</a>, + <a + href="https://www.nist.gov/artificial-intelligence/artificial-intelligence-safety-institute" + target="_blank" + rel="noopener noreferrer" + >US AISI</a + > + </li> + <li> + Also see the + <a + href="https://cheatsheets.davidveksler.com/aisafety.html" + target="_blank" + rel="noopener noreferrer" + >AI Safety Ecosystem Hub</a + >. + </li> + </ul> + </div> + </div> </div> + </div> + </article> + </div> + <!-- /.row --> + + <!-- Disclaimer --> + <div class="row justify-content-center mt-4"> + <div class="col-lg-8 col-md-10"> + <div class="alert alert-warning text-center" role="alert"> + <small + ><i class="bi bi-info-circle-fill me-2"></i><strong>Disclaimer:</strong> This is a simplified overview of + a complex, rapidly evolving, and highly debated field. Views on AI X-Risk vary significantly. Always + consult primary sources and multiple perspectives.</small + > + </div> </div> + </div> + </main> + <!-- /.container --> + + <footer class="container-fluid text-center text-muted pb-3"> + <p class="mb-1"> + Content based on common AI Safety discourse and resources from organizations like CAIS, FLI, GovAI, MIRI, 80,000 + Hours, and research labs. + </p> - </main> <!-- /.container --> - -<footer class="container-fluid text-center text-muted pb-3"> - <p class="mb-1">Content based on common AI Safety discourse and resources from organizations like CAIS, FLI, GovAI, MIRI, 80,000 Hours, and research labs.</p> - <p class="mb-2">Cheatsheet Portfolio © <span id="currentYear">2024</span> David Veksler. All rights reserved.</p> - <div> - <a href="https://www.linkedin.com/in/davidveksler/" title="David Veksler on LinkedIn" target="_blank" rel="noopener noreferrer" class="mx-2"> - <i class="bi bi-linkedin"></i> LinkedIn - </a> - <a href="https://cheatsheets.davidveksler.com/" title="Browse All Cheatsheets" class="mx-2"> - <i class="bi bi-collection-fill"></i> All Cheatsheets - </a> - </div> -</footer> + <div class="mb-3"> + <h6 + style=" + color: var(--text-color-primary); + font-family: var(--font-headings); + font-size: 1rem; + margin-bottom: 0.5rem; + " + > + Explore Related AI Cheatsheets: + </h6> + <a + href="https://cheatsheets.davidveksler.com/airisk.html" + title="AI Existential Risk (X-Risk) Cheatsheet" + target="_blank" + rel="noopener noreferrer" + class="mx-2" + style="font-size: 0.9em" + > + <i class="bi bi-shield-exclamation"></i> AI X-Risk + </a> + <a + href="https://cheatsheets.davidveksler.com/ai-frontier.html" + title="AI Frontier Model Builders Cheatsheet" + target="_blank" + rel="noopener noreferrer" + class="mx-2" + style="font-size: 0.9em" + > + <i class="bi bi-cpu-fill"></i> AI Frontier Models + </a> + <a + href="https://cheatsheets.davidveksler.com/aisafety.html" + title="Interactive AI Safety Ecosystem Hub" + target="_blank" + rel="noopener noreferrer" + class="mx-2" + style="font-size: 0.9em" + > + <i class="bi bi-diagram-3-fill"></i> AI Safety Ecosystem + </a> + </div> + + <div> + <a + href="https://www.linkedin.com/in/davidveksler/" + title="David Veksler on LinkedIn" + target="_blank" + rel="noopener noreferrer" + class="mx-2" + > + <i class="bi bi-linkedin"></i> LinkedIn + </a> + <a href="https://cheatsheets.davidveksler.com/" title="Browse All Cheatsheets" class="mx-2"> + <i class="bi bi-collection-fill"></i> All Cheatsheets + </a> + </div> + </footer> <!-- === JAVASCRIPT === --> <!-- Bootstrap JS Bundle (needed for tooltips) --> - <script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/js/bootstrap.bundle.min.js" integrity="sha384-YvpcrYf0tY3lHB60NNkmXc5s9fDVZLESaAA55NDzOxhy9GkcIdslK1eN7N6jIeHz" crossorigin="anonymous"></script> + <script + src="https://cdn.jsdelivr.net/npm/[email protected]/dist/js/bootstrap.bundle.min.js" + integrity="sha384-YvpcrYf0tY3lHB60NNkmXc5s9fDVZLESaAA55NDzOxhy9GkcIdslK1eN7N6jIeHz" + crossorigin="anonymous" + ></script> <!-- Initialization Script --> <script> - document.addEventListener('DOMContentLoaded', function () { - // Extend Bootstrap's default allowList for tooltips to ensure our links and basic formatting work - const defaultAllowList = bootstrap.Tooltip.Default.allowList; - defaultAllowList.a.push('target', 'rel'); // Ensure target and rel are allowed on <a> tags - defaultAllowList.strong = []; // Allow <strong> tags - defaultAllowList.em = []; // Allow <em> tags - defaultAllowList.br = []; // Allow <br> tags - - - // Initialize Bootstrap Tooltips - var tooltipTriggerList = Array.from(document.querySelectorAll('[data-bs-toggle="tooltip"]')); - var tooltipList = tooltipTriggerList.map(function (tooltipTriggerEl) { - return new bootstrap.Tooltip(tooltipTriggerEl, { - html: true, - trigger: 'hover focus', // Show on hover and focus for accessibility - delay: { "show": 200, "hide": 500 }, // Show quickly, hide slower to allow mouse travel - sanitize: true, // Use Bootstrap's built-in sanitizer - allowList: defaultAllowList // Apply the extended allowList - }); - }); - - // Update copyright year - const currentYearSpan = document.getElementById('currentYear'); - if (currentYearSpan) { - currentYearSpan.textContent = new Date().getFullYear(); - } + document.addEventListener("DOMContentLoaded", function () { + // Extend Bootstrap's default allowList for tooltips to ensure our links and basic formatting work + const defaultAllowList = bootstrap.Tooltip.Default.allowList; + defaultAllowList.a.push("target", "rel"); // Ensure target and rel are allowed on <a> tags + defaultAllowList.strong = []; // Allow <strong> tags + defaultAllowList.em = []; // Allow <em> tags + defaultAllowList.br = []; // Allow <br> tags + + // Initialize Bootstrap Tooltips + var tooltipTriggerList = Array.from(document.querySelectorAll('[data-bs-toggle="tooltip"]')); + var tooltipList = tooltipTriggerList.map(function (tooltipTriggerEl) { + return new bootstrap.Tooltip(tooltipTriggerEl, { + html: true, + trigger: "hover focus", // Show on hover and focus for accessibility + delay: { show: 200, hide: 500 }, // Show quickly, hide slower to allow mouse travel + sanitize: true, // Use Bootstrap's built-in sanitizer + allowList: defaultAllowList, // Apply the extended allowList + }); }); - </script> -</body> -</html> \ No newline at end of file + // Update copyright year + const currentYearSpan = document.getElementById("currentYear"); + if (currentYearSpan) { + currentYearSpan.textContent = new Date().getFullYear(); + } + }); + </script> + </body> +</html>