#452 – Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity

Sobre este episódio

<p>Dario Amodei is the CEO of Anthropic, the company that created Claude. Amanda Askell is an AI researcher working on Claude&#8217;s character and personality. Chris Olah is an AI researcher working on mechanistic interpretability.<br /> Thank you for listening ❤ Check out our sponsors: <a href="https://lexfridman.com/sponsors/ep452-sc">https://lexfridman.com/sponsors/ep452-sc</a><br /> See below for timestamps, transcript, and to give feedback, submit questions, contact Lex, etc.</p> <p><b>Transcript:</b><br /> <a href="https://lexfridman.com/dario-amodei-transcript">https://lexfridman.com/dario-amodei-transcript</a></p> <p><b>CONTACT LEX:</b><br /> <b>Feedback</b> &#8211; give feedback to Lex: <a href="https://lexfridman.com/survey">https://lexfridman.com/survey</a><br /> <b>AMA</b> &#8211; submit questions, videos or call-in: <a href="https://lexfridman.com/ama">https://lexfridman.com/ama</a><br /> <b>Hiring</b> &#8211; join our team: <a href="https://lexfridman.com/hiring">https://lexfridman.com/hiring</a><br /> <b>Other</b> &#8211; other ways to get in touch: <a href="https://lexfridman.com/contact">https://lexfridman.com/contact</a></p> <p><b>EPISODE LINKS:</b><br /> Claude: <a href="https://claude.ai">https://claude.ai</a><br /> Anthropic&#8217;s X: <a href="https://x.com/AnthropicAI">https://x.com/AnthropicAI</a><br /> Anthropic&#8217;s Website: <a href="https://anthropic.com">https://anthropic.com</a><br /> Dario&#8217;s X: <a href="https://x.com/DarioAmodei">https://x.com/DarioAmodei</a><br /> Dario&#8217;s Website: <a href="https://darioamodei.com">https://darioamodei.com</a><br /> Machines of Loving Grace (Essay): <a href="https://darioamodei.com/machines-of-loving-grace">https://darioamodei.com/machines-of-loving-grace</a><br /> Chris&#8217;s X: <a href="https://x.com/ch402">https://x.com/ch402</a><br /> Chris&#8217;s Blog: <a href="https://colah.github.io">https://colah.github.io</a><br /> Amanda&#8217;s X: <a href="https://x.com/AmandaAskell">https://x.com/AmandaAskell</a><br /> Amanda&#8217;s Website: <a href="https://askell.io">https://askell.io</a></p> <p><b>SPONSORS:</b><br /> To support this podcast, check out our sponsors &#38; get discounts:<br /> <b>Encord:</b> AI tooling for annotation &#38; data management.<br /> Go to <a href="https://lexfridman.com/s/encord-ep452-sc">https://encord.com/lex</a><br /> <b>Notion:</b> Note-taking and team collaboration.<br /> Go to <a href="https://lexfridman.com/s/notion-ep452-sc">https://notion.com/lex</a><br /> <b>Shopify:</b> Sell stuff online.<br /> Go to <a href="https://lexfridman.com/s/shopify-ep452-sc">https://shopify.com/lex</a><br /> <b>BetterHelp:</b> Online therapy and counseling.<br /> Go to <a href="https://lexfridman.com/s/betterhelp-ep452-sc">https://betterhelp.com/lex</a><br /> <b>LMNT:</b> Zero-sugar electrolyte drink mix.<br /> Go to <a href="https://lexfridman.com/s/lmnt-ep452-sc">https://drinkLMNT.com/lex</a></p> <p><b>OUTLINE:</b><br /> (00:00) &#8211; Introduction<br /> (10:19) &#8211; Scaling laws<br /> (19:25) &#8211; Limits of LLM scaling<br /> (27:51) &#8211; Competition with OpenAI, Google, xAI, Meta<br /> (33:14) &#8211; Claude<br /> (36:50) &#8211; Opus 3.5<br /> (41:36) &#8211; Sonnet 3.5<br /> (44:56) &#8211; Claude 4.0<br /> (49:07) &#8211; Criticism of Claude<br /> (1:01:54) &#8211; AI Safety Levels<br /> (1:12:42) &#8211; ASL-3 and ASL-4<br /> (1:16:46) &#8211; Computer use<br /> (1:26:41) &#8211; Government regulation of AI<br /> (1:45:30) &#8211; Hiring a great team<br /> (1:54:19) &#8211; Post-training<br /> (1:59:45) &#8211; Constitutional AI<br /> (2:05:11) &#8211; Machines of Loving Grace<br /> (2:24:17) &#8211; AGI timeline<br /> (2:36:52) &#8211; Programming<br /> (2:43:52) &#8211; Meaning of life<br /> (2:49:58) &#8211; Amanda Askell &#8211; Philosophy<br /> (2:52:26) &#8211; Programming advice for non-technical people<br /> (2:56:15) &#8211; Talking to Claude<br /> (3:12:47) &#8211; Prompt engineering<br /> (3:21:21) &#8211; Post-training<br /> (3:26:00) &#8211; Constitutional AI<br /> (3:30:53) &#8211; System prompts<br /> (3:37:00) &#8211; Is Claude getting dumber?<br /> (3:49:02) &#8211; Character training<br /> (3:50:01) &#8211; Nature of truth<br /> (3:54:38) &#8211; Optimal rate of failure<br /> (4:01:49) &#8211; AI consciousness<br /> (4:16:20) &#8211; AGI<br /> (4:24:58) &#8211; Chris Olah &#8211; Mechanistic Interpretability<br /> (4:29:49) &#8211; Features, Circuits, Universality<br /> (4:47:23) &#8211; Superposition<br /> (4:58:22) &#8211; Monosemanticity<br /> (5:05:14) &#8211; Scaling Monosemanticity<br /> (5:14:02) &#8211; Macroscopic behavior of neural networks<br /> (5:18:56) &#8211; Beauty of neural networks</p>