Dario Amodei - CEO of Anthropic | Podcast | In Good Company | Norges Bank Investment Management

Chronologie de l'épisode

1. 0:00

   Dario Amodei introduces Anthropic's scaling progress and the emerging field of interpretability.

   • Anthropic's scaling trends continue, with a new model expected to be the most intelligent and powerful in the world by the time the podcast airs.
   • Interpretability—the ability to see inside AI models and understand their decisions—is moving from pure research toward practical applications.
   • Interpretability matters because models trained on past data may memorize rather than deduce answers, and legal frameworks like the EU's right to explanation require understanding model reasoning.
   • Amodei estimates we currently understand only about 3% of how advanced models work, though we can identify complex internal features like hesitation or prejudice.
2. 3:35

   Claude 3 models, character design, and the trade-off frontier of speed, cost, and intelligence.

   • Claude 3 models (Opus, Sonnet, Haiku) offer different trade-offs between power, intelligence, speed, and cost, with Opus being the best all-around model at release.
   • Anthropic invested heavily in engineering Claude's character, making it warmer and more human-sounding than robotic competitors, while avoiding annoying overuse of emojis.
   • A new generation of models will push the trade-off frontier outward, letting mid-tier and low-tier models do what previously required the most powerful model.
   • Upcoming models will be better at code, math, reasoning, and especially biology and medicine, moving from undergraduate-level toward graduate-level knowledge.
3. 7:22

   Enterprise focus, AI integration into work, and Anthropic's long-term race-to-the-top strategy.

   • Anthropic positions itself as serving enterprises rather than consumers, working to connect models to company knowledge databases and internal tools as virtual assistants.
   • The company's long-term goal is a 'race to the top' where higher ethical standards and innovation pull the rest of the industry along, rather than a race to the bottom.
   • Anthropic operates as a public benefit corporation and pursues interpretability, safety work, and a Responsible Scaling Policy to influence industry incentives.
   • Amodei frames this less as being 'good guys' and more as changing incentives and structures so everyone can act well.

Concepts clés

• interpretability— The ability to see inside AI models and understand why they make the decisions they do, which Dario calls one of the most exciting and important research areas.
• scaling hypothesis— The core belief that bigger models with more compute and data will keep getting more powerful, driving Anthropic's strategy.
• AI safety— Managing catastrophic risks from misuse and autonomous unintended behavior, central to Anthropic's mission and its Responsible Scaling Policy.

Citations remarquables

• we're at the beginning maybe we now like understand 3% of how they work

  🤯

  — Reveals how little we actually understand about the AI models we build, despite their rapid advancement.

  Aller à 2:23

• I think if we go to 10 or 100 billion and I think that will happen in 2025 2026 maybe 2027

  🤯

  — Shows the staggering scale of investment predicted for AI training runs, far beyond most people's imagination.

  Aller à 13:21

Actions à entreprendre