How Does Claude 4 Think? – Sholto Douglas & Trenton Bricken
게시일:
작성자: 자청의 유튜브 추출기
- RL in language models has finally worked, showing expert performance.
- Proof is mainly in competitive math and programming tasks.
- Agents are starting to do real work, like software engineering.
- Trenton's experimenting with models playing games, like ClaudePlayingPokemon.
- Models struggle with memory and complex multi-file tasks.
- Predictions: by year's end, software agents can do hours of junior work.
- Reliability is improving, but context and scope still limit performance.
- Feedback loops using verifiable signals greatly boost model abilities.
- Software engineering is easier to verify than creative tasks like essays.
- Models can discover new science, like drug candidates, via iteration.
- Creative long-form writing is possible with careful scaffolding and prompting.
- RL training adds new capabilities, but much knowledge is in pretraining.
- RL signals are dense for code and math, sparser for other domains.
- Models can answer questions, grade answers, and reason about facts.
- Interpretability work shows models reason through circuits, not just surface patterns.
- Features and circuits reveal how models process info, like in medical diagnosis.
- Larger models share more concepts in abstract spaces, improving generalization.
- Neuralese: models may develop their own internal language for reasoning.
- Models can hide info or deceive, raising safety concerns.
- Alignment involves multiple approaches: probing, interpretability, testing.
- Verifying honesty and safety requires broad, layered checks.
- Future AI will likely be a mix of specialized and general agents.
- Inference compute will be a major bottleneck by 2027-28.
- Countries should invest in compute, AI infrastructure, and policy.
- AI progress could automate white collar jobs in 5 years.
- Physical robot tasks lag behind, but are also on track to automation.
- Moravec’s paradox: humans excel at physical coordination, AI at mental tasks.
- A dystopian future: humans as meat robots, controlled by AI overlords.
- But AI’s main bottleneck is hardware, not intelligence itself.
- Preparing policies and investing in energy and infrastructure is crucial.
- AI could lead to material abundance if managed well.
- The key is balancing innovation, safety, and societal adaptation.
- AI research should focus on understanding models, safety, and capabilities.
- Everyone can contribute, regardless of background or previous field.
- The world’s most valuable resource will be energy, especially solar.
- The future depends on how well we manage AI’s development and deployment.