Demis Hassabis：通往 AGI 之路与缺失的拼图

Demis Hassabis: The Path to AGI and Missing Pieces

杰米斯·哈萨比斯 Demis Hassabis · Y Combinator · 2026-04-29 · 约 41 分钟 · 原视频 ↗

本期速览 · Overview

Demis Hassabis 讨论当前 AI 范式、持续学习与记忆等缺失组件，以及通往 AGI 的路径。

Demis Hassabis discusses the current AI paradigm, missing components like continual learning and memory, and the path to AGI.

要点 · TL;DR

AGI 需要持续学习、长期记忆和超越当前大模型的推理能力。
AGI needs continual learning, long-term memory, and reasoning beyond current LLMs.
智能体是 AGI 的关键但还很初级，需要真正的记忆创新。
Agents are key to AGI but still primitive; true memory innovation is needed.
像 Gemini 这样的多模态模型对理解物理世界至关重要。
Multimodal models like Gemini are essential for understanding the physical world.

核心观点 · Key points

持续学习、长期推理和记忆仍未解决，是 AGI 所必需的。
Continual learning, long-term reasoning, and memory are still unsolved and required for AGI.
当前的预训练、RLHF 和思维链等技术将成为最终 AGI 架构的一部分。
Current techniques like pre-training, RLHF, and chain of thought will be part of the final AGI architecture.
智能体对于实现 AGI 至关重要，但我们才刚刚起步。
Agents are essential for achieving AGI, but we are just at the beginning of their development.
蒸馏技术可将前沿模型能力压缩到更小、更高效的模型中，用于边缘部署。
Distillation allows packing frontier model capabilities into smaller, efficient models for edge deployment.
像 Gemini 这样的多模态模型对于理解物理世界和机器人技术至关重要。
Multimodal models like Gemini are crucial for understanding the physical world and robotics.
将 AI 与科学结合的深度技术领域具有防御性，并能带来长期影响。
Deep tech areas combining AI with science are defensible and offer long-term impact.

反共识 · Contrarian takes

大上下文窗口是蛮力方法；智能体需要真正的记忆创新。
Large context windows are brute force; true memory innovation is needed for agents.
前沿模型仍会犯基础推理错误，表现出锯齿状智能。
Frontier models still make elementary reasoning errors, showing jagged intelligence.
AI 系统尚无法根据高层描述发明出像围棋这样的东西。
AI systems cannot yet invent something like the game of Go from a high-level description.
由于芯片生产等物理瓶颈，推理永远不会基本免费。
Inference will never be essentially free due to physical bottlenecks like chip production.
完整的虚拟细胞模拟大约还需 10 年，受限于数据和成像技术。
A full virtual cell simulation is about 10 years away, limited by data and imaging.
AGI 可能在深度技术旅程中途出现，需要战略性考量。
AGI may appear in the middle of a deep tech journey, requiring strategic consideration.

本期章节 · Chapters（共 26）

0. 德米斯·哈萨比斯简介 Introduction of Demis Hassabis
1. 当前范式与AGI缺失环节 Current paradigm and missing pieces for AGI
2. 持续学习与记忆挑战 Continual learning and memory challenges
3. 上下文窗口限制与智能体系统 Context window limitations and agentic systems
4. 深度强化学习与智能体哲学 Reinforcement learning and agent philosophy at DeepMind
5. 从游戏到通用世界模型 From Games to General World Models
6. 蒸馏与小模型 Distillation and Smaller Models
7. 对开发者生产力的影响 Impact on Developer Productivity
8. 智能体的上下文与持续学习 Context and Continual Learning for Agents
9. 推理差距与思维链 Reasoning Gaps and Chain of Thought
10. 智能体能力与炒作 Agent Capabilities and Hype
11. 自主与增强智能体 Autonomous vs. Augmented Agents
12. AI工具与创造力 Creativity with AI tools
13. 开源与开放权重 Open source and open weights
14. 多模态Gemini优势 Multimodal Gemini and its advantages
15. 推理成本与未来可能 Inference cost and future possibilities
16. AlphaFold 3与虚拟细胞 AlphaFold 3 and virtual cell
17. 虚拟细胞与数据挑战 Virtual Cell and Data Challenges
18. 五年内最具变革科学领域 Most Transformative Scientific Domain in 5 Years
19. 普罗米修斯性质与责任使用 Promethean Nature and Responsible Use
20. 给AI科学初创公司的建议 Advice for AI for Science Startups
21. AI的信念与热情 Belief and Passion in AI
22. AlphaFold式突破模式 Pattern for AlphaFold-style Breakthroughs
23. AI用于科学推理 AI for Scientific Reasoning
24. 结束致谢 Closing Thanks
25. 给年轻建设者的建议 Advice for young builders

阅读全文双语转录 →