Richard Socher：递归自我改进将解锁科学发现新范式Richard Socher: Recursive Self-Improvement Will Unlock a New Paradigm for Scientific Discovery
The Takeaway：任何可模拟的领域，AI最终都会解决；递归自我改进（RSI）将成为科学加速的最大解锁。
Richard Socher是AI领域被引用次数最多的研究者之一，现任Recursive联合创始人（刚完成约6.5亿美元融资），并即将出版新书《The Eureka Machine》。他指出，科学进步已明显放缓：知识从“知识体”变成了“知识迷宫”，3.4万本期刊像贴了“禁止入内”的牌子，跨学科通才几乎不可能。AI将像微积分对物理学一样，把生物学等碎片化领域重新编织起来。
核心机制是下一token预测本身就能内化世界模型：模型通过海量预测自动掌握地理、蛋白质折叠甚至化学反应。可模拟且可验证的领域（游戏、数学、编程）将最先被AI超越；生物、化学等则需要更多扰动实验数据、类器官和虚拟细胞。幻觉在探索新蛋白质时反而是特征而非bug。Eureka Machine的四大支柱是：人类知识LLM、科学测量数据、高保真模拟、真实世界机器人实验，再加上智能体群和开放式发现。
Socher强调：“Anything you can simulate, AI will solve. And before you know it, you're in this recursive self improvement loop.” Recursive将先做AI-for-AI，再把能力外溢到生命科学。他不相信硬起飞，但相信现有技术已足够在疾病、材料、能源上取得实质性突破。
The Takeaway: Anything that can be simulated, AI will eventually solve; recursive self-improvement (RSI) is the biggest unlock for accelerating science.
Richard Socher, one of the most-cited AI researchers and co-founder of Recursive (which just raised roughly $650 million), argues in his forthcoming book The Eureka Machine that scientific progress has slowed because knowledge has become a labyrinth of 34,000 journals. AI will act as calculus did for physics: weaving fragmented fields such as biology back together. Next-token prediction already embeds world models (geography, protein folding, chemistry). Domains with reliable simulators and verifiers (games, math, code) will be solved first. Hallucinations can be a feature when exploring novel proteins. The Eureka Machine rests on four pillars—LLMs of human knowledge, scientific measurements, high-fidelity simulation, and real-world robotic experimentation—plus agent swarms. Recursive will first perfect AI-for-AI research, then apply the same loop to the life sciences.
Richard Socher是AI领域被引用次数最多的研究者之一，现任Recursive联合创始人（刚完成约6.5亿美元融资），并即将出版新书《The Eureka Machine》。他指出，科学进步已明显放缓：知识从“知识体”变成了“知识迷宫”，3.4万本期刊像贴了“禁止入内”的牌子，跨学科通才几乎不可能。AI将像微积分对物理学一样，把生物学等碎片化领域重新编织起来。
核心机制是下一token预测本身就能内化世界模型：模型通过海量预测自动掌握地理、蛋白质折叠甚至化学反应。可模拟且可验证的领域（游戏、数学、编程）将最先被AI超越；生物、化学等则需要更多扰动实验数据、类器官和虚拟细胞。幻觉在探索新蛋白质时反而是特征而非bug。Eureka Machine的四大支柱是：人类知识LLM、科学测量数据、高保真模拟、真实世界机器人实验，再加上智能体群和开放式发现。
Socher强调：“Anything you can simulate, AI will solve. And before you know it, you're in this recursive self improvement loop.” Recursive将先做AI-for-AI，再把能力外溢到生命科学。他不相信硬起飞，但相信现有技术已足够在疾病、材料、能源上取得实质性突破。
The Takeaway: Anything that can be simulated, AI will eventually solve; recursive self-improvement (RSI) is the biggest unlock for accelerating science.
Richard Socher, one of the most-cited AI researchers and co-founder of Recursive (which just raised roughly $650 million), argues in his forthcoming book The Eureka Machine that scientific progress has slowed because knowledge has become a labyrinth of 34,000 journals. AI will act as calculus did for physics: weaving fragmented fields such as biology back together. Next-token prediction already embeds world models (geography, protein folding, chemistry). Domains with reliable simulators and verifiers (games, math, code) will be solved first. Hallucinations can be a feature when exploring novel proteins. The Eureka Machine rests on four pillars—LLMs of human knowledge, scientific measurements, high-fidelity simulation, and real-world robotic experimentation—plus agent swarms. Recursive will first perfect AI-for-AI research, then apply the same loop to the life sciences.
The Takeaway: Anything that can be simulated, AI will eventually solve; recursive self-improvement (RSI) is the biggest unlock for accelerating science.
Richard Socher, one of the most-cited AI researchers and co-founder of Recursive (which just raised roughly $650 million), argues in his forthcoming book The Eureka Machine that scientific progress has slowed because knowledge has become a labyrinth of 34,000 journals. AI will act as calculus did for physics: weaving fragmented fields such as biology back together. Next-token prediction already embeds world models (geography, protein folding, chemistry). Domains with reliable simulators and verifiers (games, math, code) will be solved first. Hallucinations can be a feature when exploring novel proteins. The Eureka Machine rests on four pillars—LLMs of human knowledge, scientific measurements, high-fidelity simulation, and real-world robotic experimentation—plus agent swarms. Recursive will first perfect AI-for-AI research, then apply the same loop to the life sciences.
Socher stresses: “Anything you can simulate, AI will solve. And before you know it, you're in this recursive self improvement loop.” He rejects hard takeoff narratives but believes current technology is already sufficient for major gains in disease, materials and energy.
查看原文 →查看原文 →
Richard Socher, one of the most-cited AI researchers and co-founder of Recursive (which just raised roughly $650 million), argues in his forthcoming book The Eureka Machine that scientific progress has slowed because knowledge has become a labyrinth of 34,000 journals. AI will act as calculus did for physics: weaving fragmented fields such as biology back together. Next-token prediction already embeds world models (geography, protein folding, chemistry). Domains with reliable simulators and verifiers (games, math, code) will be solved first. Hallucinations can be a feature when exploring novel proteins. The Eureka Machine rests on four pillars—LLMs of human knowledge, scientific measurements, high-fidelity simulation, and real-world robotic experimentation—plus agent swarms. Recursive will first perfect AI-for-AI research, then apply the same loop to the life sciences.
Socher stresses: “Anything you can simulate, AI will solve. And before you know it, you're in this recursive self improvement loop.” He rejects hard takeoff narratives but believes current technology is already sufficient for major gains in disease, materials and energy.