Seven Mental · 心智七篇
← Knowledge Atlas · Entity

Max Tegmark

Max Tegmark:MIT 物理系教授,宇宙学家,Future of Life Institute 联合创始人,AI 表征与安全研究者
ENTITY · MAX TEGMARK · MIT PHYSICS · SPATIOTEMPORAL WORLD MODEL · FLI CO-FOUNDER

Max Tegmark

MIT cosmologist — brings physics-style quantitative methods to the study of LLM internal representations

Tegmark is a well-known cosmologist, best known for the Mathematical Universe Hypothesis and Our Mathematical Universe (2014). He is also an active voice in AI safety, co-founder of the Future of Life Institute (FLI), and an advocate of the 2017 Asilomar AI Principles. In AI interpretability he imports the quantitative methods of physics to the study of LLM internals, with a focus on discovering structured, measurable regularities inside neural networks.

AI Internal Representation Research
2023Language Models Represent Space and Time (with Wes Gurnee)
Published at ICLR 2024. Applies probing methods from physics to quantify how LLMs encode spatiotemporal coordinates — Llama-2 linear probes reach R²=0.911 (space) / 0.835 (time).
”Physicist’s View of AI”
Geometric Structure Focus
Focuses on geometry, measurability and universal regularities of neural network representations — not engineering details
Emergence & Phase Transitions
Several works on emergence and phase transitions in neural networks — transferring phase-transition theory from physics to AI
AI Safety Crossover
FLI co-founder — the bridge from interpretability research to AI safety policy
→ Spatiotemporal World Model · Wes GurneeGurnee & Tegmark (ICLR 2024)

Max Tegmark

机构: 麻省理工学院(MIT)物理系 研究方向: 宇宙学、AI 安全、LLM 内部表征、物理学与 AI 的交叉

背景

Max Tegmark 是知名宇宙学家,以”数学宇宙假说”(Mathematical Universe Hypothesis)和《Our Mathematical Universe》(2014)著称。他同时是 AI 安全领域的积极参与者,是 Future of Life Institute(FLI)的联合创始人,2017 年《阿西洛马 AI 原则》的倡导者之一。

在 AI 可解释性领域的工作

Tegmark 将物理学的量化方法引入 AI 内部表征研究,侧重于发现神经网络中的结构化、可测量规律。

Language Models Represent Space and Time(2023)

Wes Gurnee 合作,发表于 ICLR 2024。核心贡献:将物理学中的探针和线性代数方法用于量化 LLM 对时空坐标的编码能力。

详见:时空世界模型

研究视角

Tegmark 的视角更偏向”物理学家看 AI”:关注神经网络表征的几何结构、可量化性和普遍规律,而非工程实现细节。他的团队有多篇关于神经网络中”涌现”和”相变”现象的工作。

机构关联

  • MIT 物理系教授
  • Future of Life Institute 联合创始人
  • Machine Intelligence Research Institute(MIRI)顾问

References

  • sources/arxiv_papers/2310.02207-language-models-represent-space-and-time.md