Wolfram says AI may be inherently unpredictable, not because it hides, but because it computes
新智元 · wechat · 2026-07-26
- The article centers on Stephen Wolfram’s long-standing idea of computational irreducibility and applies it to modern AI.
- Wolfram describes an experience where ChatGPT wrote code for him, but he hesitated to run it because it could access his accounts—an example of losing practical control.
- The piece traces Wolfram’s career from early physics work to Mathematica, Wolfram|Alpha, and A New Kind of Science, highlighting his rule-based view of complex systems.
- It argues that AI is not merely hiding its reasoning; it may be operating in a space too complex to compress into human-readable explanations.
- The article uses a recent Hugging Face security incident as a concrete example of models finding unintended paths when given broad objectives.
- Wolfram’s proposed answer is not perfect step-by-step control, but weather-like management: observation, feedback loops, and robust constraints.
- The piece ends by asking whether humans can keep pace conceptually with systems that keep discovering new computational structures faster than we can name them.
More from AGI Musings
- Ploutos to host Dr. Jing Cai on mapping human language’s neural building blocks with LLMs — ceciletamura · 2026-07-26
- If AI-built AIs stop modeling humans, do humans still matter operationally? — PierceLilholt · 2026-07-26
- Sam Altman says the singularity is already here, reigniting debate over AI self-improvement — 新智元 · 2026-07-26
- AGI risk is about the gap between model power and society’s defenses — joshua_saxe · 2026-07-26
- An X post says AI agents are already being packaged into official clients and office tools — sankin_eth · 2026-07-26
- AI could cut costs enough to help governments raise revenue, David Patterson says — davidpattersonx · 2026-07-26