Breaking

Self-Improving LLMs Cannot Automatically Solve Physical AI, Says World Mechanics CEO

October 20, 2026 at 07:28 EDT

Foundation Models
AI Agents
Research & Papers

Sonia Joseph, co-founder of World Mechanics, challenges the prevailing assumption that recursive self-improvement in language models will automatically translate into solving real-world physics problems. In a recent interview with The Information, Joseph argues that the physical world operates under fundamentally different constraints than the digital domain where large language models excel.

The Information · October 20, 2026

Language Models Alone Cannot Solve Physical AI

Sonia Joseph, World Mechanics co-founder, argues that recursive self-improvement in LLMs lacks the architectural foundations to reason about physical constraints, time, and causality in the real world.

The Core Limitation

LLMs operate through next-token prediction: they guess the next word in a sequence. This is fundamentally different from physical reasoning, which requires predicting continuous state changes, maintaining temporal awareness, and respecting physical laws over time scales ranging from milliseconds to hours.

Architectural Constraints of LLMs vs. World Models

Key LLM constraints vs. world models

Domains World Mechanics targets (climate, astrophysics, power grids)

Physics Emergence in Neural Networks

Physical concepts (velocity, direction, permanence) emerge sharply in mid-layer representations, distributed across many neurons as population codes.

Why Self-Improvement Alone Is Insufficient

The LLM Path

Scaling + self-refinement
Next-token prediction
No temporal grounding

World Models Path

Interpretability-first design
Explicit state + dynamics
Physical constraint awareness

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.