Compression Is Understanding: Two July Papers That Prove AI Gets Smarter by Compressing Physics
Compression Is Understanding: Two July Papers That Prove AI Gets Smarter by Compressing Physics There's a thesis I keep coming back to: AI doesn't understand by accumulating information — it understands by compressing it. A model that can reconstruct a concept from fewer bits has, in a real sense, captured its structure. Compression forces you to find what matters and discard what doesn't. July…
Two recent papers suggest artificial intelligence achieves greater intelligence by compressing physics.
The first paper, PhiZero from the Chinese Academy of Sciences, builds a world model around what they call a "physical language." It predicts what happens next in a video using a compact discrete vocabulary of physical language tokens. This reduces the 44,800 visual tokens of a 33-frame, 4-second clip to just 256 tokens, a 175x reduction.
A reasoning component then predicts the next token sequence, allowing the model to reason about the world's evolution in compressed form. The compressed tokens can be re-rendered with different appearances, transferred across different embodiments, and moved from simulation to real-world visuals.
The second paper comes from Zhang Hongliang, a researcher at Fudan University recognized by MIT's TR35 in 2026. He applies AI to predict the safety margins, structural integrity, and service life of structural materials for next-generation advanced nuclear reactors. By learning to compress the mechanisms of irradiation-induced interface evolution at the micro/nano scale, Zhang can predict decades of material behavior computationally instead of waiting for decades of heavy testing. Like PhiZero, Zhang compresses the relevant information while discarding what doesn't matter.
Both papers share a common structure: identify the expensive thing to compute (pixels or decades of irradiation damage), find what truly drives the state transitions (motion, contact, interaction or microstructural evolution), and compress that essential information. The discarded elements are appearance, lighting, and brute-force physics simulation. The payoff is dramatically reduced token counts, transferable across embodiments, and predictions without waiting for expensive real-world testing.
Within the realm of embodied AI, PhiZero's compressed world model directly aligns with the need for robots to understand physics before acting. Cross-embodiment transfer solves the data scarcity problem by allowing a robot to plan manipulation, locomotion, and interaction based on a compressed state-transition structure. The model-based RL theory that PhiZero exemplifies is precisely the focus of Berkeley's CS 285 lectures.
The discrete symbolic space of physical-language tokens is also more verifiable, with conservation laws, contact constraints, and causality that can be checked against the tokens.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — it may contain errors, so check the original before relying on it.
