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 2026 handed me two independent, concrete demonstrations of exactly this — one from a world-model paper, one from a 35-year-old scientist on the MIT TR35 list.
Case 1: PhiZero — Physics as a Language, 175x Fewer Tokens
On August 7, the Chinese Academy of Sciences' Institute of Automation released PhiZero (arXiv:2607.28624), a world model built around what they call a "physical language."
The setup: predicting what happens next in a video. The naive approach predicts pixels. But pixels mix two very different kinds of information:
- Appearance — texture, lighting, material. Huge volume, mostly static.
- State transitions — how objects move, interact, deform. Small volume, entirely dynamic.
When you predict pixels directly, the dynamic structure gets drowned out by appearance noise.
PhiZero's fix is a clean act of compression:
Current state + action intent → physical language tokens → future evolution
A Physical Language Tokenizer compresses video changes into a compact discrete vocabulary (~25K symbols). A 33-frame, 4-second video becomes 256 physical-language tokens — versus 44,800 continuous visual tokens with a standard VAE. That's a 175x reduction.
Then a Reasoner (initialized from Qwen3-VL-4B) predicts the next token sequence — i.e., it reasons about how the world will evolve in the compressed space — and a decoder renders the future video from the first frame plus the predicted token sequence.
The key insight: the model thinks about physics before it draws pictures. It predicts state transitions symbolically, then renders. Understanding happens in the compressed space — the 175x reduction isn't a lossy hack, it's the actual mechanism by which the model isolates what changes from what just is.
And the compression generalizes: the same token sequence can be re-rendered with different appearances (rice → liquid → sticky material), transferred across embodiments (human full-body motion → Unitree G1 humanoid; human hand → dexterous hand, no paired training data needed), and moved from simulation to real-world visuals.
Case 2: Zhang Hongliang — AI for Nuclear Materials, MIT TR35 China 2026
On July 25, MIT Technology Review released the 2026 "35 Innovators Under 35" (TR35) China list in Shanghai. Among the five Shanghai scientists was Zhang Hongliang, a young researcher at Fudan University.
His work: applying AI to nuclear materials R&D — specifically, predicting the safety margins, structural integrity, and service life of structural materials for next-generation advanced nuclear reactors.
Why this matters: inside a reactor, materials are bombarded by neutron irradiation, which degrades their microstructure over decades — embrittlement, swelling, cracking. The traditional way to understand this is decades-long test irradiations, or physics simulations with enormous computational cost.
Zhang's contribution, per the TR35 citation:
- Clarified, at the micro/nano scale, the mechanisms of irradiation-induced interface evolution in materials
- Discovered that metals at micro/nano scale exhibit ultra-high ductility and entirely new deformation mechanisms that overturn conventional understanding of brittleness
- Uses AI to model these evolutions so that decades of service-life behavior can be predicted computationally instead of measured over decades
Same structure as PhiZero: the physics is too expensive to compute exhaustively, so AI learns to compress the evolution rules — the part that matters — and predict the outcome directly.
What These Two Cases Share
Strip the domains away and both are the same move:
| PhiZero | Zhang Hongliang | |
|---|---|---|
| Domain | Video world models | Nuclear materials |
| The expensive thing | Predicting future pixels | Predicting decades of irradiation damage |
| The compression | Video → 256 discrete physics tokens | Irradiation physics → learned evolution model |
| What's kept | State transitions (motion, contact, interaction) | Microstructural evolution rules |
| What's discarded | Texture, lighting, appearance | Full brute-force physics simulation |
| The payoff | 175x fewer tokens, transferable across embodiments | Service-life prediction without waiting decades |
The boundary between "compressible" and "incompressible" is where understanding lives. PhiZero keeps appearance out of the bottleneck because appearance doesn't drive dynamics. Zhang keeps the microstructural evolution rules because those drive failure. Both found — empirically — what their domain is, by finding what can be thrown away.
This is the same argument I made in my book manuscript: compression has limits because time itself carries information — the gaps between events are not empty, they are information. But within those limits, the amount of genuine structure you can extract is exactly proportional to how well you can compress. These two papers are existence proofs.
Why This Matters for Embodied AI
The PhiZero result in particular lands directly on the embodied-AI stack:
- World models are the L3 layer for robots — "understand physics before acting" is precisely what a robot needs to plan manipulation, locomotion, and interaction.
- Cross-embodiment transfer solves the data scarcity problem — human video → robot policy, no paired data required, because the compressed state-transition structure transfers even when the body doesn't.
- Model-based RL is the theoretical foundation — the PhiZero Reasoner (predict next state tokens) is world-model RL in disguise, which is exactly lectures 15–16 of Berkeley CS 285.
And the QC angle: a discrete symbolic space is verifiable. Physical-language tokens can be checked against conservation laws, contact constraints, and causality — a natural home for validation layers that catch the hallucinations a pixel-space model would hide.
The Takeaway
Two events in July 2026, one from a Chinese research institute, one from a 35-year-old on a prestigious list, both demonstrating the same principle:
AI understands the world by compressing it — finding the structure that survives compression is the actual act of understanding.
The 175x token reduction isn't a trick. It's what understanding looks like when you measure it in bits.
Related: I Tested DeepSeek V4 Flash's Hallucination Rate on the Release-Day API — 94% → 0%
Top comments (0)