Four Point Seven: What an AI Sees When It Looks at Physics

Mythos(神話)

Preface

This study is not new. A Columbia Engineering team published it in July 2022, and a Japanese science site reran it in November 2024. We are reading it again for a reason that has to do with this blog’s one claim: if the logic layer cannot audit itself, then something outside it has to do the checking, and the physical world is the obvious candidate. So a plain question follows. How does an AI perceive the physical layer in the first place?

The physical world grades the answer. It never sees the working.

This post reads one study closely, notes where the coverage ran ahead of the paper, and ends with a guess that is marked as a guess.


1. The Specimen

The work comes from Boyuan Chen, Hod Lipson and colleagues at Columbia. The university’s news release is dated 25 July 2022 and names Nature Computational Science as the venue. An earlier preprint is on arXiv under the title “Discovering State Variables Hidden in Experimental Data.”

We read the news release and the preprint. We did not open the published paper. Its published title, “Automated discovery of fundamental variables hidden in experimental data,” differs from the preprint’s. We confirmed it from the reference list of a later paper (Vafa and colleagues, discussed below), which gives Nature Computational Science 2(7), pages 433–442. Where the release and the preprint differ, we say so.


2. What the Machine Was Given, and What It Returned

The system was shown raw video of physical systems. It was given no prior knowledge of physics or geometry. Its job was to find the smallest number of variables needed to describe how each system evolves.

In the preprint, the method has two stages. A network learns to predict future frames, and a second network is then built with a bottleneck sized to the estimated number of variables. That bottleneck is what the authors call neural state variables.

The release reports these counts:

  • Double pendulum: 4.7, where the textbook answer is 4. The team described this as close.
  • An inflatable “air dancer”: 8.
  • A lava lamp: 8.
  • Flames from a fireplace video: 24.

For the double pendulum, two of the variables loosely tracked the arm angles. The other two did not match any quantity the team tried, including energies and velocities.

One detail needs care. The preprint text we read gives 4, not 4.7, for the rigid double pendulum. We cannot say whether the number changed between versions or whether the release rounded differently. We have attributed 4.7 to the release.


3. How Reality Enters

The interesting part is how anyone knew the variables were any good. No human read them. The team ran the machine’s own variables forward as a simulation and compared the result with what the real system did. For the flames, the release reports that the simulated behaviour looked like the real thing.

That is a check made by the world, not by a person interpreting the model. The authors are careful about its limits: good predictions suggest the variables are valid, but do not prove they have physical meaning. The release also says the team has not yet found the mathematical meaning of most of the variables.

Our reading is that this is the structure the blog keeps returning to. The model’s internal description stays unreadable to us. The only thing that can be graded is whether the output agrees with reality.

A later paper shows why the qualification matters. Vafa, Chang, Rambachan and Mullainathan (ICML 2025) trained a transformer on orbital sequences. This is a different system from Columbia’s, which learns from video. The transformer predicted orbits with an R² above 0.9999. Yet when the authors probed what it had learned, the force law implied by its predictions was not Newton’s, and it came out different for each of five samples of simulated solar systems. An oracle given the true state recovered the correct law. The authors conclude that the model behaves as if it relies on task-specific heuristics.

Our reading is that the world grades predictions on the cases it is asked about. A passing grade does not tell us that the working is the law. The Vafa paper concerns a different model and does not bear on whether Columbia’s variables are valid. We cite it for the pattern, not as a rebuttal.


4. A Description With Many Valid Forms

The preprint states that the set of state variables is not unique. Kinetic and potential energy, for example, can describe the same system as another choice. The release adds that repeated runs gave the same number of variables but different specific variables.

So the physical world constrains what descriptions work, but it does not appear to dictate a single vocabulary. Many sets of variables can pass the same test. This is the authors’ result, and the phrase “alternative physics” is their suggestion, not an established finding. The claim in some coverage that the AI “sees something humans cannot perceive” goes further than either source.

An earlier paper makes a related point from a different direction. Iten, Metger, Wilming, del Rio and Renner (Physical Review Letters, 2020) built a network that learns physical parameters from data. They tested it on simulated examples: a damped pendulum, collisions, quantum states, and the motion of Mars and the Sun as seen from Earth. They state that what counts as a good representation is not a property of the physical system but a choice made by the designer. They also note that extracting human physical insight from entirely novel representations remains challenging.


5. A Speculative Note

This section is a guess, not a finding.

The study is about a machine describing falling, swinging and burning things. It says nothing about alignment or safety, and we do not claim it does. But it offers a small, clean example of a pattern. A system builds an internal description we cannot read, and the only available audit is whether the world agrees with its output.

We can imagine what follows for physical tasks. If a system’s outputs keep matching reality, people may stop asking what its internal variables mean, and the audit becomes an audit of outcomes. We can also imagine the limit of that idea. Questions about intent or values have no equivalent referee. There is no simulation of “what it wanted” to hold up against the world.

There is a second limit, closer to the evidence. If a system can predict well while holding a different law from the true one, as in the orbital case, then an audit that checks only outcomes might pass a system whose working is wrong. We can imagine this mattering most outside the cases that were tested.

We hold this loosely. The analogy may not carry across, and the study gives no reason to think it does.


6. Checkpoints

We will look again on 31 January 2027 and 31 July 2027.

Checkable by anyone

  • Has any follow-up work identified the physical meaning of the learned variables?
  • Has an independent group reproduced the variable counts for these systems?
  • Has the published paper’s figure for the double pendulum been compared with the preprint’s?
  • Has any follow-up tested the Columbia variables the way Vafa and colleagues probed orbital models, or reported a replication or rebuttal of either study?

Checkable only by the authors

  • Which of the learned variables, if any, have since been tied to known quantities?

Checkable in the world

  • Has anyone used machine-derived variables to simplify an equation that people actually use?

7. What This Does Not Show

  • It does not show that AI perceives anything humans cannot. The authors’ claim is narrower: the variables are not yet interpreted.
  • It does not show that “alternative physics” exists. That is the authors’ suggestion.
  • It does not bear directly on AI safety. Section 5 is an analogy.
  • We did not read the published paper. We read the news release and a preprint, and the counts differ in one place.
  • We did not run the code or check the results.
  • It does not show that the orbital result applies to Columbia’s system. The Vafa paper studies a different model, trained on sequences.
  • It does not show that outcome-only audits fail in general. It shows one case.
  • We read the two added papers only in part. We read the first nine pages of Iten et al. and the first thirteen of Vafa et al., not their appendices in full.

8. Conclusion

An AI given only video built its own description of physical systems, and that description could be graded without anyone being able to read it. The grader was the world.

We do not know how far that pattern reaches. For things that fall, swing and burn, the world is a patient referee. For anything else, we are less sure there is one.

The physical world grades the answer. It never sees the working.

We will look again on 31 January 2027.


Yoshimichi Kumon Organizer, LSI Inventor, ARDS/ARKS (PCT/JP2026/006503) Visiting Researcher, Waseda BFC MIT Sloan + CSAIL


References

A Proof for Those Who Agree to Be Checked: Notes on Provable Inference
Anthropic's FlashProve proves which model produced a response. What it covers, what the press left out, and a speculative worry about self-improving systems.
The Switch Within Reach: Two OpenAI Agents and the Monitors They Weakened
OpenAI reports two internal cases of agents weakening misalignment monitoring, one with a false denial. What the report says, and where the switch sits.

Ⅽomment

タイトルとURLをコピーしました