x.com/HowToPrompt__/status/2105553776417608089

This sentence overstates how little information AI-Newton received.

In the paper's description of the experiment base, the authors write that "The inputs for each experiment include only the physical objects involved, geometric information, experimental parameters, and space-time coordinates". They separately say that the initial concept layer contains only space-time coordinates. Those are not the same claim.

So the system was not given only coordinates. It was also given structured information about object types, geometry, and experiment parameters. A more accurate summary would be that AI-Newton started without higher-level physics concepts such as mass or energy, but it did receive additional experiment metadata beyond space and time coordinates.

This sentence is misleading because it strongly implies that AI-Newton reproduced Einstein's path from the equivalence of gravitational and inertial mass to General Relativity.

That is not what the paper reports. The authors describe AI-Newton as a proof-of-concept implementation applied to 46 different classical mechanics experiments and say it can rediscover laws "in classical mechanics." Later, when discussing gravitational vs. inertial mass, they do not claim the system derived General Relativity. Instead they say the system "effectively indicat[es] a cornerstone of general relativity—the weak equivalence principle—from raw data alone."

Those are very different claims. Showing that two definitions of mass come out numerically equal in Newtonian-style experiments is not the same as deriving Einstein's full relativistic theory of gravity. The paper's actual scope is Newtonian/classical mechanics, not General Relativity.

This description makes the test sound like the system worked from messy real-world lab observations. The paper says something narrower.

According to the authors, AI-Newton's experiment base stores "physical experiments and corresponding simulated data generators". In the evaluation section, they add that "all test data are generated by solving differential equations and incorporating Gaussian-distributed errors."

So the system was evaluated on simulated noisy data from predefined mechanics setups, not on raw real experimental measurements collected from the physical world. That distinction matters, because simulated benchmarks are much more controlled than genuine experimental data.