An AI system turns a table of measurements into a short equation. The expression is readable, the predictions match the supplied rows and the result looks ready for a science headline. Before calling it a new law, ask what was actually demonstrated: recovery of a known formula, a useful predictive relationship or evidence about a physical mechanism.
Symbolic regression is an unusually interesting part of AI research because its output can be inspected as mathematics. That visibility gives readers something concrete to question. It also leaves an important distinction intact: an equation’s agreement with observations and the evidence for its interpretation are separate matters.
What AI Feynman set out to do
Silviu-Marian Udrescu and Max Tegmark’s AI Feynman paper, published in Science Advances on 15 April 2020, describes finding symbolic expressions from tables of numbers. Its method combines neural-network fitting with techniques that exploit structure, including physical units, symmetry and separability, to break harder problems into simpler ones.
The authors report recovering all 100 equations in their test set drawn from the Feynman Lectures on Physics. This is an important formula-recovery result on a specified benchmark. The target equations were already known; the experiment should not be described as the discovery of 100 previously unknown physical laws.
The distinction does not diminish the technical achievement. Recovering an explicit relationship from data is a concrete task. For a reader assessing a claim, the useful next questions concern the source of the data, the available variables and what would count as success beyond the benchmark.
Two formulas can agree with the supplied rows
Here is a deliberately small, original illustration. Both x and y are dimensionless numbers. The supplied table has only two rows: x = 1, y = 2, and x = 2, y = 4. This is an algebra example, not a physical dataset or a test of AI Feynman.
The formula y = 2x matches both rows. So does y = x² − x + 2. At x = 1 the second expression gives 1 − 1 + 2 = 2; at x = 2 it gives 4 − 2 + 2 = 4. Their agreement on those rows does not make the formulas identical.
At x = 3, the first formula predicts 6 and the second predicts 8. A new observation there could distinguish these two candidates. That is a more specific proposal than collecting another row without considering where their predictions differ.
Even a third matching point would not establish a unique unrestricted formula. Consider y = 2x + (x − 1)(x − 2)(x − 3). It agrees with y = 2x at x values 1, 2 and 3, because one factor of the added term is zero each time. At x = 4, however, it gives 14 rather than 8.
The arithmetic was checked for this article. It demonstrates a limit of what these particular observations settle. It does not show that a symbolic-regression method must choose the longer expression, or that additional measurements are pointless. A choice of allowed formulas and other justified constraints changes the problem.
Simplicity is a stated selection rule
The follow-up AI Feynman 2.0 paper at NeurIPS 2020 explicitly addresses accuracy and complexity together. It seeks Pareto-optimal formulas: candidates offering the best accuracy at a given complexity. The research also extends the use of structure identified from a neural-network fit.
For the toy table, preferring y = 2x is an understandable choice of a simpler candidate. The editorial lesson is to state that preference and the allowed search space. A method that selects a compact expression has made a model-selection decision; the size of the expression is not itself an experimental observation of nature.
Ask how complexity is measured and whether the candidate list includes the relationships of interest. If a search permits only straight lines, its best straight line answers a narrower question than a search over a much wider family. Describe the question actually posed before interpreting the winning equation.
Keep the physical interpretation beside the mathematics
Now imagine a fictional laboratory team receiving a candidate equation for a sensor. The team wants to know whether it remains useful at temperatures outside the range used to generate it. For this proposed review, preserve the original measurement range and identify the new conditions to be examined.
The team could plan observations where plausible candidate equations disagree, while respecting the equipment’s operating limits. Record the measurement procedure and the result separately from the prediction. This is an editorial proposal for evaluating a claim, not a validated protocol for a particular instrument.
For this fictional case, a temperature label in a spreadsheet would not by itself establish how temperature was measured or controlled. The interpretation needs that record. If a headline claims a mechanism, ask which evidence addresses that mechanism rather than merely repeating that predictions were close.
The dimensionless algebra example above deliberately avoids assigning a real quantity to x or y. Do not reuse its formulas as a model of the sensor. An illustration of competing expressions supplies no evidence about the behaviour of an actual device.
Give a research claim a reproducible description
For independent creators explaining research, or business owners reviewing a proposed analytical tool, a concise evidence note can prevent the headline from outrunning the experiment. Ask for the following details before deciding what the result supports.
- The variables supplied, their meanings and how the data were produced.
- The candidate formula, its allowed inputs and the range examined.
- The criteria used to select it, including the accuracy and complexity measures.
- The observations reserved for checking the result and the errors found.
- The additional evidence needed for any claimed physical interpretation.
These questions do not turn every reader into a specialist. They make the claim specific enough to discuss with one. A summary that says the system recovered a known benchmark equation is useful. A different summary would be needed for a new laboratory relationship with independent experimental support.
There is a related distinction in MaryChuks.com’s article on AI geometry proofs: proposed constructions and the deductions that establish a proof perform different roles. In equation discovery, keep the candidate expression and the evidence used to assess its physical meaning equally visible.
Symbolic regression gives AI research a valuable kind of output: a formula people can inspect, compare and test. Let that inspectability do its work. State the benchmark, show the selection rule and describe what further observations establish. A readable equation is an invitation to a precise scientific question.
Sources
Primary papers checked on 31 October 2026. The algebra examples are original and their arithmetic was checked locally. The sensor review is hypothetical; no AI Feynman code, laboratory instrument or model-training experiment was run.
- Udrescu and Tegmark: AI Feynman: A physics-inspired method for symbolic regression — Science Advances, 15 April 2020.
- Udrescu et al.: AI Feynman 2.0: Pareto-optimal symbolic regression exploiting graph modularity — NeurIPS 2020.
Discover more from Marychuks.com AI, Psychology, Business & CreativeVerse
Subscribe to get the latest posts sent to your email.
One thought on “AI Equation Discovery: Ask What the Formula Establishes”