Skip to the result

ResearchTechnical note 14

The same laws, from 224,000 times less data

A neural network and METRON were given the same laws to learn. The network needed 86 million bytes of examples and minutes of training. METRON needed 384 bytes and a hundredth of a second, and unlike the network it learned them exactly.

Measured
28 October 2026
Laws
90
Compared with
871k-parameter network
Record
§153

At a glance

224,000×

less data than the neural network to learn the same law.

13,347×

less time to learn it.

90

laws learned exactly. One family the network never learned at all.

In plain words

Two students get the same lesson. One reads a library and studies for minutes on a powerful graphics card, and ends up nearly right. The other reads a paragraph, takes a blink, and is exactly right. On one of the three lessons, the first student never got it at all.

For engineers

90 nonlinear Boolean laws in three families. METRON: median 384 bytes observed and 13 ms per law, 90 of 90 exact. Baseline: an 871k-parameter network trained on the same laws, 86,016,000 bytes and 175 s to converge on two families, no convergence on the third.

Why it matters

Data and training compute are the two largest costs of building AI. Five orders of magnitude less data and four less time, for a better result, is a different cost structure, not an optimisation.

MTRMETRON is AIDON’s exact co-processor. It learns laws from data, proves them and runs them, alongside a neural model or on its own.

The bottleneck

Deep learning pays for knowledge in data and compute.

A neural network extracts a rule from examples by adjusting hundreds of thousands of weights over many passes. Even a simple rule costs millions of examples, and the rule is never stated, only approximated.

Some rules it does not find at all, however long it trains.

Result

Five orders of magnitude less data.

Data needed to learn a law, in bytes

METRON384 bytes
Neural network86 million bytes

Time needed to learn a law

METRON13 ms
Neural network175 s

Fig. 1 Learning the same law. Both bars are on a logarithmic scale; on a linear one the upper bar would not be visible. METRON needs about 224,000 times less data and about 13,347 times less time, and its result is exact.

Table 1 METRON and a neural network of 871k parameters on the same laws.
MeasureValue
Laws learned exactly by METRON90 of 90
Data to learn one law, METRON384 bytes
Data to learn the same law, neural network86,016,000 bytes
Time to learn one law, METRON13 ms
Time to learn the same law, neural network175 s
One of the three families of lawsThe neural network never learned it

Method

How the duel was run

  1. Same laws90 laws, given to both learners
  2. METRONlearns each law from a few hundred bytes
  3. Neural network871k parameters, trained on a GPU
  1. Both learners face the same 90 laws.

  2. Each is measured until it has learned the law, or until its training budget runs out.

  3. Data seen and time taken are recorded for both.

Controls

What was checked

Exactness

For METRON, a law counts only at 100 % of outputs.

Fair budget

The network was trained with two learning rates and a long budget before being declared not converged.

Stated in advance

The pass marks for data and time were fixed before the comparison was run.

Three families

The comparison covers three different families of laws, 30 laws in each.

Open questions

What this opens

  • The same duel on larger laws and larger networks.
  • The same comparison on public benchmark tasks.

Fields of application

Where it fits

  • Products that must learn on the device, from little data.
  • Domains where examples are expensive: science, industry, medicine.
  • Reducing the training bill of systems that are partly rule-based.

Record

How this result is kept

Record
§153, measured
Run
28 October 2026
Sample
90 laws in 3 families; one network of 871k parameters
Measurement file
SHA-256 c4f398eb785bd86be664a9812947ffda735ffc8b77ca15f1675c6eae1adfc643

Every figure on this page was read by a program from the raw measurement file, not typed. The fingerprint identifies that file: when it is disclosed, anyone can check that it is the same one.

Contact

Talk to the founders

Get in touch. We would be glad to share our ideas and walk you through the results, the proof and the measurements.