Skip to the result

ResearchTechnical note 8

A 32-bit multiplier, assembled from learned laws

Multiplication is the operation on which language models fail first and worst. METRON learned it in 32 steps, each one exact, and the steps together multiply any two 32-bit numbers without error.

Measured
9 November 2026
Operation
x · y mod 232
Learned steps
32
Record
§178

At a glance

65,536

products right when every pair of 8-bit numbers is tried, with none left out.

8,192

random pairs of 32-bit numbers multiplied exactly.

0

bits lost when the learned multiplier is compiled to plain bitwise code.

In plain words

Long multiplication at school is a few simple steps repeated. METRON learned each of those steps from examples, got every one exactly right, and chained them into a multiplier that does not make mistakes.

For engineers

Multiplication mod 232 composed from 32 learned stages; each stage is exact on all of its output bits. Verified on 8,192 random pairs, exhaustively on 8×8 bits, on 8,192 16×16 pairs and on 121 edge cases; the compiled form is bit-identical.

Why it matters

A multiplier is the heart of every processor and every accelerator. Showing that one can be learned from data and still be exact is the step between a learning system and a chip.

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

Multiplication is where approximate systems break.

Every output bit of a product depends on many input bits at once. A model that predicts digits loses the thread almost immediately: in our own test, a public language model was at zero from 6 digits.

Learned as a single function, multiplication is out of reach for METRON too. The lowest bits come out; the rest do not. The result here comes from how the problem is learned.

Result

Thirty-two exact steps. One exact multiplier.

Output bits of the product that are exact, out of 32

METRON32 / 32
Same data, no structure4 / 32
Scrambled control11 / 32

Fig. 1 32-bit multiplication. Learned as one black box, only the 4 lowest bits come out right. Learned the way METRON does it, all 32 are exact.

Table 1 The multiplier assembled from learned laws, against the processor’s own multiplication.
MeasureValue
Learned steps, each exact on all 32 bits32 of 32
Random 32-bit pairs8,192 of 8,192
Every pair of 8-bit numbers, exhaustive65,536 of 65,536
16-bit by 16-bit pairs8,192 of 8,192
Special cases: zeros, ones, powers of two, limits121 of 121
Compiled to bitwise code, bits that differ0
Whole run53 s

Method

How the multiplier is built

  1. Examplesof each step of a multiplication
  2. METRONlearns each step exactly
  3. Multiplier32 steps in sequence, compiled
  1. Each step is learned from examples, and accepted only if it is exact.

  2. The steps are chained into a full 32-bit multiplier.

  3. The result is compiled to plain bitwise code and compared again, bit by bit.

Controls

What was checked

Exhaustive check

All 65,536 pairs of 8-bit operands, with no sampling.

Edge cases

121 special pairs (zeros, all ones, powers of two, the largest values), all exact.

Scrambled control

With the examples shuffled, a step reaches only 11 of 32 bits.

Compiled equals learned

The compiled multiplier and the learned one agree on every bit of every test.

Open questions

What this opens

  • Wider multipliers, already measured up to 4,096 bits in technical note 03.
  • Division and the remaining arithmetic of a processor.

Fields of application

Where it fits

  • Arithmetic units for processors and accelerators.
  • Digital signal processing, where multipliers dominate the circuit.
  • Any AI system that has to return a product, not an estimate of one.

Record

How this result is kept

Record
§178, measured
Run
9 November 2026, 53 s
Sample
32 steps; 8,192 + 65,536 + 8,192 + 121 operand pairs
Measurement file
SHA-256 c72c05a4b199a1447e1d0d7f745444bf56a831f32dfaa51cc45d02a8fdf27016

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.