Try it

Give it a rule it has never seen

Type a function of a and b. Your browser draws 999 examples of it plus 16 new questions, and the model, which never sees your formula, reads the examples and answers every bit of every question in one pass. It also says how sure it is of each bit. Everything runs on your device; the 1,145,339-parameter model is a download of about 8 MB. You can also upload your own data as a CSV file and download its answers.

Loading the model…

The hidden rule

c =

Known weak spots:

Random rules, made up in your browser:

It can multiply a by b, but not values it has to compute first, such as (a + b) * a, where it gets the lowest few bits and guesses the rest. Rotations are new to it as well.

Numbers of small numbers (a, b under 256)

Ask your own questions too

One a, b per line, decimal or 0x…, up to 16. Random questions fill up the rest of the 16, and yours never appear among the 999 examples.

What you can write

Python-style: + - * // % **, & | ^ ~ << >>, comparisons (1 or 0), not and or, x if cond else y or cond ? x : y, and min max popcount reverse rotl rotr, and three kinds of loop: sum(f for i in range(n)) adds up f for i from 0 to n − 1; loop(n, start, f) starts with x = start and sets x = f n times; rec(n, x0, y0, f) starts with x, y = x0, y0 and sets x, y = f, x n times (the answer is x). Inside a loop, i is the step number. A loop runs at most 255 times and loops can't be nested. Numbers are unsigned and wrap at the width after every step, as in training, and division by zero gives all ones. Don't expect much on rules built to scramble bits, like hashes or ciphers.

Exactly right —

Bits right —

Bits recovered (of 32) —

Sure of every bit · right when sure —

a	b	true c	its answer	each bit, high → low (dark 1, light 0, in between unsure)	bits recovered

All 999 examples it read

#	a	b	c

Bits recovered is the width plus log₂ of the probability the model gave the true answer. A blind guess scores 0 and certain-and-right scores the full width, while a confident mistake drags the score down. Addition gets the full width; a hash gets about 0.

Use your own data

Upload a CSV file with columns a, b, c. The first 999 rows are the examples it reads; every row after them is a question for it to answer. Leave c empty on the question rows, or fill it in to see how often it gets them right. Numbers are whole and unsigned, decimal or 0x…, below 2³². The file never leaves your device.

Choose a CSV file (made from the rule above: 999 examples and 200 questions)

Questions answered —

Sure of every bit —

Exactly right —

Confidence is the probability of its least certain bit; an answer counts as sure when every bit is at 0.99 or above. Questions are answered 128 at a time: each is answered on its own (the questions in a pass never see each other, so the answers are the same as asking one by one), but one pass reads the 999 examples once for all of them, which is about 30 times faster.