041

Knowing a lot isn't the same as being intelligent.

Intelligence can be defined as "skill-acquisition efficiency" (François Chollet, creator of ARC-AGI): how fast you learn something new, not how much you already know.

Today's AI does the opposite. It gets better by getting bigger: train a larger model on more of the internet and it gets better at questions that look like ones it has already seen.

That's why we test our model on chess and Sudoku. You can't memorise your way through them, because every position and every puzzle is new. Our model learned Sudoku Extreme from 1,000 example puzzles. In chess, it finds the best move in positions it has never seen more often than DeepMind's method, at a twelfth of the size.

General intelligence won't come from a bigger memory. It will come from models that learn from a few examples by working things out. That's reasoning.

The goal is general intelligence: a reasoning engine that works out the rules of problems it has never seen, from code to science.

Our results · Chess

The model is the whole engine.

Chess engines are usually hybrids of code and AI. AI evaluates positions and code handles the rest. We have built a chess engine that is pure AI, where the model is the whole engine.

Why this matters beyond chess: the model is the reasoning engine. The logic that normally is hand-written is now learned. Chess tests the ability to think multiple steps ahead, which is fundamental for intelligence.

Play it yourself on chess.041.io
Chess without the codeHow often the model picks the best move in a position it has never seen.Both learn only from example positions: no chess strategy is written in.0%10%20%30%40%50%60%051015202530Chess positions studied (millions)041 chess  52%DeepMind  44%041 chess (4M parameters)DeepMind's method (50M parameters)Both learn from the same Lichess positions with Stockfish's best moves, and are tested on the same held-out positions.DeepMind's method is our reimplementation of "Grandmaster-Level Chess Without Search" (2024).

Our results · Sudoku

It learned the rules from examples.

Sudoku Extreme is a set of the hardest Sudoku puzzles, a standard test of how well AI reasons. Our model learned to solve them from 1,000 example puzzles, with no Sudoku rules written in. At 99.8%, it beats every model we tested against, including ones up to 28 times bigger.

Why this matters: you can't memorise your way through Sudoku. ChatGPT or Claude could write you a program that could solve Sudoku, but they can't solve it themselves. Every puzzle needs step-by-step logic. Learning that logic from a thousand examples, in a small model, is the skill we're taking to harder problems.

Sudoku ExtremeGeneric model architecture: nothing about Sudoku is given.Share of test puzzles solved in one try.Log scale: 90 → 99 → 99.9% are equal steps0%90%99%99.9%55HRM27M86.6GRAM10M87.4TRM5M93.0EqR5M99.8041955k99.2FRM7MparamsTrained on 1,000 puzzlesSudoku ExtremeGeneric model architecture:nothing about Sudoku is given.Share of test puzzles solved in one try.Log scale: 90 → 99 → 99.9% equal steps0%90%99%99.9%55HRM27M86.6GRAM10M87.4TRM5M93.0EqR5M99.8041955k99.2FRM7MTrained on 1,000 puzzles.Under each name: parameters.