041

From reasoning towards superintelligence

Our approach to reasoning produced the top Sudoku model and a chess model that learns faster than DeepMind’s method. Winning at games is the test, not the goal.

A few weeks ago, we had a breakthrough in our research. Since then, we’ve built the world’s best model at solving Sudokus, and a model-only chess engine that beats DeepMind’s transformer-based version. Games are fun, but why does this matter? Just like we use games to teach children, they hold lessons for artificial intelligence.

How did we end up here? When founding the lab, our key insight was a better shape for logic and the code. We had long struggled to make this work, and had already explored a bit of other venues. Then, a curious thought came on a Friday evening, and I let the agents return to our old friend, the neural interpreter. By Sunday, we had stunning results. Our architecture was able to learn much faster than a transformer. It also generalized to samples more complex than anything in the training data.

Neural interpreterHow often each model computes the exact output of a small Lisp program.Depth 6 is deeper than any program seen in training.041 modelTransformerMambaDepth 3TRAINED0%50%100%02.557.510Training steps (thousands)100%67%33%Depth 6NEVER SEEN0%50%100%02.557.510Training steps (thousands)96%2%2%Small Lisp programs; exact-output accuracy, mean of 2 seeds.041 Lisp synthesis report, 19 Oct 2026. Early result, small testbed.Neural interpreterHow often each model computesthe exact output of a small Lisp program.Depth 6 is deeper than any programseen in training.041 modelTransformerMambaDepth 3TRAINED0%50%100%02.557.510100%67%33%Depth 6NEVER SEEN0%50%100%02.557.510Training steps (thousands)96%2%2%Small Lisp programs; exact-output accuracy,mean of 2 seeds. 041 Lisp synthesis report,19 Oct 2026. Early result, small testbed.

Results always come with some uncertainty. Something could be off with the data, or perhaps the approach might work with only a very specific type of problem. We went on to test the architecture beyond code, choosing Sudoku as the next target. There is a wide range of published results for Sudoku, and it’s a quick one to work with as it doesn’t need much compute. The same thing as with code happened again - our model was extrapolating beyond the data it had seen. Today it solves 99.8% of Sudoku Extreme, the best result of any learned model. We also do it with fewer parameters than anybody else on the rankings.

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.

In the meantime, we had also started working on chess. The shape of reasoning in chess is more general than in Sudoku. The game requires thinking in consequences, and then aggregating the findings to choose the best move. This is similar to working through problems and solutions - your opponent poses problems, you solve them. This is the shape of intelligence needed throughout life: figuring out what are the problems worth solving, and then solving them. The work around chess is still active. We have already demonstrated a model-only chess engine that learns faster than our reimplementation of DeepMind’s transformer-based variant, with less than 10% of the parameters.

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).Chess without the codeHow often the model picks the best movein a position it has never seen.Both learn only from example positions:no chess strategy is written in.041 chess (4M parameters)DeepMind's method (50M parameters)0%20%40%60%0102030Positions studied (millions)041 chess 52%DeepMind 44%Both learn from the same Lichess positionswith Stockfish's best moves, and are testedon the same held-out positions. DeepMind'smethod is our reimplementation of"Grandmaster-Level Chess Without Search" (2024).

Our upcoming research includes more reasoning tasks, and gradually more focus on the areas that truly matter: language and code.

I’m often asked which part of AI we are solving. We are here to solve it all. The easier paths like becoming a Jev-like restricted logical engine will never win in the long run. The only way to truly compete is to build the best overall intelligence. We are not publishing the methods, as the race is serious. We are running towards intelligence beyond the best human experts. The “why” behind the 041 reasoning engine is superintelligence.

← ALL POSTS