# Introducing Nous: A Different Kind of Neural Chess Engine

*July 22, 2026 · By Balaji Daggupati*

> Every strong chess engine wins by search. Nous does not: it imagines a whole line of play at once, judges it, and rewrites any part, the way a person does. Building it that way also made it an instrument, because a system that imagines, judges, and revises let us measure something the rest of AI is betting on but cannot usually see, whether a system can trust its own judgment. Here is what it showed.

Nous is a chess engine that plays on intuition, not [search](https://en.wikipedia.org/wiki/Minimax). That is an odd thing to build. For seventy years, every engine that mattered has won by search, racing through millions of positions a move: [Deep Blue](https://en.wikipedia.org/wiki/Deep_Blue_(chess_computer)), [Stockfish](https://stockfishchess.org/), [AlphaZero](https://en.wikipedia.org/wiki/AlphaZero), [Leela](https://en.wikipedia.org/wiki/Leela_Chess_Zero), all of them. Nous does none of that. It imagines an entire line of play at once, as a rough sketch, then rewrites any part of that sketch, including its own next move, in light of how the line ends. We named it Nous, an old word for the kind of knowing that arrives whole, without being reasoned out step by step, which is close to how it plays. It plans the way a person does, not the way a computer does. Under the hood it is a [diffusion model](https://arxiv.org/abs/2406.07524), the kind of system usually pointed at generating images, here pointed at a chessboard, and it plays in public on [Lichess](https://lichess.org/@/PebbleTraining), where it has climbed from losing every game to a solid club level, around 1600 [Elo](https://en.wikipedia.org/wiki/Elo_rating_system) across roughly two thousand games.

We built it this way because it plans the way people do and today's dominant models cannot. In [de Groot's studies](https://en.wikipedia.org/wiki/Adriaan_de_Groot) a grandmaster looks at only a few candidate lines, follows one a little deeper, notices the ending is bad, and goes back and changes an earlier move. That last step, revising a decision already made, is exactly what a [left to right](https://en.wikipedia.org/wiki/Autoregressive_model) language model cannot do once it has committed a word, and what a search tree does not do, since it discards its bad lines rather than editing them. Nous can. As far as we know, planning a whole line by imagining and revising, with no search anywhere, is an unusual way to build a chess engine, and getting to watch it work was reason enough to try. (If doing [implicit search on a chessboard by denoising](https://arxiv.org/abs/2502.19805) sounds like a rabbit hole, it is a good one.)

![How three engines think: Stockfish and Leela search; Nous imagines one line and revises it.](/diagrams/how-engines-think.svg)
*Three ways to pick a move. Every strong engine wins by searching. Nous does not search at all; it drafts a whole line and revises it, the way a person does.*

![The Nous loop: imagine a line, judge it, revise any part, repeat, then play.](/diagrams/nous-loop.svg)
*How Nous thinks. No tree and no millions of positions: a draft, a verdict, a rewrite, over and over.*

## How it plays

Watching Nous is nothing like watching Stockfish. It plays a recognizably human game: principled, sensible openings, where it is strongest and regularly outplays higher rated bots straight out of the gate, and a soft spot in the calculation dense middlegame, where it will start a tactic on a hunch and not always finish it. It hangs the occasional piece and misses the occasional one move trick, the way a talented, slightly over eager human does. Its best phase is the opening and its worst is the middlegame, the exact reverse of a brute force engine, and the fingerprint of a player running on intuition rather than calculation.

## Can a machine trust itself?

It also turned out to be more than a chess engine. A system that imagines something, judges it, and then keeps or rewrites it is a working miniature of the loop the whole field is now scaling. [Reasoning models](https://arxiv.org/abs/2501.12948) generate many attempts and keep the ones they score highest. [Reinforcement learning](https://en.wikipedia.org/wiki/Reinforcement_learning_from_human_feedback) rewards the answers a model, or a learned judge, prefers. Test time compute is a model thinking longer and then trusting what it thought. All of it rests on one assumption almost nobody can test: that a system's judgment of its own work is good enough to act on. In language there is no answer key, so the assumption goes unchecked. Chess has one, instantly and for free, and for simple endings [tablebases](https://en.wikipedia.org/wiki/Endgame_tablebase) make it provably exact. So we used Nous as an instrument, and watched it imagine, judge, and revise where every judgment can be checked. Three things came back. None of them is only about chess.

## A verdict comes before a vote

We gave Nous a judge, a small network trained to score positions, and let it help choose moves. It failed in a way that took us a while to believe. Whenever we let the judge rank the candidate moves and pick the best, the engine got *worse*. Whenever we used the exact same judge only to veto the obvious disasters and otherwise trust the model's instinct, it got much better. We saw it again and again, across every method we tried to make it rank. A judge whose scores tracked the truth closely, correlated with a strong engine at about 0.86, still hurt the moment we asked it to *select* rather than to *filter*.

![Veto only helps for both a decent and a sharper judge; ranking hurts under the decent judge but flips to helping once the judge is sharper.](/diagrams/veto-rank.svg)

*The same judge, two jobs. A middling evaluator helps when it only vetoes blunders but hurts when it ranks and chooses. Halving its error against a strong engine flips ranking from harmful to helpful.*

> A learned judge can be trusted to throw out disasters long before it can be trusted to pick winners.

Then we sharpened the judge, cutting its error against a strong engine roughly in half, and the sign flipped. Now ranking helped, and a move it used to get wrong in a live game it now got right, with everything else held fixed. So there is a threshold, and it is not one threshold but two: **a model earns the right to catch its own worst ideas long before it earns the right to choose among its good ones.** Trusting your own judgment is not a switch that flips on when a model is finally good enough. The bar to filter sits far below the bar to steer. This is the same wall that [reward model overoptimization](https://arxiv.org/abs/2210.10760) and [Goodhart's law](https://en.wikipedia.org/wiki/Goodhart%27s_law) describe in language models, except that here we can watch a judge cross it, and name the exact game and move it turned on: move 27 of [a real game](https://lichess.org/MF4ws2kc), a winning position, where the earlier judge chose Ne3 and let the win slip, and the sharper judge chose Qxg5, the winning move. Same player, sharper judge, wrong turned right.

## Reward reorders more than it creates

The next hope, and one I shared, is that reinforcement learning teaches a model new abilities. We ran it carefully, with [verifiable rewards](https://arxiv.org/abs/2402.03300) that were exactly checkable rather than guessed at, and it worked in the narrow sense: Nous got measurably better at tactics. Then we asked the question the answer key finally let us ask, which is *where* the improvement came from. For every tactic it newly solved, was the winning move already somewhere in its head, ranked too low, or was it a move it simply could not produce before?

![Of the tactics reinforcement learning fixed, 93 percent were reordered from moves the model already had; only 7 percent were new.](/diagrams/reward-split.svg)

*Almost all of the gain was reordering moves the model already generated. Only a sliver was ability it did not previously have.*

Reinforcement learning was overwhelmingly taking candidates the model already possessed and promoting them into a better order, not discovering candidates it lacked. It sharpened what the model already had far more than it created anything new. This is what a recent line of work finds in language models too: [reinforcement learning with verifiable rewards](https://arxiv.org/abs/2504.13837) mostly biases a model toward reasoning paths already inside its base distribution rather than expanding the boundary of what it can do. And there was a sharper detail underneath, one closer to the noise than the headline numbers but pointing the same way: the model gained the least when the reward came from its *own* judgment rather than from external truth. As far as we can tell, the useful signal has to come from outside. A model cannot bootstrap far by judging itself against itself, which is the first finding again in different clothes.

## Thinking longer is only as good as what you are thinking about

The last one reorganized how I think about the rest. We kept trying to make Nous calculate deeper, and it kept refusing to improve, and for a while we blamed the machinery. We were wrong. Nous plans over its own imagined continuation of the game, its little [world model](https://worldmodels.github.io/), so we measured that imagination directly: when it pictures the next few moves, how often does the line stay faithful to how the game goes?

![The imagined line stays fully consistent 73 percent of the time one move ahead, falling to 28 percent by five moves.](/diagrams/imagination-decay.svg)

*The imagined line decays with depth. By five moves out, only about a quarter of it survives contact with reality.*

By a few moves out, only about a quarter of the imagined line is still true. That single fact explains the rest. **If the future you are imagining is mostly fiction, then past a shallow depth thinking longer stops helping and starts to hurt, because you are carefully refining a line that was never true in the first place.** Extra refinement passes helped for a while and then flatly stopped, and asking it to calculate several moves deep scored worse than simply trusting its single best move. The bottleneck here was never how cleverly Nous explored its imagined future. It was whether that future was real enough to be worth exploring at all.

This is the one to sit with, because it puts a quiet condition on the whole excitement about models that think longer. Longer only helps when what is being thought is true. We could see it only because we had the answer key. In language, a model reasoning at length over a false premise looks, from the outside, exactly like one reasoning over a true one, right up until it hands you the wrong answer in immaculate steps.

> **More curious about the technical side?** The companion [technical report (PDF)](/planning-by-revisable-imagination.pdf) states every claim precisely and makes each one falsifiable: the architecture, the training recipe, the judge threshold traced to the exact game and move, the reorder versus create decomposition, the imagination probe, and a controlled scaling study, with the conditions under which each result would count as wrong.

---

## The one thing all three share

Line them up and they are the same lesson told three ways. A judge helps only once it clears a threshold. Reward mostly rearranges what is already there. And thinking longer pays off only over an imagination that is true. Every one of them comes back to a single dependency: **a model can only improve on its own judgment to the extent that judgment stays tied to something true from outside the model.** Cut the tether and each mechanism turns on you. We ran a scaling study too, training a model twice the size on identical data, and it did sharpen the imagination, which was encouraging, but it did not raise the ceiling on how far ahead Nous could usefully think. Scale makes the daydream more accurate. It does not hand you the answer key.

That is why chess was worth doing at all. It is one of the last places where the truth is cheap, checkable, and public, so a model that imagines and judges and revises can be watched doing each of those things and checked at every step. Nearly everything else we are building runs that same loop out in open water, where the truth is expensive or missing entirely, which is exactly why the loop is so hard to reason about. A small, checkable experiment cannot settle the big questions. But it can drag a few of them out of the realm of intuition and into the realm of things you can measure, and that alone is worth something.

None of this makes Nous a grandmaster, and the mechanisms that might get it there at real scale we have not yet demonstrated. We wrote down exactly what would prove us wrong before we ran the experiments that might, which matters to us more than being right. Nous is still playing, in public, while we chase the next question, and you can watch it, which is the other reason we built it this way. When a machine reasons in moves on a board instead of numbers in a vector, you can see the moment it flinches at a bad line and takes it back. It is, for now, one of the few kinds of machine thinking you can sit and watch happen.

## Go play Nous

Nous takes all challengers, day and night, on Lichess as PebbleTraining, where you can watch its games live or take it on yourself. Watch it sketch a plan, second guess itself, and take a move back. Fair warning: it hangs the occasional piece, and it will still probably annoy you in a sharp position. [Play Nous on Lichess](https://lichess.org/@/PebbleTraining).

## Down the rabbit hole

If any of the above pulled at a thread, here is where those threads go.

**How Nous thinks**

- [Simple and Effective Masked Diffusion Language Models](https://arxiv.org/abs/2406.07524) — Sahoo et al., the recipe the generator is built on.
- [Implicit Search via Discrete Diffusion: A Study on Chess](https://arxiv.org/abs/2502.19805) — Ye et al. (ICLR 2025), the closest predecessor to what we built.
- [Planning with Diffusion for Flexible Behavior Synthesis](https://arxiv.org/abs/2205.09991) — Janner et al., planning as generation rather than search.
- [World Models](https://worldmodels.github.io/) — Ha and Schmidhuber, on agents that plan inside an imagined model of the world.

**Chess as a laboratory**

- [Amortized Planning with Large-Scale Transformers: A Case Study on Chess](https://arxiv.org/abs/2402.04494) — Ruoss et al. (DeepMind), grandmaster strength without search.
- [AlphaZero](https://en.wikipedia.org/wiki/AlphaZero) and [Stockfish](https://stockfishchess.org/) — the two great search engines Nous deliberately is not.
- [Adriaan de Groot](https://en.wikipedia.org/wiki/Adriaan_de_Groot) — the psychology of how grandmasters search, which is barely at all.
- [Endgame tablebases](https://en.wikipedia.org/wiki/Endgame_tablebase) — where chess truth becomes provably exact.

**Judging, rewarding, and thinking longer**

- [Scaling Laws for Reward Model Overoptimization](https://arxiv.org/abs/2210.10760) — Gao et al., the filter versus steer wall, measured. See also [Goodhart's law](https://en.wikipedia.org/wiki/Goodhart%27s_law).
- [Does RL Really Incentivize Reasoning Beyond the Base Model?](https://arxiv.org/abs/2504.13837) — Yue et al., the language-model version of reorder versus create.
- [DeepSeekMath](https://arxiv.org/abs/2402.03300) and [DeepSeek-R1](https://arxiv.org/abs/2501.12948) — verifiable rewards and the reasoning-model loop at scale.

Nous plays continuously at [lichess.org/@/PebbleTraining](https://lichess.org/@/PebbleTraining). The [full technical report (PDF)](/planning-by-revisable-imagination.pdf) carries the architecture, the experiment ledger, and the exact conditions under which each claim would count as wrong.

---

**Keywords:** Nous, discrete diffusion, self verification, reward models, world models, chess