A sad and a great message.

LLMs have incredible abilities but also serious limitations. They do all their real learning in their training phase and can't adapt to new circumstances that they haven't been trained to handle. As top War.app play is not going to be in an LLM's training set I can't see it being very good at this game.

I'm certain with the right resources you could develop an AI that would play this game better than any human, but no one with those resources are spending them on this game.

I trained a neural network to play War.app. Astra helped me with the technical details while I mainly served as domain expert. The result was attrociously bad. I do not believe the reason for the bad play was me lacking compute but I saw that this was going nowhere reasonable. Skill issue probably but then Astra also had skill issues I guess.

That was the sad message. Now the great message. Here GothamChesss got cocky against Astra and was playing at the edge of defeat: https://www.youtube.com/watch?v=PVVA-_wvKks . So I guess time for another shift of goal posts?
I feel that Astra played actually better against him than against me. But pretty sure the reason was that while playing against me it had to multitask with computer use on Windows, controlling OBS studio and playing the game. Maybe the thinking effort also played a role.

Edited 2026-10-10 19:44:00

I don't think you are drawing the right conclusions.
Even Astra beating strong grandmasters wouldnt be that interesting, or, it would be only interesting if it makes it in a humane way rather than chess engine way (maybe it's already playing it like that but I don't know how to check/test and I think it's unlikely). And even without Astra/LLM beating, it is fair to estimate they will be used in improving current chess engines.
(It would be also interesting if it could beat strongest engines currently available but I don't see it happening soon)

The issue is when it comes to warapp, it has more limitations. The source of information (related what's good and what's bad) is very limited and not always generally applicable, the game is not a perfect information game and the "move" is, I don't know it's exact categorisation/terminology but I'd call it semi-simultaneous (vs turn based in chess, go etc.) So even though the games are similar, there will be more time needed for that to happen. (It will definitely happen in the future, though.)

I also would like to see how Astra performs in other chess variants or stuff like go, reversi etc.

if it makes it in a humane way rather than chess engine way (maybe it's already playing it like that but I don't know how to check/test and I think it's unlikely).

Astra thinks in a human way in that it has best practices baked in. I believe someone at OpenAI actually trained it on chess for some reason. It is not terrible but also not very good. It will never beat StockFish obviously which can, given the right hardware, crunch through hundreds of millions of positions within 1 second. You can actually see Astras rough thinking traces when it plays chess. The actual thinking is hidden. I have not used an Open Source models for some time but I believe when you use one of the Chinese models, you can see their real thinking traces since they do not hide them.

and the "move" is, I don't know it's exact categorisation/terminology but I'd call it semi-simultaneous (vs turn based in chess, go etc.)
Yeah, that is a huge hindrance in War.app. The fog is actully not that huge of a problem but the fact that the moves happen simultanously prohibits any form of minimax.

So even though the games are similar, there will be more time needed for that to happen. (It will definitely happen in the future, though.)
Maybe someone who actually knows what he is doing can train a neural network. I did but it was terrible. Personally I think differently. I know that LLMs will strugle hard to beat the best of the best at War.app or chess for the foreseable future. But then I used Astra to build me a chess computer and it played at roughly 2500 Elo apparently (I joined some bot tournaments). If Astra can build me a War.app bot that is also fine.

Edited 2026-10-10 22:50:57

I'd like to see a head to head -- Claude vs. Astra on the classic 1 v. 1 format.