https://www.youtube.com/watch?v=zoFEPcPT6Io

GPT-6 Astra playing War.app via computer use and without any handholding. It should be easily repeatable with a subscription. This here is really just letting the model cook on its own, so I wonder what it can achieve with some human guidance.

how far can 25 of these go in clan wars🙃

I'm pretty sure Astra could win clanwars.

I guess this is expected, given reports like https://www.notebookcheck.net/GPT-6-Astra-AI-model-beats-Valve-s-Portal-game-solo-in-24-hours.1392390.0.html and https://talkchess.com/viewtopic.php?t=86707 etc.

Still kind of crazy. Maybe you could have it commentate/ evaluate a top ladder game(s) and see if it finds improvments to how they play?

Hm, you have to be careful with AI skills to not compare apples to oranges. E.g. for football (soccer) it would be very easy for a robot team to beat a human team. All you need to do is to give the robot team the ability to shoot super hard. But this does not mean that the robot team learned any human intricacies of the game.

As for chess, even I could write a chess computer by hand which easily beats myself. But if GPT-6 just beats you by writing a chess computer or worse just loading an instance of Stockfish, you would call that cheating. For War.app I would accept an AI to write a program to help it to make decisions.

What you saw in my video is Astra just playing out of the box. This is far from reasonable human level but the challenge was here more that Astra can play in the first place and this at a somewhat reasonable level.

Maybe you could have it commentate/ evaluate a top ladder game(s) and see if it finds improvments to how they play?
Out of the box Astra is just plain to weak, absolutely 0 way. What I see is the possibility that with the help of Astra someone develops an extremelly strong War.app bot which plays at top human level. We currently have only human intuition and simple human calculations dominating this game. If we could combine this special human skill with a computers ability to brute force then I see a possibility.

Writing a chess computer which beat me would take... IDK there is boilerplate of allowed moves but ignoring the boilerplate maybe 100 lines of actual code. War.app would take multiple 10k lines of code.

Can it play Nations Cup? We need one more player for a 2nd team. 🫣

I feel like this post has started to make things sound reasonable or possible that simply wont happen.

For context, whenever i talk about 'Astra cant play this game' i automatically mean at an decent clearly 'understanding the game' level, let alone possibly coming up with new stuff...
Can it play in general? Sure.. the same play it can play chess.. i.e. it eventually starts making shit up.

Astra and for that matter any LLM will never be good at warzone for the simple fact that it is trained to predict the next token in a text (in simpler non-technical terms, it predicts the next word)
Thats all an LLM is, including Astra.

You can write tools around it, have brute force algorithm around it, give it all the information in the world to it. At the end, its still just an LLM and simply predicts the next word. It will and simply cannot ever understand this game.

It doesnt even understand Chess well enough to not start to eventually cheat, and there are billions of games of Chess, and proper text notation for it which it learned from. And why? Because it can learn how to write Chess Text notation, but it doesnt understand it. Because it doesnt understand Chess.
The same thing here.

It is entirely possible to write a NN in general and have it learn specifically for this game and specifically with this games data.
THEN it is possible to train an AI to play this game... it will need to learn through self-play, thats for certain. The raw data this game has is not enough to teach an AI, but possible.
Basically what AlphaZero was to Chess, Go and Shogi

But importantly, that would be a model Specifically trained on the game (likely specifically for a certain template) and not an LLM.

No LLM, and if its a new 'Super Ultra Extra Deluxe Astra+++' LLM model, it still wouldnt be able to play this game well.. as it wouldnt understand it. it would just make stuff up.

iirc Alphagozero was not game specific, it could learn anything. On that basis, I suggest that it could learn War.app and dominate.

> iirc Alphagozero was not game specific, it could learn anything. On that basis, I suggest that it could learn War.app and dominate.

Alphagozero was only the game GO, as shown in the name.
You are thinking of AlphaZero, that learned Go, Chess and Shogi, which i have named.
It should be mentioned tho that only the Framework was identical, not the actual Model that ended up playing each game, those were different.

It should be also noted that Alphago Zero uses massive amount of hardware to process information about game. Not available to anyone.

@Roi Joleil: I believe you are heavily underestimating tools like Astra. But I will setup a chessboard and play against it I guess. I even know that you can find online videos about atrocious AI chess however I have doubts about how non programmers have set this up and it might forget that AI is on an exponential curve.

A few days ago, AI has solved a millennium price problem. Where I agree with you is that LLMs have a certain architecture and e.g. will never make Stockfish obsolete which is hyper optimized to crunch millions of chess positions per second. But humans also lack that ability.

Edited 2026-09-10 16:12:28

It’s not a coincidence that Shogi/Chess/Go were chosen by the DeepMind team - these are perfect information games where a lack of “intuition” can be compensated for with an extreme advantage in “calculation”.

Warlight is a much harder problem in my opinion and each template would have to be individually trained.

Leon with data from 2017, ty

you need to train your agent, not just plug and play

Edited 2026-09-10 19:28:02

Yes, you need to train your agent/model/engine to be really really really good at calculating and evaluating positions. If the engine gets good enough at this, it no longer needs to “intuit” to beat the best human players. This was proven to work in perfect information games not by DeepMind in 2016 but by DeepBlue in 1996.

Warlight is not a perfect information game though. There’s simply not that much to calculate in a fog-of-war game. I hypothesize that there simply isn’t enough to calculate in order to compensate for a lack of intuition.

That’s true for NNs and equally true for LLMs. It’s a completely different problem. I think it would make sense to see how these models do on a fogless template where they aren’t significantly disadvantaged relative to humans.

you are saying it, like it's a big deal..

go past games, read 10000 games, come back.. maybe even gpt6 can do it

plus, all games are actually publicly available

Edited 2026-09-10 20:40:30

Game against GPT-6 Astra is done. The game took 45 minutes. I will speed it up and upload a video. I will also see if I can get a StockFish analysis for both my and the GPT-6 game.

Edited 2026-09-10 21:19:06

I dont have a lot of knowledge in the field of AI, but I am also very sceptical about a LLM dominating this game. For similar reasons already said.

If we want to have an AI that is good at playing War.App, I think we have a better shot at training an AI specifically for playing War.App. Of course, there are a lot more difficulties to overcome for training such an AI for this game than for example chess. But with technology advancing, I believe this has a better chance of ever getting close to beating everyone in this game than a LLM.

https://www.youtube.com/watch?v=TxOUh991K6E

GPT-6 Astra won against me in chess. Here the video. I played directrly in the ChatGPT browser app, so you can see the bots thoughts simulatenously. I forgot to talk about this in the video but I believe that when non programmers make such videos they might make mistakes in their setup. So first I let Astra build a chess board locally with only allowed moves. And then I played against it with the computer use skill. There were some compactions during the game, probably due to computer use being heavy. Anyhow, if I had setup the game without Astra being able to see the board, it might just lose track after compaction.

Astra made 1 bizarre blunder but it still managed to win against me. StockFish believes that I am a 1050 Elo player and that Astra is a 1350 Elo player (probably not worth too much but more for the direction). 71.1 accuracy for me and 79.7 for Astra.

Here the chess.com analysis: https://www.chess.com/analysis/game/pgn/3D6kqDEXrv/analysis

Edited 2026-09-10 22:39:42

BTW, since I wrote both War.app bots and a chess bot for fun before AI was even a thing, here my thoughts on the difficulties of both.

- I believe LLMs will heavily strugle to play chess above a certain level. Maybe enough to beat Magnus Carlsen but this is not how chess computers work. Chess computers work generally with very simple evaluation functions and by analyzing a ton of future positions like a couple hundred millions. There is a tradeoff between the evaluation function and the ability to go deep into the search tree. If your evaluation function is too complex it takes too much time to calculate. This means however that the way an LLM can play superb is by not playing itself but by just programing a simple chess computer.

- War.app is different. It is impossible to minimax into the future however there is less need to brute force. The reason is that the current board state already contains the information you need. So if you e.g. manage to break an opponent bonus, that is always a good thing. You do not need to look into the future to find out that e.g. if you take the opponent pawn with your queen, this will lead to a forced move sequence where you lose your queen. The way I see War.app is that you need an extremelly strong evaluation function and you support that with localized calculations. E.g. you calculate which bonuses you can take the next 3 turns. It is way more difficult than in chess to formalize the game properly and (at least for me) the maybe main difficulty is finding out what I even want to calculate in the first place. Even for just 1 move it makes no sense to calculate all my options and all opponent options but you need to prune obvious bad play. And even if I had calculated all the outcomes of the possibilities for 1 move, the question is which move to pick since I do not know how the opponent plays. This would need to involve some randomness on my part since the mathematical issue we have here is a mixed strategy nash equilibrium where there is no "best move" but optimal War.app play involves throwing a weighted dice.

Edited 2026-09-10 23:12:02

The chess game makes it clear to me that out-of-the-box Astra is not going to be a competitive player. Chess is almost certainly “easier” for an LLM than Warlight and yet it’s straight up hanging minor pieces.

I am curious what level of play we get to from a training set of ~10,000 Fogless Fighting games. It’s probably still too small to make a big impact but I’m trying to be pragmatic and recognize that no one is going to fund the tokens for millions of self-play games.