we built an action model playing Magic like a human introducing ANA: a normal adversary - 5M parameters - runs on a CPU - 12ms per decision - predicts human pick 90% of the time games today. robots tomorrow. this is step one towards people-shaped policies & you can... try it now 👇show more

play with ANA: blog post: and if you want policies that act like humans: DMs are open

If this training process can help game devs make smarter AI players/combatants in their games, it will be a hit. Some franchises like Total War suffer player churn from flawed AI, as well as performance issues when there are too many variables (Stellaris).

that's exactly the first kind of application we're targeting! it's both a real problem afflicting games and the perfect trial for this kind of model for us

Yes! So many strategy and RPG games would be far better if NPCs and armies played their cards better, as a real player would, and not rely on cheese-able scripted logic that leans on cheating with stats/resources.

This is a really cool project.

Thanks! We’re aiming to deploy a lot of these action models in the near future 😁

On CPU!!! This is cool!

@fdotinc im begging you. make a model that can play Total War games.

@fdotinc that would be sick! just need to find/make the right dataset for it

@fdotinc One way would be Total War games on youtube, other battle replays that game allows to save. Best would be to ask someone from gamecommunity like Turin on youtube. He has direct connection also to Creative Assembly producers of these games.

Weights ? 🧐

just shared them!

@fdotinc did you escape prison?

@fdotinc plot twist, moving to SF was just going to Alcatraz

lfg

the end of lousy bots is near 🚀🚀

90% prediction on just a CPU is genuinely impressive.