Hi all, I am going to keep this short and sweet since it seems like there have been a lot of posts like this recently.
Numbrion reached rank 1 on ranbat doubles at a peak elo of 2136. The bot had an average gxe of 88.5 and an average elo of 2043. I have never made a neural net before so this could not have been done without claude code, but I do have a technical background. This was a little over a week of work and was trained in about 10 hours on a macbook pro m1. I actually have a better model than what I used to get that score but I could not check the actual results since my account is on the bot ladder now and there are not many people playing bots at such high elo.
(this is an old image I was planning to take a new image when my new model took number 1 but since the bot change was made I can't really climb)
Basically if you though that 2v2 pokemon was safe from bots because there are more options, you were wrong. As seems to be the rule here I will keep this closed source. I also won't be making more models because my method scales with more compute and I don't really feel like renting computing space to make a better model. I will also be kind of light on the technical details because the jaxcalibur technical doc was enough for me to copy the architecture of that model and improve on it. I really love pokemon so I hope the bots can eventually be a positive force in the pokemon community. For now it's clear you only need a claude code sub and a bunch of free time (in my case unemployment) to make a model on the human limit.
Here are some of the things that were a big unlock for the model:
- Size can be extremely small even 0.8M parameters was enough to get rank 1.
- Training enviroment speed is everything you basically need to rewrite the showdown engine in rust or another fast language.
- Search is important you can even do 2 layer search if you engine is efficient enough.
- You need a teacher to show your model what is right and wrong instead of PPO.
- Teacher should be normal model with search on shallow setting to generate examples fast.
That is about it for me. Ask questions if you have any.
Numbrion reached rank 1 on ranbat doubles at a peak elo of 2136. The bot had an average gxe of 88.5 and an average elo of 2043. I have never made a neural net before so this could not have been done without claude code, but I do have a technical background. This was a little over a week of work and was trained in about 10 hours on a macbook pro m1. I actually have a better model than what I used to get that score but I could not check the actual results since my account is on the bot ladder now and there are not many people playing bots at such high elo.
(this is an old image I was planning to take a new image when my new model took number 1 but since the bot change was made I can't really climb)
Basically if you though that 2v2 pokemon was safe from bots because there are more options, you were wrong. As seems to be the rule here I will keep this closed source. I also won't be making more models because my method scales with more compute and I don't really feel like renting computing space to make a better model. I will also be kind of light on the technical details because the jaxcalibur technical doc was enough for me to copy the architecture of that model and improve on it. I really love pokemon so I hope the bots can eventually be a positive force in the pokemon community. For now it's clear you only need a claude code sub and a bunch of free time (in my case unemployment) to make a model on the human limit.
Here are some of the things that were a big unlock for the model:
- Size can be extremely small even 0.8M parameters was enough to get rank 1.
- Training enviroment speed is everything you basically need to rewrite the showdown engine in rust or another fast language.
- Search is important you can even do 2 layer search if you engine is efficient enough.
- You need a teacher to show your model what is right and wrong instead of PPO.
- Teacher should be normal model with search on shallow setting to generate examples fast.
That is about it for me. Ask questions if you have any.
Last edited:
