After Nessie and Miku, i was thinking, I have experience in this area (AI/ML etc.). How long would it take for me to get something "ok" working? The answer was 1 week.
This bot will never be let on any human ladders, or the source code wont be public, ever.
This is a VGC bot trained on my local computer and it can only access my showdown instance that i privately host.
Pfft, there are already bots that are between 1200~1300 why are we ringing alarm bells? Because this bot is not done training. The previous attempts are known to peak around those elos. My bot is missing most of the features from Jaxcalibur and Nessie. E.g. a pUCT has been shown to add 100~200 elo to strength of the bot. It still wouldn't be super human but training it more and adding search would beat more and more humans. Still worse compared to Nessie and Miku but i have more reasons to ring the alarm bells.
AFAIK, all the previous implementations (Jaxcalibur, MikuMiku37 for sure) used a custom port of showdown server code to something faster than javascript via LLM asisted coding. My code is running on the showdown code itself. So development time of a bot is reduced by a lot. Training becomes exponentially harder when the bot gets better at play. I think without a faster server implementation it is really hard to reach super human level, but this doesn't mean "vibecoders" (deragotary) won't flood the ladders with 1300~1500 elo bots.
Currently my bot is a buggy mess, but it proves the concept. If a "beta" version takes 1 week to make, it will only get faster to do. More tips and tricks will be found to optimize training.
In my opinion, best bots will be great training buddies. Inversely, they will will severely damage the community. People will start prepping offline. Why would you prep online if you can train against every possible meta team and solve the meta privately? Why would you need to share the team and report before bagging a regional win? If you don't need to get advice from friends to make the team better, your team cannot leak before regionals. It is not going to happen tomorrow, but it is in the horizon.
I'm not a doomer about this topic but but we have many problems that we need to solve fast. I think it will evolve the quality of the game really fast but at the same time damaging the social aspects of the game.
One catch here is i know top VGC players might mark their accounts as bot to hide info before a tournament since they are excluded from places.
Sub dividing ladders is on paper easy but there is so many open questions on the implementation as well.
This bot will never be let on any human ladders, or the source code wont be public, ever.
This is a VGC bot trained on my local computer and it can only access my showdown instance that i privately host.
I deliberatly don't want to give any important details on what i did so people won't flood the ladder with "slop" bots.
All i did was writing some PPO self-play trainer that plays on the source code of pokemon showdown. It is very bare bones without any extra features to boost training quality and play quality. By my current estimation my bot is around 1200~1300 elo after merely 1 day of training. Everything was done locally. In my balcony, so my laptop doesn't overheat. I'm really scared about "state of play" after seeing just a proof of concept beating me (I'm not the best player on Reg M-C but I'm decent)Pfft, there are already bots that are between 1200~1300 why are we ringing alarm bells? Because this bot is not done training. The previous attempts are known to peak around those elos. My bot is missing most of the features from Jaxcalibur and Nessie. E.g. a pUCT has been shown to add 100~200 elo to strength of the bot. It still wouldn't be super human but training it more and adding search would beat more and more humans. Still worse compared to Nessie and Miku but i have more reasons to ring the alarm bells.
AFAIK, all the previous implementations (Jaxcalibur, MikuMiku37 for sure) used a custom port of showdown server code to something faster than javascript via LLM asisted coding. My code is running on the showdown code itself. So development time of a bot is reduced by a lot. Training becomes exponentially harder when the bot gets better at play. I think without a faster server implementation it is really hard to reach super human level, but this doesn't mean "vibecoders" (deragotary) won't flood the ladders with 1300~1500 elo bots.
Currently my bot is a buggy mess, but it proves the concept. If a "beta" version takes 1 week to make, it will only get faster to do. More tips and tricks will be found to optimize training.
Why are you sharing a half-baked post like this?
I wanted to warn community on how easy to train something if you have access to middle of the pack Macbook Pro. It is not cheap but it is not hugely expensive as well. Accesibility to a tool like this is here now. We need to take some measures!In my opinion, best bots will be great training buddies. Inversely, they will will severely damage the community. People will start prepping offline. Why would you prep online if you can train against every possible meta team and solve the meta privately? Why would you need to share the team and report before bagging a regional win? If you don't need to get advice from friends to make the team better, your team cannot leak before regionals. It is not going to happen tomorrow, but it is in the horizon.
I'm not a doomer about this topic but but we have many problems that we need to solve fast. I think it will evolve the quality of the game really fast but at the same time damaging the social aspects of the game.
What should we do? My proposal
I believe most people are not malicious in competitive pokemon communities thus i suggest this as a solution until we see iritating people flooding the ladder with slop bots.- Users can mark their account as bot. a.k.a. Self registring the bots with flag in their account.
- We should have a checkmark under "Don't allow spectators" that says "Don't allow bots" in Battle! area
- If a user is marked as bot, then their ladder ranking is skipped when we go to "Ladders" tab. Thus we exclude those accounts from ladders tours or elo-based voting in other ladders etc.
One catch here is i know top VGC players might mark their accounts as bot to hide info before a tournament since they are excluded from places.
Sub dividing ladders is on paper easy but there is so many open questions on the implementation as well.
Acknowledgements
- All elo estimates are not science based, since im not letting it to the ladder, we cannot know for sure. I'll let my top tier play friends test it after i work on it a little bit more to get an estimate. But when people know they are playing against bots, they play differently.
- All the elo estimates are "peak elo", it doesn't consistently hit it. This bot is prone to "brain farts".
