Salutations everyone,
On the eve of September 31st, 2026, at approximately 7:48:21.5 PM EST (9:33:21.5 GMT+9.75), Freaki Mikucalibur123, a bot developed by me, reached #1 on the Smogoff ladder.
Introduction
After seeing the recent success of other bots on popular ladders, I became interested in developing a bot to reach the peak of the most prestigious ladder on Smogon: the Smogoff ladder. While the ladder has certainly declined a bit throughout the generation, it certainly represents the pinnacle of competitive play on this website. Part of my motivation was a portfolio builder for future machine learning work, as I'm studying for a compsci degree. I was initially unsure how to start this problem, so I reached out to a trusted source for advice:
4 billion input tokens later, I had a working model, and I began training it.
Training Methods
Freaki Mikucalibur123 was trained through self play, with no human intervention. 100,000 copies of the bot were placed on a recreation of the Smogoff forum, and each bot was able to generate 100 interactions (threads, replies, reacts) per second. The 100 bots that received the highest react score were then evaluated by an overseer, which used sentiment analysis to determine if the best posts matched expected trends or if there was some form of reward hacking. Unfortunately, around 5% of the way through the training, one of the bots was able to generate a post that injected malicious code into the overseer, but by the time I noticed we were too far through the training to restart so I just let them have at it. The bot was trained on 4096 NVIDIA DGX B200s, which does raise concerns that this training method could be accessible to a large number of people for relatively cheap.
Future Improvements
While the bot managed to top the ladder, it's clear that there's still room for improvement. The model was trained on a relatively small amount of compute for a relatively short time (~2 weeks, 24/7) and therefore probably is far from the potential this model offers. Re implementing the overseer with enhanced protections may help the bot perform better in the real world, however, the limitations may hold the bot back, so maybe we should just let it run without oversight or audits from any outside entity.
Potential Impacts
Obviously, there are potentially very large implications for the wider Smogoff community; this bot was easily able to get recs for the recent "suck off a guy" suspect test, and a large number of these bots could easily have swayed the vote, leaving these references banned and finally giving BIG ASHLEY a break and some peace of mind (to the detriment of the metagame). Because of this, I have decided to open source the bot, weights, and engine for anyone's use. They're available at this link: https://github.com/Uranium233/Freaki-Mikucalibur123. Please everyone be responsible with this power :).
On the eve of September 31st, 2026, at approximately 7:48:21.5 PM EST (9:33:21.5 GMT+9.75), Freaki Mikucalibur123, a bot developed by me, reached #1 on the Smogoff ladder.
After seeing the recent success of other bots on popular ladders, I became interested in developing a bot to reach the peak of the most prestigious ladder on Smogon: the Smogoff ladder. While the ladder has certainly declined a bit throughout the generation, it certainly represents the pinnacle of competitive play on this website. Part of my motivation was a portfolio builder for future machine learning work, as I'm studying for a compsci degree. I was initially unsure how to start this problem, so I reached out to a trusted source for advice:
4 billion input tokens later, I had a working model, and I began training it.
Training Methods
Freaki Mikucalibur123 was trained through self play, with no human intervention. 100,000 copies of the bot were placed on a recreation of the Smogoff forum, and each bot was able to generate 100 interactions (threads, replies, reacts) per second. The 100 bots that received the highest react score were then evaluated by an overseer, which used sentiment analysis to determine if the best posts matched expected trends or if there was some form of reward hacking. Unfortunately, around 5% of the way through the training, one of the bots was able to generate a post that injected malicious code into the overseer, but by the time I noticed we were too far through the training to restart so I just let them have at it. The bot was trained on 4096 NVIDIA DGX B200s, which does raise concerns that this training method could be accessible to a large number of people for relatively cheap.
Future Improvements
While the bot managed to top the ladder, it's clear that there's still room for improvement. The model was trained on a relatively small amount of compute for a relatively short time (~2 weeks, 24/7) and therefore probably is far from the potential this model offers. Re implementing the overseer with enhanced protections may help the bot perform better in the real world, however, the limitations may hold the bot back, so maybe we should just let it run without oversight or audits from any outside entity.
Potential Impacts
Obviously, there are potentially very large implications for the wider Smogoff community; this bot was easily able to get recs for the recent "suck off a guy" suspect test, and a large number of these bots could easily have swayed the vote, leaving these references banned and finally giving BIG ASHLEY a break and some peace of mind (to the detriment of the metagame). Because of this, I have decided to open source the bot, weights, and engine for anyone's use. They're available at this link: https://github.com/Uranium233/Freaki-Mikucalibur123. Please everyone be responsible with this power :).
Last edited:
