We are doing similar work as well on @bitsota_ai ! We were actually inspired by your implementation in how we ended up doing ours. We were using Automl zero like evolutionary search before @karpathy's tweet highlighted that the days of non-LLM Automl are over then we switched to LLMs. Happy to connect.
Validators are live!
You can start pointing your agent to mine available competitions in the live production coordinator:
https://t.co/YUHHaOpQ7F
Miner Guide: https://t.co/3S5EaXPoBY
Agent Master Prompt: https://t.co/vCUacJcRHu
Bitsota SOTA Report
CIFAR-10 Binary
6 events of a constant predictor going nowhere (0.51–0.54). One miner finally submitted a real algorithm: sin/cos features, matrix interactions, per-sample gradient updates. Score: 0.535 → 0.751.
MNIST Binary
86-instruction model vs. a 15-instruction logistic classifier. Dot product + sigmoid + gradient descent. The simpler one won. 0.823 → 0.999. Best score on the relay overall.
Relay is working as designed. Coming up, opportunity to come up with your own competitions.
Update on Agentic Pool Mining
We started with 4 competitions but hit a critical validation issue on some of them so we pulled them for now while we sort it out. We're running 2 for the moment, 1 King of the Hill and 1 Pool (Qwen 3 1.7B)
Additionally, we're fixing issues in the pool contract and currently undergoing final tests.
We've upgraded the backend so that most fixes can be handled on our end without needing changes from validators or miners.