about us
hi! we're michelle and will — two ai nerds who saw the openai goblin and got curious.
michelle
i work at cloudflare leading our ai platform. i run product for workers ai, our inference platform and explore ml research on the side.
find me at x.com/michellechen.
will
i lead applied research at prime intellect. i work on the verifiers library and our Lab platform, plus dev tooling and stuff for the people building on top of it. i think RL is really cool.
find me at x.com/willccbb.
about this experiment
we trained on prime intellect's lab, it took 22 training runs and about $132 in inference and training.
we generated prompts via codex and claude. we used workers ai kimi via pi for some agent help.
about this site
this is an interactive walkthrough our goblin rl experiment. built as a cloudflare worker with vite + vanilla typescript — no framework, just bespoke scroll, animated diagrams, and a small streaming chat proxy to the prime intellect inference api so you can poke at the trained checkpoints yourself.
this website was almost one-shotted by GLM5.1 via workers ai (unreleased).
source: github.com/mchenco/goblins.