Hi all — I’m a systems engineer most of the time, and I LOVE Mojo.
Last year I found myself wanting to understand what’s happening when a model does something it shouldn’t. Gemma-4-31B is a great model, but it will confidently make up papers it’s never read and I needed to know why. That pulled me down a genuinely fun, rewarding rabbit hole into the internal state of the model.
Today I published my hand-rolled kernel. I built it, along with the serving path and a training pipeline (not yet published) by deploying and orchestrating a whole fleet of agents to work alongside me. The part I’m most excited about: the interpretability tooling plugs straight into the kernel and lets me read what’s happening LIVE, during the forward pass.
The kernel and serving path are open now (MIT); the interpretability tooling and training pipeline are next. Work’s here: GitHub - lightofbaldr/nomos-nvfp4: Pure-Mojo 4-bit (NVFP4) inference kernel for Gemma-4-31B on NVIDIA Blackwell, with lossless speculative decoding. · GitHub
(the model + drafter it runs are linked from the repo, on HF). Research writeups: Research — Topology of Thought — Nomos Logos
Would love any thoughts.