model-graph

model-graph blog

notes on live LLM interpretability, from building it

See your model think: live LLM internals in your browser

model-graph streams every layer of a running language model — norms, attention, logit lens, residual trajectories — live into a chat UI. OpenAI-compatible, open source, MIT.

2026-07-20

The logit lens, live: watching a 1.2B model settle on its answer

What the logit lens shows, how model-graph renders it as a per-token ladder with flip dots, and what settle depth reveals about per-token difficulty and depth-as-compute.

2026-07-20

Making ONNX observable: residual streams for in-browser interpretability

Compiled ONNX graphs only expose declared outputs. With ~40 lines of graph surgery, the residual streams already flowing through fused SkipLayerNorm nodes become visible — enabling WebGPU in-browser interpretability.

2026-07-20