AI research lab

Why it exists

I want a working model of how inference actually behaves, precise enough to teach, humble enough not to fake a benchmark.

Problem being investigated

Technical explanations drift. They recompute the prompt from scratch in the reader's imagination, they treat attention as comprehension, or they present sampling controls as if they were layers of the network.

Model

The active artifact is the prompt-to-token explainer: a pipeline, an attention sketch labeled as conceptual, a sampling playground labeled as illustrative, and a prefill/decode comparison that says "commonly" where the literature says the bottleneck depends.

Lessons

• The correction is often a sentence, not a new visualization. "Not recomputed from scratch" matters more than another animated neuron.
• Illustrative numbers have to be branded as illustrative in the same view as the chart, not in a footnote.