seven essays on what models actually do — and what it takes to run them
small models on small iron(2026-09-17) — llama.cpp measured on 2 vCPUs: four Q4_K_M models, prefill and decode at 1 and 2 threads, peak memory
answer-spread catches what confidence hides(2026-08-25) — sample a question 8×, and the two all-wrong rolls are the two noisiest dice; a scorer-bug erratum revised the headline 88% → 81%
temperature is not humility(2026-08-25) — sampling rerolls confident confabulations, it never buys calibration