Writing
Notes on evaluating AI systems, mostly where models fail. I build and benchmark agent systems, retrieval pipelines and the evals that check whether any of it works.
- Oct 26, 2026 On a laptop, the YOLO version is the smallest factor in how fast it runs YOLO26 vs YOLO11 across eight runtimes on one laptop: the model version changes latency by about 5%, the runtime by 5×, and Windows' power mode by up to 8×. It nearly made me publish the wrong result.
- Oct 26, 2026 A 2019 classifier ties a 2026 decision model on agent guardrails Zero-shot, an open 2026 decision model does no better than a 2019 NLI classifier on 244 agent-guardrail decisions, and neither beats always giving the same answer.