Latest / Logosofia no Dia a Dia
When Your AI Says No to Everything
Why does your AI assistant refuse to answer harmless questions? This episode explores the hidden failure mode of over-refusal in large language models — when safety guardrails fire on innocent prompts like "how to kill a mosquito." We break down three key benchmarks: OR-Bench (80,000 prompts), XSTest (the predecessor now considered "solved"), and PHTest (model-specific pseudo-harmful prompts). The core finding: there's a near-perfect correlation between refusing toxic prompts and over-refusing benign ones (Spearman rank 0.89). Claude models show the highest safety but also the highest…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.