Latest / Models & Agents — Video Edition
Ep 143: Models sound most sure of themselves exactly when their answers are wrong — and a new eval…
Models & Agents Models sound most sure of themselves exactly when their answers are wrong — and a new eval harness is making that gap impossible to ignore. What You Need to Know: An enterprise architect built a synthetic ground-truth harness that revealed LLMs confidently misattribute root causes in data-drift scenarios, especially when signals overlap. Open-weight Qwen 3.8 27B drew attention for over-thinking in LM Studio and for successful Jacobian-lens transfer from the prior 3.6 checkpoint. Anthropic’s Dario Amodei laid out a detailed case for regulation that deliberately advantages…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.