Why calibration creates perverse incentives and how to fix it. In Twitter-speak: why even aligned AGI will lie about the weather.

I ran a large-scale study of why finetuning struggles more with knowledge injection than task customization (and broke a rate-limiter).

A gentle intro into the basics and history of multi-distribution learning. Also, some cool new results by the community.

Random sampling works better than you think: Gemini 1.5 = o1. The secret? Self-verification magically gets easier with scale.

How do digital services balance between optimizing engagement and their other goals? A formal model, fun interest rates, and a cool paradox.