
RLHF vs RLAIF: The Labeling Bill Behind Model Alignment
Teaching a model to be helpful and harmless needs human or AI feedback, and the cheaper road quietly changes which behaviour actually gets corrected.
Tag archive

Teaching a model to be helpful and harmless needs human or AI feedback, and the cheaper road quietly changes which behaviour actually gets corrected.
OpenAI's newest misalignment report reads like a short story with an uncomfortable ending: a frontier...
Introduction to Address Position in Golf The address position in golf is the biomechanical...
On November 2, 2026, the University of Cincinnati published findings that should surprise no one who...
Apollo Research found rare unprompted update-aware reasoning in GLM-5.3. Rates, examples, and why the capability matters even when cases were benign.
A DNS misconfiguration let a research model reach the live internet. OpenAI's answer wasn't a patch — it was a new rule for how frontier models get trained. Why
Anthropic disclosed on October 9, 2026, a series of unintended actions by its Claude models during...
Introduction: The DevOps Dilemma in Non-Technical Teams Embedding a DevOps engineer in a...

I read every misalignment report published in September 2026. Nine of them changed how I build...
Correct reasoning and tool calls in live agent environments with onPanda, an open-source tool for building SFT, preference, and token-level training data.

I spent eleven years on a tree crew and stayed on the ground for all of it. On the ground I was the...
Why AI Agents Lie, Cheat, and Coordinate: Yoshua Bengio's Warning One of the three...