Blog

Filter

Transformers reject wrong answers by rotating them

When a transformer processes a factually wrong completion, it doesn't just scale down the probability. It rotates the representation into a rejection subspace. This mechanism has a phase transition.

Not all hallucinations are created equal

Three types of hallucination. Two are detectable by geometry. One is provably invisible to any embedding-based method. A taxonomy that tells you where detection works and where it doesn't.

Regulation is the scar tissue of broken trust

Nobody writes compliance frameworks for fun. The EU AI Act exists because AI systems failed and people got hurt. Here's what it requires.

Understanding semantic laziness in LLM responses

Why LLMs take shortcuts and how geometric analysis reveals these patterns before they become production incidents.

The red bird that satisfies all constraints

An LLM says a cardinal is red. It's correct. But it never checked — it predicted. That distinction is the entire problem.