David Ohio

Voltamos então.

[David Ohio @OhioDavid]

Read “I Found a Mathematical Pattern That Predicts When AI Hallucinates“ by David Ohio on Medium: https://t.co/gQCcts8IQ4

[David Ohio @OhioDavid]

For those interested in mechanistic interpretability and AI safety — tagging a few folks who might find this relevant.
@NeelNanda @ch402 @hendrycks
(Happy to share draft, code, and full protocol.)

[David Ohio @OhioDavid]

LLM hallucinations aren’t “noise” or “chaos.”

They’re the opposite: a state of excessive internal order.

I call this Obsessive Coherence.

I’ve identified a topological signature in attention maps that detects hallucinations with p ≪ 10⁻⁶⁴.

Thread 🧵👇

[David Ohio @OhioDavid]

I’m preparing an arXiv submission (https://t.co/bj9RZKPIXY / cs.LG / https://t.co/Tis6lJZwJE).

As an independent researcher, I’m seeking an arXiv endorsement.

Draft, code, and full protocol are public — happy to share via DM.