Florian Dietz

I'm an independent AI alignment researcher with 15+ years in AI. My ICML 2026 paper, "Split Personality Training," trains a second "auditor" persona into an LLM that catches hidden misbehavior the standard auditing setup misses entirely, with no measured cost to the model's capabilities. It also generalizes to detecting alignment faking. I was a MATS scholar under Evan Hubinger, and I now mentor alignment researchers through MARS, SPAR, and SAIGE. I built Comgra, an open-source tool for interpreting neural networks. At the Foresight Institute I'm developing an epistemic memory system that tags an agent's knowledge with its source and confidence. A thread runs through all of this: I look for techniques where safety and capability improve together. When the safer option is also the better-performing one, it gets adopted because it pays for itself, not in spite of a cost.

LinkedIn

LinkedIn

LessWrong

https://www.lesswrong.com/users/florian_dietz

Projects

Epistemic memory for AI agents: making the safe path profitable

Team Member

Getting Split-Personality Training ready for deployment

Team Member

No grants recorded.