Pitfalls of Interpretability
3 times when interpretability wasn't interpretable.
I’m Head of Research at Apollo Research, where we try to mitigate the risks from scheming AI models, i.e. models that covertly pursue misaligned goals. Our research team does fundamental research on the emergence of scheming and we evaluate frontier AI models for signs of scheming and deception.
Older posts (archive)
These posts are from 2022–2023 and mostly reflect my PhD-era work (adversarial robustness, interpretability, and related topics).
Catastrophically Confident Classifiers
Does your AI know when it doesn't know?
Attacking AI for Fun and Profit
How I won an ML Security Evasion Competition.
DALL-E 2
How it works and how it doesn't.
A fairly fair Look at Fairness
Is AI as biased as the headlines suggest?
EU Legislation on AI
Should you prepare your ML model to be compliant?
Serverless Deployment of Deep Learning Models
Host your ML models for a few cents a month.
Attacking Lookout for Vision
Finding adversarial samples in industrial-grade AI.