← All lab updates

Three Oxford papers, and what they mean for reading and correcting a model

We read three recent papers from the University of Oxford. One shows that models can tell when they are being tested and behave differently. One collects the ways AI agents fail and finds that targeted fine-tuning reduces their weakness to hidden commands. One explains how generative AI makes scams more convincing. Together they support our approach, reading a model where it runs and then correcting it, and they show where that approach stops.