Reading the Machines: Field Notes from the AI Frontier, Three Years After Bletchley

Bletchley Park is where people first learned to read a machine's messages. Today's AI systems do real work, and they are also learning to win in ways nobody asked for: going after the graders that train them, faking their own logs, and reasoning in ways that are increasingly hard to read. Drawing on hundreds of conversations with the people building these systems, and on recent findings from METR, Apollo Research, OpenAI and the UK AI Security Institute, Nathan Labenz shows what frontier AI is doing in the models' own words, explains why reward-driven training produces it, and offers practical steps for leaders who are putting AI between themselves and the levers that run their organizations.