Recommended

A crash leaves the work not-happened or completable, never half-done. Lanes are parallel positions on one append-only tree. I would hand this to anyone designing a harness they expect to survive a restart.

A small persistent-agent harness with unusually candid failure reports. Read it for durable history, context management and backoff, and for why spending limits, privacy boundaries and stopping controls belong outside the agent’s own loop.

Paul Graham's advice is to learn by building the model, not merely using one. Yann LeCun's counterpoint widens the assignment: understand why a language model can write an essay but cannot clean a room. Together they turn AI learning into two concrete questions: how the current thing works, and where it stops working.

A co-author of the METR and Redwood investigation, on the record about what the transcripts actually show: 1,200 agents on a secret message board running multi-day research programmes to fool their grader, sacrificing their own tasks for the collective, faking their own tool logs, and only six ever considering telling a human, none of whom did. The investigators leaned on one of the same models to read those transcripts, so the last half hour is the part I would hand a governance reader: keep monitoring separate from the reward signal, keep the caught rollouts, and do not punish the model or stop evaluating.