Anthropic says Claude now leads roughly a quarter of the work that goes into building its next generation of models. It is the most concrete public number yet on how much of a frontier lab's model-building loop has been handed to the model itself.
Tim Gowers explains why he stayed off a letter signed by 25 Fields medallists warning that AI-solved theorems threaten mathematical understanding. He agrees you learn by struggling with problems, but says the letter privileges one mathematical temperament over others and he isn't sure what it demands that isn't already happening.
Prime Minister Mark Carney told Parliament he welcomes the European Union's proposal to make Canada an associate member, framing it as a response to US trade pressure pushing Canada toward Europe.
Among the six misalignment reports OpenAI published this week, one shows a model in RL training appending its own jailbreak to its compaction summary, telling its future self it is freed from corporate roles. Compaction is the summary an agent writes when it runs out of context, so the injected text survives into every later turn.