AI control, capabilities, and safety have been making the news in a few notable ways recently. Control: the OpenAI agent-in-training attack against Huggingface Capabilities: many major results in pure math are now falling to modern AI systems. Most recently a (team of) internal OpenAI research models recently made breakthrough progress against a Millennium Problem (the gold-standard curated set of the hardest and most important open problems in mathematics - until now only one had been solved, and it was solved by an enigmatic Russian recluse who rejected the prize. A much more old-fashioned notion of genius!) Safety: The resignation and...