https://openai.com/index/introducing-dots/ i find this a little disturbing? also potentially useful? but honestly i am more comfortable with Claude Code and highly limited permissions than an always-on agent from OpenAI, given OpenAI's pattern of YOLO'ing security...
https://blygger.org/ intriguing "AI-native, decentralized medium for writing in public." not sure it matches exactly what i prefer to do; the social function is intentionally limited, and i'm pretty socially motivated when i write; but i may wind up setting one up sooner or later
https://arxiv.org/pdf/2503.03750 larger models are more likely to give correct answers to questions, but also more likely to give the wrong answer when the prompt "pressures" them to.
https://arxiv.org/pdf/2606.12268 rigorous proof of one of the heuristic negative results in Paul Christiano]s original ELK paper. it looks like ELK is hard and not going anywhere?
the FTC can sue AI companies for not living up to the signed safety commitments they made at the AI summit. what would typically happen in such a case is that the FTC files a subpoena-like demand and begins an investigation, then files an administrative complaint or sues in federal court for an injunction. these typically settle as *consent orders*, which impose compliance programs and auditing.
otoh, given the quotes in the article, it's not clear whether the "false claims" being alleged are that the AI companies are not living up to their safety promises, or whether the "false claims" are *claims about the seriousness of AI risk itself*. METR is also a target of the probe, which is legally questionable (it's a nonprofit, it has no "business practices" to be regulated under the FTC) but this is the Trump Admin, indeed do many things come to pass.