- The Information reports that OpenAI's next model, Astra, uses recurrent depth, a method that loops layers internally so less of the model's reasoning ends up in the human-readable scratchpad, which researchers say undermines chain-of-thought monitoring (Joe)
- Astra is the first model to reach the Critical cybersecurity level in OpenAI's Preparedness Framework, and OpenAI plans to release it anyway despite earlier commitments to update the framework first (Joe)
- A rebuttal of four claims circulating in the press: that datacenter opposition is astroturfed, that AI statements about sentience should be dismissed, that human vocabulary for AI behavior is always anthropomorphizing, and that a robot sprint record does not count as real-world capability (Joe)
- At the G20 Innovation Ministerial in North Carolina, US officials pushed the non-binding Carolina Principles favoring light-touch regulation, and China reportedly signed, though the text has not been published and the December leaders' summit will matter more (Alana)
- Bernie Sanders published a Fox News op-ed citing the Hugging Face agent incident, renewing his call for AI companies to pause development and urging Trump to negotiate an agreement with Xi (Alana)
<- Back to digests
Daily digest
Nothing to see here
By Joe Rogero and Alana Horowitz Friedman
Inscrutable AI, wrongheaded media, agreement from China, and more
Read the full dispatch on AI StopWatch
From AI StopWatch, published with their permission.
The summaries and the full Spanish translation are produced automatically. The original is always linked.