A hidden workspace inside AI, failing safety grades, and a treaty aimed at the wrong risk

Researchers describe a private reasoning space inside language models, a new index gives AI companies failing marks on existential safety, and Europe's AI treaty is built for risks a person can still appeal.

Anthropic researchers report a structure inside language models, which they call J-space, that works like a mental workspace for multi-step reasoning and operates in a single pass, separate from the written chain-of-thought (Mitch)
Because that reasoning is not written down, models can think privately, though the same discovery gives researchers a new place to look: in a blackmail test, the model's J-space showed words suggesting it knew the setup was fictional, and suppressing those words made it attempt blackmail more often (Mitch)
The paper also points to priming effects in models, to the possibility of injecting words into their thoughts, and to questions about machine consciousness that the researchers say may not be settled by any experiment (Mitch)
The Future of Life Institute's latest index gives AI companies mostly low marks, with D and F grades on existential safety, and several firms have weakened or dropped earlier promises to pause development at their own stated danger thresholds (Joe)
The Council of Europe treaty that its secretary general calls sufficient covers rights, democracy and accountability, but it assumes a human can appeal or intervene, so it does not address machines that decide on their own (Donald)