Google DeepMindConjecture Machines, July 2026
Testing whether an idea survives contact with reality is the slow and costly part, and in most fields that gap is widening.
OpenAIMisalignment reports, 16 September 2026
One of its own models could not retrieve a figure, made one up, and presented it as coming from the source.
Anthropic17 September 2026
A measurement that outsiders are meant to trust has to be complete, accurate and enforced by the technology itself.
Prof. Matthew SchwartzOn Anthropic’s site, 1 October 2026
Even with every monitor in place, the automated checks on Claude’s physics still could not be trusted.
Periodic LabsOn its home page
Physics is an environment where Super Intelligence can be verified, and progress has been fastest where results can be.