Sep 11, 2026

Mathematician Andreas Thom wrote that after OpenAI's non-sofic group result he asked Mark Sellke and Sébastien Bubeck whether months of his ChatGPT discussions on expander matching had entered training or the solving process. The whole reply: 'that did not happen'. A second case after Navier-Stokes

Why it matters: Top replies note Thom hadn't opted out of training until late June, so the answer is obviously yes; another likens it to a collaborator publishing along the lines of your talks without credit. Others warn there is no evidence yet, only claims. The default-on training setting also came up

Raschka's essay: Astra is the best model he has used, unusually strong at 3D rendering and animation, at the frontier on Artificial Analysis without pulling far ahead. He explains the rumored 'recurrent depth' (looped transformer) design and asks whether looping blocks hides the chain of thought

Why it matters: The top reply says Astra was 'insane until Monday', then something changed and it now feels like Sol. Others credit Raschka with defusing The Information's 'secret technique' framing; one argues that looping a whole model on itself is hidden reasoning by definition. Slowness complaints too

An interactive explorer from Anthropic's economics team. Jobs are bundles of O*NET tasks; set how much AI augments or automates them and get 2030 US growth, unemployment and wages. Up to twice normal growth, unemployment stays in range; beyond that, knowledge workers' wages suffer

Why it matters: The top reply: if AI lets one nurse do two nurses' work, the first pressure is fewer nurses, so the framing is naive. Others say the worst scenario is 'LLMs change nothing', with none for damaged education or trust, or 'what if Anthropic is wrong'. Finding 4, capital's rising share, is a must-read