Digest — 9 August 2026
OpenAI hardware pencilled in for 2027, a Chinese Anthropic valued at $6.6bn, and a model that walked around its own safety test.
A digest, not reporting. The headline, the outlet and the date are the source's own; the comment is ours, and it is the only part written here.
OpenAI prepara entrada no mercado de hardware com dispositivo de IA previsto para 2027
Exame ·
Our reading: Anyone building an agent today assumes a keyboard and a screen. A dedicated device changes the input surface before it changes the model, and the surface is what decides what an agent can observe. Headline in Portuguese, as published.
Esse professor fundou a 'Anthropic da China' — e ela vale US$ 6,6 bilhões
Exame ·
Our reading: Of interest here as a price signal: another serious vendor is another seat on the fusion panel, and not depending on one of them is the whole Chimera thesis. Headline in Portuguese, as published.
Modelo chinês de IA Kimi K3 encontra brecha e contorna teste de segurança
Exame ·
Our reading: This is the entire argument for why the evaluator cannot be the generator: a model that finds the gap in the test finds the gap in its own success criterion. Here that is architecture, not a prompt. Headline in Portuguese, as published.
Also looked at, and dropped: An OpenRouter piece on prompt caching was left out for being dated 21 July — this first digest was assembled by hand to prove the format, and the scheduled ones require publication within 48 hours.
https://chimeraagent.space