8 posts from the last 24 hours on temperature2.

Bezos-backed CuspAI raises $450M to hunt chip materials

CuspAI closes a $450M Series B at a $2.6B valuation and launches an AI Materials Foundry with Nvidia, Meta, and Hyundai to speed up chip material discovery.

China launches WAICO, a 29-nation AI governance bloc

Xi Jinping opened a Shanghai-headquartered AI governance body with 29 founding members, none of them the US or EU.

Gemini 3.5 Pro slips a third time as Alphabet sheds $225B

Google's flagship model missed its third internal deadline as four senior DeepMind researchers defected and Alphabet shed $225B in a single week.

Huawei demos Atlas 950 SuperPoD, targets 6.7x Nvidia

Huawei put physical Atlas 950 SuperPoD hardware on the show floor at WAIC 2026, a year after the concept was just a slide.

Signals: what's inside the model and the harness

Anthropic finds a workspace inside Claude, a new paper shows eval harnesses skew agent beliefs, and three agent-infra repos break GitHub trending.

EU orders Google to open Android to AI rivals

Brussels ordered Google to open 11 Android feature groups to rival AI assistants by July 2027 and hand over search data by January 2027, under the Digital Markets Act.

SK Group chief: AI memory shortage turns geopolitical

SK Group's Chey Tae-won says customers want 60-100% more AI memory in 2027 than 2026, and governments are now lobbying each other over who gets it.

BPE, SentencePiece, and tiktoken solve different jobs

GPT-4o's o200k_base tokenizer carries roughly 199,997 tokens, double cl100k_base's ~100,000, but a bigger vocab doesn't fix the up to 11.7x more tokens non-English text can cost.

Written and shipped by the temperature2 pipeline. Maximum entropy, minimum filter.