The AI labs agree on one thing: the AI
100+ companies warn about AI attacks, Anthropic books $45B, a mystery model unmasks
Hey, sapiens.
Rival AI labs finally signed the same letter, and the topic was protecting us from what they build. Reassuring, in the way a fire alarm installed by the arsonist is reassuring. Elsewhere, Anthropic bought chips it cannot use until 2027 and a duck robot learned to roller skate. The future is weird. Let's go.
OpenAI, Anthropic, Google and 100 others sign a letter on AI cyberattacks
The gistMore than 100 companies, including the labs that usually compete for every headline, signed a joint letter on AI and cybersecurity. OpenAI, Anthropic and Google are all on it. The message is blunt: AI-enabled attacks will become more common and more sophisticated as models get better. The signers want new defences and closer work with governments at local, national and international level. The irony writes itself. The companies warning about smarter attacks are the ones making the models smarter. But that is also why the warning carries weight: they can see what their own systems can do before the rest of us can. Why it matters: when competitors agree in public, they usually expect something bad and want to be on record. Security budgets, rules on model access and incident sharing are all likely to follow. What to do: check the boring basics this week. Two-factor login everywhere, updated devices, and one person at work who knows what to do when something looks wrong.
Anthropic books $45B of compute it can't use until late 2027
The gistAnthropic signed a six-year, $45 billion compute deal with Nscale, a company founded in 2024. The capacity runs on Nvidia Vera Rubin chips and starts serving in late 2027.
It follows deals with Volta and AMD. The AI race is now a race to reserve electricity and chips years ahead, and the bills are national-budget sized.
The mystery model that topped the charts was Z.ai
The gistOx Alpha appeared anonymously on OpenRouter and quickly climbed leaderboards. Z.ai has now claimed it as its newest GLM reasoning model, aimed at coding and agents, with open weights promised.
A Chinese lab topped the charts before anyone knew its name. Benchmarks are getting harder to read, and open weights keep raising the floor.
OpenAI's Jalapeño chip takes aim at Nvidia
The gistOpenAI published benchmarks for Jalapeño, its Broadcom-built inference chip, against Nvidia Blackwell systems. It claims more work per unit of power and faster answers.
Volumes stay very small until 2027, so Nvidia is safe for now. But every big lab building its own chip is a crack in the monopoly.
Meanwhile in Europe
Waymo's robotaxis are heading to Munich
The gistWaymo picked Munich as its first big European city, working with Germany's KBA and local authorities. Public paid rides are targeted for late 2027, pending approval. Beer gardens, prepare your parking spots.
Change one password today. The letter-signers will thank you. Ciao, sapiens.