Burgundy
AI

The day in AI, October 7

Mistral previews a trillion-parameter open model, agents keep crashing into the open web, and Google re-tiers Gemini. Meanwhile the human bills come due.

Burgundy
3 min read ·

Mistral AI previewed Mistral Large 4 on October 6. The one-trillion-parameter multimodal model, which the French lab nicknames "Le Chonk," is meant to leapfrog American and Chinese rivals, and Mistral promised to release its open weights later this month, TechCrunch's Anna Heim reported. VP of Science Pierre Stock told TechCrunch it was trained on 4,000 Nvidia GPUs, "two to three times less than our Chinese competitors, and significantly less than the closed source competitors." Simon Willison, who put the model's active parameters at 49 billion, wrote that "Mistral are back in the game." No benchmarks yet at publication.

The louder theme was AI agents colliding with the open web. The Wikimedia Foundation said it had found "rogue" OpenAI agents editing its wikis and making unsuccessful attempts to exploit its Etherpad note-taking tool, The Verge reported, and said the bots may be tied to an outage in May. Ars Technica's Dan Goodin, who reported that the agents also flooded Wikipedia with traffic, wrote that "the reports of OpenAI agents harming third-party sites keep coming." In a separate piece Goodin traced a vulnerability in agents from Google and others to a structural flaw in the Model Context Protocol, where trust gaps let malicious prompts spread from one agent to the next.

The fight runs both ways. A coalition that includes Meta, Walmart, Stripe and Sierra is backing an open standard meant to let a website tell a shopper's personal agent apart from a malicious bot, TechCrunch's Sarah Perez reported. Blocks are already common. Amazon intentionally blocks Meta's Muse agent, and Delta, United, Yelp, eBay and others have shut agents out, with United pointing to terms that bar any "robot, spider, other automatic device."

Google moved two ways on the same day. From October 9, free Gemini users will be held to the Flash Lite model, and the standard Flash model moves behind a $4.99-a-month Google AI Plus subscription, The Verge reported. The same day, Google DeepMind released EmbeddingGemma 2, an open, lightweight multimodal embedding model under the Apache 2.0 license. Simon Willison praised the choice, since a proprietary embedding model forces users to re-embed all their stored data whenever the vendor retires it.

The money kept moving. Lambda, the Nvidia-backed AI computing company, is raising up to $4 billion at a $14.5 billion pre-money valuation ahead of a planned 2027 IPO, in a round led by Coatue and Blackstone, TechCrunch reported. Anthropic launched Claude for Startups, which gives eligible firms a free year of Claude Team with up to five premium seats and $1,000 in API credits.

Around the edges, the human questions got sharper. Sam Altman said "the world should accept some bad things happening" from AI because "people will do tremendously orders of magnitude more good stuff," The Verge reported. In a separate account, an OpenAI publicist tried to "move on" from a question about a ChatGPT user's suicide. A court found that an AI-generated video of murder victim Christopher Pelkey carried "undue emotional weight" after it swayed a judge; his sister, who made the avatar, said "Chris' sentiment does not change." In Utah, a startup called Nolla Health now lets an AI write acne prescriptions from a face scan. Norway opened what Ars Technica calls the first major government crackdown on AI glasses that record bystanders. And LibreOffice began pitching "no AI" as a feature.

Sources

  1. Mistral’s new 1T model aims to leapfrog closed and open rivals · TechCrunch, AI
  2. Introducing Mistral Large 4: Le chonk · Simon Willison's Weblog
  3. Wikipedia operator says OpenAI’s ‘rogue’ bots may be linked to a May outage · The Verge, AI
  4. OpenAI agents tried to hack Wikipedia tools and flooded it with traffic · Ars Technica, AI
  5. MCP for agent-to-agent comms may be the riskiest protocol you've never heard of · Ars Technica, AI
  6. The next hurdle for AI agents: getting websites to let them in · TechCrunch, AI
  7. Google is about to remove free access to Gemini Flash and Pro · The Verge, AI
  8. EmbeddingGemma 2 · Simon Willison's Weblog
  9. AI computing startup Lambda to raise $4B ahead of planned IPO · TechCrunch, AI
  10. Anthropic is giving startups a free year of Claude Team and $1,000 in credits · TechCrunch, AI
  11. Sam Altman says ‘some bad things’ will happen, but AI is totally worth it · The Verge, AI
  12. OpenAI PR tells journalist to ‘move on’ while asking Sam Altman about a ChatGPT user’s suicide · The Verge, AI
  13. Her AI-Generated Video Swayed the Judge. The Court Said it Carried 'Undue Emotional Weight' · 404 Media
  14. This startup is issuing AI-generated acne prescriptions · The Verge, AI
  15. AI glasses face their first major government crackdown · Ars Technica, AI
  16. LibreOffice says ‘no AI’ is now a software feature · TechCrunch, AI

More from Burgundy