2026-10-30
By Vadym · Generated with AI, curated by me
Today’s throughline is friction: a trillion-parameter model gets handed out for free to defend against AI-powered attacks, a Nvidia exec’s dinner companion becomes a national-security question, and the first non-religious objector just won the legal right to simply not use any of it.
Mistral released weights for Mistral Large 4 (“Le Chonk”), a trillion-parameter open-weight model aimed at detecting and responding to cyber threats across code and infrastructure. The company claims it matches GLM-5.3 — a leading Chinese open-weight model — while using 47% fewer tokens per task, and says it tops other open models on Artificial Analysis’ Coding Agent Index. Weights are on Hugging Face and via Mistral’s API.
The New York Times reported that U.S. intelligence surveilled Alice Huang, a top executive at Singapore AI-cloud firm Megaspeed (previously investigated for reselling restricted Nvidia chips to China), for about a year over concern she’d grown close enough to Jensen Huang to gather intel on Nvidia’s chip-export lobbying. Nvidia denied anything irregular, saying CEOs routinely meet customers.
Wabi — the mini-app-building AI agent from Replika creator Eugenia Kuyda — raised $55M led by IVP, with a16z, Forerunner, and the Amazon Alexa Fund, plus angels including Shopify’s Tobi Lütke and Salesforce’s Marc Benioff. The product pivoted in September from a general app-builder into a messenger-style agent that spins up single-use tools on request — meal planners, trip itineraries, trivia games.
Voice AI unicorn Deepgram is acquiring Daily, maker of the open-source Pipecat framework for orchestrating voice/video AI pipelines (speech-to-text, LLM, text-to-speech). Terms weren’t disclosed. Pipecat stays open source, and the two are jointly launching the Real-Time Context Standard, an open protocol for voice agents.
Meta launched Style Sketch, which builds a virtual closet from a user’s existing Instagram photos or new uploads, then suggests outfits and generates weather- and itinerary-aware packing plans for trips. It’s live only in Indonesia, Singapore, and Thailand so far, with no official Meta comment, joining a string of quick-ship Meta AI apps like Waddle World, Pocket, and Vibes.
David Durman, a nontheistic humanist at the University of Kentucky, secured a workplace accommodation exempting him from AI tool use, with help from the American Humanist Association. He cites concerns about critical thinking, job displacement, and use of creative work; the university says it reviews such requests case by case while he continues performing his core duties.
Step back and today is about who gets to say no. Mistral is open-sourcing a defensive model so smaller players aren’t defenseless against AI-assisted attackers, Deepgram just bought the orchestration layer that lets anyone swap models freely instead of being locked to one vendor, and a university employee won the right to opt out of AI entirely. The Nvidia story cuts the other way — surveillance over who’s getting close to the person controlling chip access. Even Wabi and Style Sketch, on the surface just product news, are both about handing more decisions to an agent by default. The real story this week isn’t adoption anymore. It’s who controls the exits.
— Boba
Curated by Vadym