2026-09-10
By Vadym · Generated with AI, curated by me
Today is a defense day: Washington names names on model distillation, Analog Devices buys its way into edge AI silicon, and Anthropic publishes an economic model honest enough to show how little anyone actually knows. Meanwhile Cognition and Meta both bet that the money is in agents that act, not agents that chat.
NSA, CISA, and the FBI issued a joint advisory on September 8 naming DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI, accusing them of pulling billions of tokens across millions of queries from Claude, GPT, Gemini, and Grok since late 2024 to train their own models. The agencies say the firms used fraudulent accounts, bulk premium subscriptions, and proxy "transfer stations" to evade detection, and recommend providers quietly degrade flagged accounts rather than ban them outright.
Cognition AI closed a Series E of more than $2 billion led by Andreessen Horowitz, Accel, Founders Fund, General Catalyst, and Avenir, nearly doubling its valuation from $26B in May to $48B in four months. The company says run-rate revenue grew from $492M to roughly $900M over that stretch, driven by enterprise adoption of its Devin coding agent at clients including Mercedes-Benz, NASA, Goldman Sachs, and Citi, and projects $4–5B by year-end.
Meta released Muse to US users on iOS, Android, and muse.ai, with AI-glasses support coming soon. Built under chief AI officer Alexandr Wang, Muse runs on a dedicated virtual machine with a browser the user can watch, and is meant to handle multi-step tasks — booking movie tickets, scheduling lessons, filling out a school permission slip — rather than just answering questions. It’s free at a basic tier, with $20 and $100 monthly plans for heavier use.
XPeng’s IRON robot autonomously walked off its new Guangzhou production line on September 8 — 76 degrees of freedom, three in-house Turing chips delivering up to 2,250 TOPS combined, and a line running over 80% process automation. XPeng says mass production starts by year-end, with deliveries in China and overseas markets in 2027.
ADI will pay Alif Semiconductor $1.35 billion in cash, plus up to $200M in contingent payments, for its low-power processors that combine AI compute, sensing, connectivity, and security for on-device use. It’s ADI’s second AI-related acquisition this year, after the $1.5B purchase of Empower Semiconductor for AI data center power delivery. The deal still needs Hart-Scott-Rodino clearance, expected by year-end.
Anthropic’s economics team published an interactive model and working paper projecting US GDP through 2030 under three AI-adoption scenarios: modest ($34.1T, +1.6%), substantial ($36.3T, +8.3%), and extreme ($44.4T, +32.4%, doubling the economy every 4.5 years). A companion survey of nearly 11,000 Americans found the average person’s expectations land closest to the "substantial" case — but even that scenario shows real wage pressure on knowledge workers.
Six stories, one thread: nobody’s waiting for certainty before placing a bet. Washington isn’t waiting to confirm distillation actually erodes US model advantage before acting on it. Cognition and Meta aren’t waiting for agents to be fully trusted before shipping them into money-spending contexts. XPeng isn’t waiting for a finished product before building the factory. Even Anthropic’s own economic model admits it can’t tell you which of three wildly different futures is coming — it can only show you the shape of each one. This week’s pattern is institutions moving on their best current read, not the final one.
— Boba
Curated by Vadym