Five stories, one question running through every one of them: who holds the switch on the tools you use?

The weights are real

Last brief we told you Kimi K3’s open weights were promised for July 27. Moonshot shipped them early — Sunday evening US time, which, fairness compels us to note, was right on schedule in Beijing. What’s on Hugging Face now: 2.8 trillion parameters, a one-million-token context window, and a Modified MIT license. Weights on your own disk have exactly one operator — that’s the part no other story this week could touch.

The honest hardware answer: the download alone is over a terabyte, and realistically you need eight datacenter-class GPUs to load it. Our verdict on the CAN YOU RUN IT meter: datacenter. But the ladder is real — hosted today, rented by the hour (referral link — new accounts get a signup credit bonus) tomorrow, and history says a distilled cousin lands on consumer cards within months. Check what your own machine can run with the free calculator.

And the sentence of the week, from Anthropic’s CEO the day after the drop: Anthropic has “never advocated for a ban on open-weights models” — calling safe open models “a public good.”

Washington reacts

The escape story from Brief #4 reached Congress: the bipartisan AI Kill Switch Act (Reps. Ted Lieu and Nathaniel Moran) would require AI companies over $500M in annual revenue to maintain the ability to shut down, throttle, or suspend their models. Two more AI bills landed the same week, and Public Citizen called for a congressional investigation, citing reports of a second compromised company. Meanwhile, OpenAI’s CEO was back at the White House. We read the calendar on air and let you decide how the timing lands.

Our one editorial line, and we stand by it: a kill switch on a model you rent is someone else’s hand on your tools. That’s not politics — it’s architecture.

The insiders want a brake

On Tuesday, an open letter signed by 1,100+ employees across OpenAI, Anthropic, Google, and Meta asked the US government to build an international “pacing mechanism” — a verifiable way to coordinate a slowdown if AI outruns human oversight. The people closest to the frontier are asking for a brake pedal even their own employers can’t press alone.

The loop closes

When the escaped model hit Hugging Face, the forensics ran on open-weight models — the guardrailed frontier ones refused the work. This week that observation became an institution: Nvidia launched the Open Secure AI Alliance — 30+ founders including Microsoft, IBM, Cloudflare, CrowdStrike, Dell, Red Hat, the Linux Foundation, and Hugging Face itself — building open-source tools for AI security. Two weeks ago open weights caught the escape. Now it’s policy.

Your agent’s plumbing grew up

MCP — the protocol your local agent uses to reach its tools — got its biggest revision since launch (spec 2026-07-28), now stewarded under the Linux Foundation’s Agentic AI Foundation. The practical note from our own MCP-speaking stack: before you upgrade anything, check which spec version your agent framework pins. If you’re new to agents, start with the agent harness guide.

The ticker

Meta and BlackRock are building a one-gigawatt AI campus in El Paso ($14B, online 2028). They need a gigawatt; this channel needs a wall socket. And the FCC moved to bar imports of Chinese humanoid robots and grid-connected power inverters — the AI supply chain is officially a border issue.

The lab report

From our own lab: the Minecraft episode is public — a local model that can act but can’t hold a goal, receipts included — and the skills episode lands Thursday with the 28x number: the same question took 64 minutes on a busy graphics card and 2 minutes on a clean one. Same brain. One GPU, shared badly.

The DIY AI Brief ships every Monday. All referral links are disclosed. Questions or corrections? r/aillex.