A Chinese lab just handed the world a frontier AI model and asked for nothing back! 2.8 trillion parameters. A terabyte and a half. Free to download, and within half an hour, it was the number-one model on Hugging Face. So now you own it. Congratulations. But it hurts a little to say: Owning it turned out to be the easy part.
Episode 4. The license nobody read. You could have 64 high-end chips, or just wait a day and a half and eight companies were serving it. The 90-second swap into Claude Code. The nonprofit that hired a lock-picker. Uber’s 50,000 agent sessions a day that the security tools can’t see, and why defense is about to be the biggest product category in AI. The one story I can’t stop thinking about? Your encrypted app isn’t encrypted if an AI is watching your screen. The lock on your front door is perfect. But the AI is already in your living room.
Tell me what you think. I’m reading everything.
The swap essay
Kimi K3’s weights, license, and model card
Simon Willison on the release
The license, read closely (revenue-tiered, not open source)
Third in the world, at half the per-task cost of Opus
Nathan Lambert on the open-weights escalation
1.56 terabytes in 96 shards
Kimi K3 on OpenRouter
Moonshot’s own Claude Code guide
The AI Security Leaderboard
FAR.AI’s announcement
The methodology
Hugging Face’s incident report (still the best short story of the year)
Anatomy of the intrusion
Hugging Face’s own lesson
Uber’s 50K sessions a day
The peer-reviewed paper behind Uber’s number
Solving the agent identity crisis (what Uber built instead)
Harbor (AI as a browser capability - your models, your credentials, your tools)
Qwen-CUA (screenshots in, clicks out)
The paper (it beats the best closed models at operating real software)
Meredith Whittaker at SXSW (brain in a jar, and the blood-brain barrier)
Her “dangerous honeypot” quote
By default, Signal doesn’t recall
Mozilla’s State of Open Source AI report
