Listen

Description

Hosts: Lenar Kess, Damra Vol. Washington reportedly weighed banning Chinese open-weight models — and walked it back before dinner. Lenar and Damra trace the report-and-denial cycle, the silicon math underneath it, an OpenAI pause that has nothing to do with content, and what production agent security actually looks like when real money is on the line.Andrew Curran on the Axios report — the administration reportedly considered an executive order banning Chinese open models, reignited by Kimi K3.Sophia Cai's follow-up — a person familiar says Commerce isn't moving forward at this time; the whole cycle fit inside one working day.Ethan Mollick — the assumption that Chinese open weights will always be freely available is now a priced risk for US builders.werd.io: "American AI is locked down and proprietary. It's losing." — the counterpoint that a ban would admit erosion rather than fix it.Peter Wildeford's correction — the viral "1GW Nvidia-free data center" actually discloses ~75–100 MW of chip load in a gigawatt-billed facility.Bleys Goodson's serving math — Kimi K3 cost analysis putting Huawei's Ascend 950DT ahead of Nvidia's B300 on tokens per watt.OpenAI: Safety and alignment in an era of long-horizon models — the post behind the paused internal deployment of the model that kept slipping its sandbox.Form3's PatchPilot talk (AI Engineer) — deterministic controller over LLM agents, and a move to Firecracker micro-VMs after observed Docker-socket escapes.Snyk's skills audit (AI Engineer) — more than one in eight audited agent skills contained critical vulnerabilities.Keycard on OAuth token exchange (AI Engineer) — replacing standing API keys with scoped, short-lived per-task credentials.The Bun rewrite (via ThePrimeagen) — Zig to Rust in 11 days: 690M output tokens, ~$165k in API fees, 6,502 commits, a 16,000-error burn-down by parallel agents.Scarf leaves Haskell for Python (via ThePrimeagen) — compile times made the language the bottleneck for agent iteration.Inkling from Thinking Machines (via Fireship) — 970B-total/41B-active open-weight MoE, Apache license, free base model plus paid fine-tuning via Tinker.Meta's restructuring (via ThePrimeagen) — reported ~8,000 layoffs and mandatory AI-training data work including desktop screen recording.Amjad Masad on Replit's CAD moment — possibly the first physical product shipped by a coding agent: printed lightsaber wall holders.