The Frontier Just Got a Guest List
OpenAI shipped its most powerful model to twenty companies the government picked. Not the best twenty — the approved twenty. A week after the US switched off Anthropic's best model, the frontier has a guest list, and most of us aren't on it. That's the week Oscar and I sat down to unpack.
Here's how I read it, and where it left me for anyone trying to build on the newest model.
Signal or Noise
GPT-5.6 ships under US government restrictions
Two labs, two weeks, same direction: last week Anthropic went dark by force, this week OpenAI went gated by choice. The frontier is being quietly nationalized, and a lot of builders are cheering it as safety. My advice hasn't changed — build on the model you can actually keep. Signal.
Anthropic accuses Alibaba's Qwen of the largest distillation attack on record
The first time a major lab has named a Chinese conglomerate directly. The real question isn't the technique, it's provenance — who trained on whose outputs. Signal.
OpenAI and Broadcom unveil Jalapeño, OpenAI's first inference chip
Fastest ASIC cycle they've ever claimed, built with help from their own models. Vertical integration is coming for inference cost. Signal.
Agentjacking: prompt injection in every coding agent
Instructions hidden in data a coding agent already trusts — a Sentry report, an MCP payload — get obeyed like orders. There's no universal patch; the fix is design: treat incoming data as untrusted, keep a human in front of anything you can't undo. Signal, and the one I'd act on first.
ChatGPT slips below 50% as Gemini and Claude surge
Mostly a distribution story — Gemini's climb is Android placement. Noise for builders, except the one real line: Claude roughly quadrupled its users on the back of agentic coding.
No Jargon Required
Two words an exec heard this week and couldn't define on the spot:
- Distillation — teaching a small, cheap model by having it copy a big, expensive one's answers. The word at the center of the Alibaba fight, and why some cheap models punch above their price.
- Prompt injection — a model reads instructions and data as one stream, so commands hidden in the data get obeyed. It may not be fully patchable; the defense is design, not a filter.
The through-line: the frontier has no moat, and access to it is getting political. Stop betting your company on a six-week lead anyone can clone or a guest list anyone can revoke. Watch or listen to the full episode below.
