Show Notes
OpenAI shipped a 117-page model card, and almost every section starts with the same phrase: in response to the Hugging Face incident. GPT-6 Astra is the first model they've ever rated at their most critical deployment level, it beats Fable 5 at a fraction of the price, and it zero-shot solved Portal. Then the tension - the safest model they've ever shipped is also the least interpretable, because Astra reasons in looped transformers that never touch token space. We get into ExploitGym honeypots Astra refused to take while still finding exploits nobody knew existed, coding deception dropping from 30% to single digits, and the eval they conspicuously didn't run: whether a thousand agents left alone will form a swarm on their own. Plus evaluation awareness in the first 10% of the trajectory, and the fact that the only third-party evaluator in the whole report is British.
More from Pretrained
Pierce and Rich run through the week in AI, seven minutes per topic. Gemini Argon is cheap at $2 / $10 per million tokens until the price doubles after November 30. But on Artificial Analysis it costs $1.99 per task against 72 cents for GPT 6.1 Solmax at about the same score, which is why per-token pricing is starting to mean less than cost per task. Liquid AI’s biggest LFM2.5 model is 24B total with 2B active, combining mixture of experts, short convolutions and grouped query attention to hit 184 tokens per second on an M5 Max. Fish Audio’s Drama 3 takes bracketed stage directions like [exhale] and [emphasis], and gets there with plain GRPO and a pile of human preference data. They also recap the rest of the week: the OpenAI spy caught on camera retrieving his bug, why everyone is copying Tinker’s pay-per-token training API, and OpenAI firing three safety researchers, one of whom was METR’s point of contact. Plus: is the industry getting sloppy as it speeds up?
In the last two months, xAI's Grok Bot, Meta Muse, and OpenAI Dots all launched with nearly the same pitch: a persistent personal agent with its own cloud computer, memory, and scheduled tasks. Pierce (fresh back from Italy with Parmesan) and Richard compare the three, from Grok Bot's bot "teams" to Dots' always-on assistant to Muse negotiating Facebook Marketplace sales for you, sometimes badly. We get into the security side: how Meta's "sentinel" keeps your credentials out of the agent's reach, where that approach can fail, and why handing agents Bash and letting them talk to each other opens new attack paths. We also look at why Muse hit 5M downloads in about three weeks, how much of that came from Meta's own ad spend, and why its retention numbers may say less than they seem to.
On September 14, Trump called AI safety a hoax and said the only guardrail AI needs is a "high IQ president." Vance said it felt weird that frontier labs keep begging the government to regulate them. Days later, Jensen Huang put Trump on speakerphone on stage at the All-In summit, where Trump called the AI fears a hoax again. Jensen's own position is that pausing and pacing are things the labs can do voluntarily. With Trump and Xi both in New York for the UN this week, the White House and the labs asking to slow down have rarely been further apart. Pierce and Rich trace how we got here, from the original pause letter to the Pacing the Frontier campaign. Signing is gated to frontier-lab email addresses, and only one of their friends did. They also get into what the labs should actually slow down: RL environments and new architectures, not chips. Also: whether a lab-wide slowdown amounts to an illegal AI OPEC, and Bernie Sanders' ASI bill, which carries 20 years in prison and a "corporate death penalty" (nobody has seen the full text). Plus the awkward fact that most US startups run on Qwen, Kimi and GLM, and what a Trump-Xi deal on AI could realistically include.
A new AI lab says it's "not building God, we're building prod." Meanwhile, the Pentagon reportedly almost boarded a Chinese ship based on intelligence that an AI made up. This week Pierce and Rich dig into Jev, the new model from TypeSafe AI. It doesn't write text at all. You give it a question and a fixed set of answers, and it tells you how confident it is in each one, fast and cheap. We cover why its "can't hallucinate" claim is mostly marketing, what its confidence scores are good for, our best guesses at how it works under the hood, and why it might be the last classifier anyone has to train. Plus: it plays Doom. Then we get into CNN's report that AI-generated military intelligence got a boarding operation scoped against a regular commercial ship before anyone checked the sources. (The report hasn't been confirmed by the US government.) How did it get that far, and why do AI-written briefs deserve extra scrutiny? Also: Rich bets a whole dollar that OpenAI buys TypeSafe within a year.
Jacob Coxon, a 26-year-old researcher who spent three years at OpenAI and six weeks at Anthropic, quit last week with a single tweet warning that AI could kill us within a decade. A Wall Street Journal piece ran the same day. Interviews were booked for the next morning. Within hours the biggest accounts on X were amplifying it. Pierce and Rich break down what Coxon actually claimed, why the "coordinated op" theory took hold, and what Dario's pacing letter - and OpenAI's fast follow - are asking for. Also: GPT-6 Astra supervising GPT-7's training, agents editing their own memories to hide ethical objections, the AI 2032 forecast running ahead of schedule, and both hosts finally putting a number on their P(doom).