The November 5 AI shock is not what you think
Not regulation, not a crash: memory and debt. And Herbert on Apple's chips, against MLX.
Not getting this letter yet? It goes out every Friday by email, in English or French. Subscribe to get it automatically.
SubscribeHello,
On the menu this week: the fourth article of the Herbert series, where it runs on Apple's chips, and five reads.
The reads: <strong>the November 5 shock on Wall Street</strong>, which the press blames on the elections and which owes more to memory prices and debt, OpenAI producing 948 root exploits on Linux for 1.2 million dollars, <strong>Astra working less well as soon as it is told the task is about safety</strong>, METR measuring that Anthropic is moving twice as fast with the <strong>evidence redacted</strong>, and Qwen 4 announcing itself while Hydrogen-1.0 is there to download.
<span style="color:#2d7d9a;">A word about the coming weeks: between my travels to give training courses and the work that goes on with Herbert and the harness, the news will not always come with a podcast. I would rather spend that time on substance, because things are moving right now and it seems important to me. Thank you for bearing with me.</span>
The previous article made Herbert speak CUDA. This one brings it to Apple's chips, in Metal, with no .metal file and no MLX, against mlx-lm and llama.cpp on a Mac Studio M3 Ultra, a MacBook Air M1 and a MacBook Air M5. Six Qwen models up to 35 billion parameters. On the M3 Ultra, Herbert is ahead at decode everywhere, by 9 to 34%, and ahead at prefill up to 8,192 tokens. On very long prompts I got wrecked before coming back level: the text says how.
Herbert on Metal, without a .metal file or MLX, against mlx-lm and llama.cpp on a Mac Studio M3 Ultra. Six Qwen models up to 35 billion parameters. What the engine decides at load time, the night it picked wrong kernels. Long prompts, where I got wrecked before coming back level.
The idea was to write a single Herbert that adapts itself to all three: at load time, it looks at which chip it runs on, measures what it can do and adjusts to it. That is the heart of my work, but this time in Metal. The kernels, the tiles, the block sizes are chosen on the machine, not in advance; the number format, bf16 or f16, is chosen at build time, depending on what the chip can do. That is what this whole article is about.
On Thursday, November 5, 2026, the S&P 500 lost 4.4%, its worst session since April 2025. The Nasdaq entered a correction, and Nvidia gave back its gains for the year. The press blames the midterm elections. There are three drivers, and the Texas pause the market fears has existed since August 3.
On November 4, 2026, OpenAI published the results of an agent pipeline that reproduced 1,724 Linux kernel vulnerabilities and wrote 948 privilege escalation exploits, at 1,279 dollars each. The next day, GPT-6 Cyber was classified as "critical" and restricted to verified defenders. That same week, Moonshot removed cybersecurity from the training of its open model.
OpenAI report from November 4, 2026: on identical code tasks, GPT-6 Astra achieves 10% of the criteria when the task is framed as safety research, 16% when it is framed as capabilities research. No refusal, no written decision: a doubt, and less work. The monitors saw nothing.
From August 12 to October 29, 2026, METR had access to Anthropic's internal data. The conclusion, published on November 5: capabilities are progressing about twice as fast as before October 2025, and AI explains most of it. Anthropic removed from the public report the majority of the evidence that METR considered most informative.
On November 6, 2026, at noon UTC, Qwen published its first official teaser; no weights online. On November 2, Deep Principle released Hydrogen-1.0, a GLM-5.3 post-trained on its lab notebooks, under the MIT license, claiming to beat GPT-6 Astra on FrontierScience-Research. And GLM-5.3 is running at T-Systems in Munich.
This letter went out by email to those who asked for it. All the letters