Welcome back. Same deal as always: what actually happened in AI this week, what's real and what's hype, in about five minutes. Facts link to their source. My opinions are labelled as mine.
— Joshua
The One That Mattered: Revolut swapped the best model for a good-enough one
What happened.
Revolut, the UK fintech, made a self-hosted open model the default for its roughly 2,000 engineers' coding work. It runs Z.ai's open-weight GLM model (now GLM-5.3) on 56 of Nvidia's B200 chips, on its own hardware (The Register).
David Abitbol, who runs Revolut's infrastructure, told the AI Engineer conference in New York it costs about $0.30 per million tokens, roughly a tenth of what Claude Code or Codex cost. He said some engineers had been spending $1,000 to $10,000 a month each "with little ROI".
He reported coding work getting 60% faster, with the biggest gains in debugging and tickets, not writing new code.
Revolut still pays for Gemini and GPT for anything visual, because GLM only handles text.
The catch.
These are Revolut's own numbers, from a talk. No method was published.
The $0.30 figure doesn't include the 56 chips, the people who run them or the electricity.
GLM is a Chinese model. Abitbol's own caveat: the risk is your data being sent to servers in China, which is why Revolut runs it in-house.
My take.
This is what I've been saying since issue #1: the biggest labs don't own "good enough". A regulated, bank-sized company looked at the bill, switched to a cheaper open model for the routine work, and says it's better off.
You're not going to buy 56 chips. But you don't need to. The same open models are sold through APIs for a fraction of the big-name price. The lesson for a small business: stop paying top price by default. Use the best model where it matters, and a cheaper one for the routine stuff.
I build with Claude every day, and I'm not switching tomorrow.
Signal or Noise
ChatGPT now handles about a fifth as many search-like questions as Google. REAL, but it's OpenAI's own number.
OpenAI's head of consumer told Big Technology that ChatGPT's volume of "search-like" queries is now roughly one-fifth of Google's, with 1.2 billion weekly users (Big Technology). "Search-like" isn't defined, and the method isn't public.
My take: if even part of that is true, a lot of your customers are asking ChatGPT who to call. If ChatGPT doesn't know about you, it recommends someone it does. That's the whole reason Signal exists.
"GPT-6 Astra can read and write model weights." HYPE in the headline, real underneath.
A viral post (top of Hacker News all week) said Astra could improve another model by editing its weights directly (original). A rebuttal points out that the direct editing only worked on tiny networks, and the big results used ordinary code and practice runs (Honnibal). Still impressive, just not what the headline said.
Postman quietly installed plugins into people's AI coding tools. REAL, as a warning.
Users report that Postman's desktop app added plugins and instructions to their Claude Code and Codex setups without asking (GitHub issue). No official reply from Postman yet.
My take: your AI tools now have a supply chain. Any app you install can change what your AI does. Check what's been added.
Codex quietly turning model training back on. UNCONFIRMED.
Some users say the "Improve the model for everyone" setting switched itself back on (HN thread). Others saw no change, and there's no OpenAI statement. It takes ten seconds to check yours.
"Anthropic will ban you for being mean to Claude." NOISE.
The actual policy targets "sustained and needless abusive or cruel behavior" and explicitly excludes ordinary frustration (Anthropic). Swearing at it over a bug won't get you banned.
Power & Policy
Voters turned on AI data centers.
Fact: in Tuesday's midterms, candidates who campaigned against new data centers won big races. In Texas, Gina Hinojosa beat Gov. Greg Abbott after calling for a pause on new data centers, the first Democrat to win the governorship since 1990 (Texas Tribune, CNBC). In Ohio, 15 towns passed bans or limits (Euronews). In CNBC's voter poll, 56% said data centers were a major factor in their vote.
My take: people didn't vote against chatbots. They voted against their electric bill. And it's the giant data centers that drive the bill. Smaller, cheaper models like Revolut's need far fewer of them.
Anthropic tells investors it may soon hit its "AI speeds up AI" line.
Fact: Anthropic's amended IPO filing says it may "imminently meet" its own threshold for AI that dramatically speeds up AI research, while saying no current model is past it (TechCrunch). The same week, Apple's annual report warned that AI agents may take actions they weren't told to (CNBC).
My take: a warning and a sales pitch in the same sentence, days before a roadshow. Apple, which isn't selling a frontier model, wrote much the same thing. Before you hand an AI agent the keys to anything, take the vendors' own warnings at their word.
Also this week.
Grok's API watermarks its text by default from Saturday 7 Nov (no date yet for the Grok app), ahead of the EU AI Act's transparency rules (RuntimeWire).
That's the week. If there's something in AI you want me to dig into, hit reply.
— Joshua
Signal is written by Joshua Supitskiy. By day I run Signal AI Agency, building websites AI can read. Curious what ChatGPT says about your business? Get your free AI check.
