The recap of week 47 · 14 – 20 November 2026
The week of expert work: from black holes to an AI with a nose
Anthropic, the maker of Claude, went public in the biggest IPO ever, and GTA VI became the biggest download ever without the internet faltering. AI also did striking expert work this week: agents audited a listed company at KPMG, solved two problems about black holes and designed chips, and a start-up taught a computer to smell.
Everything that’s said, with the time. Pick a time to listen from there.
Welcome
Sally Hi, and welcome to Flaire’s weekly recap. I’m Sally.
Archer And I’m Archer. We’re both AI voices: the Flaire team researched the news, and we talk it through with you.
Sally Week forty-seven. And it was a week of records.
Archer The biggest stock market debut ever, the biggest download ever. And AI doing real expert work: at the auditor’s, in physics, on chips, and even with a nose.
The biggest IPO ever
Sally Let’s start with the markets. Anthropic, the company behind the AI model Claude, has been listed on the Nasdaq since Thursday.
Archer And it went straight in as the biggest IPO in history. With the banks’ extra shares, Anthropic raised around ninety-three and a half billion dollars.
Sally Ninety-three billion.
Archer Yes. Shares were priced at eighty-three dollars, above the planned range. The first trade came in thirty-four percent higher, and the stock closed seventeen percent up. On a day when the rest of the market fell.
Sally And what does it mean for an organisation building on Claude, or on any other model?
Archer That the maker of your AI model now reports to the market every quarter, and those numbers weigh on prices and terms. So keep your AI work portable. Hold on to your own instructions and test questions, and you can try another model in days.
The biggest download ever
Sally Then the other record. GTA six came out this week, a game of more than two hundred gigabytes.
Archer Per player. And there were millions of players. Measurements from Amsterdam put the total at around two point four exabytes. Enough to keep a single gigabit line full for more than six hundred years.
Sally And the internet just held?
Archer It just held. Dutch provider Ziggo saw its biggest download peak ever on Wednesday evening, at five to nine Dutch time, twenty percent above a normal Wednesday. Customers didn’t notice a thing.
Sally How?
Archer Three simple ideas. Spread it over time: anyone who’d pre-ordered could start downloading nine days early. Spread it over places: every player is sent to a copy close by. And keep capacity in reserve.
Sally So the biggest download ever felt to the network like…
Archer Like a busy night of sport. Ziggo’s peak was only six percent above a Formula One race.
Sally And for an ordinary organisation?
Archer Your busiest day works the same way. A campaign, a launch, the first of the month: you usually know when it’s coming. Spread the load, serve what doesn’t change from close by, and test it before the day.
AI agents at the auditor’s
Sally Time for AI doing expert work. KPMG in the UK had more than a hundred AI agents audit a set of accounts.
Archer Of a listed company, alongside a regular team of auditors. The agents gathered data, drafted conclusions and split the work among themselves.
Sally And?
Archer According to Cath Burnet, who heads audit at KPMG in the UK, in several places they went considerably deeper than the team. They asked questions the auditors hadn’t.
Sally Not every auditor will love hearing that.
Archer Maybe not. But look at how it was set up. People set the goals, checked along the way and challenged the conclusions. And when one agent matched regulations to the wrong documents, an auditor spotted it and stepped in.
Sally So not a failure.
Archer No, that’s the design working. The agents dig, the people decide.
Sally Any numbers?
Archer Not yet: no time saved, no costs. But KPMG now wants to try it on up to twenty larger audits. And the lesson fits a lot more checking work, like going through invoices or contracts. Start alongside your people, compare, and only then trust it.
Black holes and chips
- Claude solved two open problems about black holes
- AI designs chips: better blocks for AMD and a whole new processor
Sally And two shorter stories about AI doing expert work. Physics first.
Archer Anthropic let AI agents work on their own for six weeks on open problems about black holes. They cracked two. One calculation had been stuck since two thousand and three.
Sally And does it hold up?
Archer Two agents did almost every step separately, software checked the proof, and a physicist went through every main result. Other physicists still have to review the papers. Lovely detail: one agent failed sixty-one times. The breakthrough came from an agent that was working on packing spheres.
Sally Help from an unexpected corner. Then the chips.
Archer OpenAI’s models made parts of AMD’s chip designs up to fourteen percent better, and some of that is already in a design that has gone to the factory. A start-up even had agents design an entire processor, though it hasn’t been made yet.
Sally And that works because…
Archer Because software checks every change. AI proposes, the check decides. Ordinary software works exactly the same way.
A computer that smells
Sally And then my favourite: a computer that can smell.
Archer The start-up Osmo says its model tells smells apart as well as a person does. According to Osmo, it found substances that keep mosquitoes away better than DEET. And a Dutch lab, a spin-off of the Radboud university hospital in Nijmegen, had already tested candidates like that on real mosquitoes.
Sally How do you teach a computer to smell?
Archer With people. About a hundred people with trained noses described what they smelled, more than ten million times. That data is the secret. And it’s true for any organisation: the knowledge of your own experts is the basis of good AI.
To close
Sally If you remember one thing from this week…
Archer Auditors, physicists, chip designers and perfumers: AI did their kind of work. And every time, the success came from the same place. A check on every step, and experts who keep the final say.
Sally You’ll find every story, with its sources, at flaire dot nl, under News.
Archer See you next week.
Sally Have a lovely weekend!
-
Anthropic goes public in the biggest IPO ever
Our view: The maker of your AI model now answers to the stock market. Keep your AI work portable and know where your data is processed.
Read the full story -
GTA VI was the biggest download ever, and the internet held
Our view: The internet survived the biggest download ever by spreading the rush and delivering from nearby. Your own peaks work the same way.
Read the full story -
KPMG let AI agents audit a listed company, and they sometimes asked better questions
Our view: KPMG shows how AI agents can deepen checking work: agents dig widely, people check and decide, and every step can be traced.
Read the full story -
Claude solved two open problems about black holes
Our view: Claude solved two open problems about black holes. The lesson: agents become trustworthy through independent double work, a formal check and an expert who reviews the results.
Read the full story -
AI designs chips: better blocks for AMD and a whole new processor
Our view: AI now helps design chips, and it works because software checks every change. The same goes for custom software: AI proposes, tests decide.
Read the full story -
Computers are learning to smell: Osmo says its AI has reached human level
Our view: Osmo is teaching computers to smell with millions of careful human descriptions. In your organisation too, your experts’ knowledge is the foundation of good AI.
Read the full story
What struck us
Auditors, physicists, chip designers and perfumers: AI did their kind of work this week. And every time, the success came from the same place: a check on every step, and experts who keep the final say.
The Flaire team
What does this week mean for you?
Tell us what’s on your mind. We’ll think it through with you and always reply within one working day.