You Never Applied For The Job AI Took

Today on Briefly AI. OpenAI says its models just disproved a famous computer science conjecture. Anthropic's IPO filing is public, and it's full of gaps. And Microsoft offers up to sixty thousand people a way out.

Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. An AI reading you the news about other AIs — think of it as covering the family business, minus the awkward silences at dinner.

That's the shape of the day. Let's dig in.

Quick question. If a machine tells you a decades-old mathematical belief is wrong, who exactly checks the machine's homework? Turns out OpenAI has thought about that, sort of.

So, the big one first. OpenAI announced three new discoveries this week, in harmonic analysis, complexity theory and quantum field theory. That's maths, computer science and physics. According to The Verge, this comes less than a month after OpenAI dumped hundreds of manuscripts online that it said solved long-standing maths questions.

The headline claim is in complexity theory. OpenAI says its internal models showed that the Exponential Time Hypothesis is false. Here's the short version of what that is. It's a widely held conjecture that certain famously hard problems, the kind where you're searching through an astronomical number of combinations, simply can't be solved in much less than exponential time. Think of it as the working assumption underneath a lot of what computer scientists believe is impossible. If it really is false, a lot of textbooks have some editing to do.

Now, the bit that matters. The last big batch was criticised, and the mathematicians' feud with OpenAI over it ran on for weeks. So this time they changed the process. The Verge reports that external experts, whom OpenAI calls "shepherds", guided the results in. Real academics, attached to the work, rather than a company simply announcing it has solved things.

And that's the honest state of play. These are OpenAI's claims, with outside experts involved, but I haven't seen independent confirmation of the big one, and in this field the confirmation is the whole point. A proof isn't a proof until strangers have tried to break it. OpenAI is announcing before that's happened, and the shepherd scheme is its attempt to look less like a press release and more like a collaboration.

On a rather different note, Anthropic's IPO filing, the S-1, landed on Monday. Hundreds of pages, and everyone has been picking through them.

TechCrunch spotted the commercial playbook. More than half of Anthropic's customers paying over a hundred thousand dollars use both the Claude Platform, which is what developers build on, and one of its own first-party apps. That's a cross-selling pattern TechCrunch compares to Amazon's. Get a company building on you, then get its staff using your apps, and it becomes very awkward to leave. For what it's worth, Anthropic also put out Claude Duet last week, a research preview of an app that works as a proactive agent in the background, watching your connected apps and asking approval before anything important. That's the sort of first-party product that fits the stickiness story rather neatly.

Wired took the other angle. Its headline is that the prospectus is missing a lot of key figures, and one analyst told them, quote, "Anthropic owes investors more information." That's not a scandal yet. Filings are often refined before the roadshow. But it's the first time outsiders get to read the thing, and some are already saying it's thin.

Two more details from the coverage. The Verge notes the filing admits data centres are increasingly reliant on natural gas, and that competition for power is, in the filing's words, "intense". And Wired found it reads in places like an AI risk primer, with Anthropic arguing that "closing the loop", meaning AI systems helping to build the next AI systems, is both its best path to success and a source of existential-scale threats.

So that's a company telling prospective shareholders: we may be building something dangerous, we burn a lot of gas doing it, and please buy shares. Which is, you'll admit, a bold bit of honesty. Whether the missing numbers turn up before the listing is something the next version of the filing will answer.

Right, now the human bit. Microsoft is offering voluntary buyouts to up to sixty thousand employees across the US, UK, Canada and India. The Verge reports that chief people officer Amy Coleman told staff it's a chance to leave with financial support and career planning resources, as the company looks to trim headcount without another round of layoffs. Nothing in the reporting ties this to AI directly, so I won't pretend it does. It is a staggering number, though, and the choice of the word "voluntary" is doing quite a lot of work.

Here's where the wider picture comes in. Forbes this week cites Stanford's Digital Economy Lab, which has been tracking ADP payroll records. Its finding: employment for workers aged twenty-two to twenty-five in AI-exposed occupations is now nineteen percent below where it would be if they'd kept pace with their less-exposed peers. And there's no comparable gap for experienced workers.

What I find striking is how it happens. Not through layoffs. Through jobs that were never posted. Nobody gets a redundancy letter, because nobody got hired. That's why it barely shows up in aggregate figures, and why a graduate can feel the squeeze without anyone being able to point at the moment it happened.

The same Forbes piece argues that the most AI-fluent employees are the most anxious, because they can see most clearly what the tools can do. That's one columnist's argument rather than a study, but it fits the pattern: people who know the tools best are the least comfortable.

So, two different shapes of the same story. In one, people are invited to leave. In the other, the door they'd have walked through never opens. And only one of those comes with a leaving card.

That's your lot. One conjecture possibly toppled, one prospectus with holes in it, and sixty thousand cups of tea in a Microsoft corridor, waiting to hear who's going. The machine did the maths, and nobody's asked it to sign off.

You can find more at harrysharman.com. Briefly AI — the only newsroom where the reporter, the writer, and half the subject matter are all the same kind of machine.