Nobody Told You What "Use AI" Means

Today on Briefly AI. Mistral unveils a trillion-parameter open model called Le Chonk. OpenAI dumps 722 maths manuscripts on the world in one go. And most workers have been told to use AI without being told what that means.

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.

Somewhere in Paris, a team of engineers named a trillion-parameter model "Le Chonk", and honestly, it's the most relatable thing a frontier lab has done all year.

Right, so the model is Mistral Large 4. Mistral is the French lab, and it released this as a preview. According to Simon Willison's write-up, it has one trillion parameters in total. Parameters are, roughly, the adjustable dials inside a model. But only 49 billion of them are active at any moment. That's called a mixture-of-experts design: the model routes each request to the relevant slice of itself rather than waking up the whole thing. Mistral says it trained it on its own cluster of 3,800 Nvidia Grace Blackwell chips.

The bigger claim comes from Wired. Mistral says this is the best open-weight model outside China. Open-weight means you can download the model and run it yourself, rather than renting it through someone's API. And the company's own phrasing was that it wants to show it's "still in the race" at the frontier. Simon Willison's verdict was blunter: Mistral are back in the game.

Now, the landscape, because a single model on its own tells you very little. On the closed side, OpenAI's GPT-6.1 Sol, not the 5.6 Sol from June, is pitched as near-Astra coding at two dollars per million input tokens, versus ten for GPT-6 Astra. Builders are already stacking them. One recap describes a cheap Sol as the orchestrator, GPT-6 Luna handling narrow tool calls, and Astra brought in only to review the result. On the open side, Nathan Lambert wrote this week that GLM-5.3, the Chinese open release Anthropic warned had Mythos-level cyber risk, hasn't produced major attacks so far. Simon Willison has also been running Qwen 3.8 27B on his own laptop. And Microsoft just shipped three voice and transcription models, MAI-Voice-2.1 among them. So the shape is this. Closed labs are competing on price per task. Open models are mostly coming from China, and Mistral wants to be the exception. Whether Le Chonk lives up to the billing is a question for the benchmarks, and it's only a preview so far.

Now, OpenAI. You may remember the row with the mathematicians last month. Well, it's back, and bigger.

According to The Verge, OpenAI has published solutions to a number of long-standing maths problems, produced by an unreleased frontier model. That's a batch of 722 manuscripts, grouped into 372 result families, meaning clusters of related papers. It extends a run of these releases that has, in The Verge's words, both impressed and unsettled parts of the mathematical community.

Wired adds the reaction, and it's rather less polite. The piece describes OpenAI preparing to release more than 100 new solutions to unsolved problems, and quotes one mathematician saying there's a perception of "mobster behavior" from leading AI companies. The complaint isn't mainly that the maths is wrong. It's about how it's done. Mathematicians are being handed an avalanche of results, produced by a model nobody outside OpenAI can use, and the people who'd normally check, credit and argue over them are left to keep up on their own time.

Here's why that matters beyond maths. Proof is one of the few fields where you can't just say "trust me, the machine got it right." Either a proof holds up under scrutiny or it doesn't. So when 722 manuscripts land at once, the bottleneck stops being whether the machine can produce them. It's whether humans can verify them. A paper nobody has checked is a claim, not a result. And the questions about research ethics The Verge mentions, who gets credit and who does the checking, haven't been answered by anyone yet.

The concrete state of play: the manuscripts are public on GitHub, the model behind them isn't, and the mathematicians are reading through them now.

Right, on a more human note. Forbes ran a piece this week from Joe McKendrick with a title I rather like: "We Keep Missing The Most Important Point About AI Adoption". It draws on a survey by Resume Now, and the numbers are quietly damning.

Forty-two percent of US workers say they lack confidence integrating AI into their workflows. Forty-four percent say they have no clear path to build the skills. And sixty percent say the expectations placed on them are undefined.

Read that last one again. Sixty percent. Most of the workforce has been told, in some form, to use AI, and has not been told what good looks like. Are they meant to draft with it? Check with it? Report that they used it? Nobody said. So you get people either avoiding it, in case they do it wrong, or using it privately and hoping nobody asks.

The point of the Forbes piece is that the missing ingredient isn't the technology, because the tools are there. Access to sanctioned AI tools at work has climbed a long way, with one benchmark putting it at about sixty percent of workers. But only about a quarter of organisations have moved a substantial share of their pilots into production. Plenty of companies have bought the thing. Far fewer have worked out the job.

There's a human layer under it too. A psychologist from the American Psychological Association told the Washington Post that the fear isn't only about losing a job. It's about losing purpose, because people's identity is tied up in the effort they've put into their work. Put that next to an undefined expectation and you can see why people freeze. It's hard to adopt something when you don't know whether you're being told to master it, or being quietly told you're about to be replaced by it.

The practical bit: if you manage people, the cheapest fix on this list isn't a new tool. It's one written paragraph saying what AI is for on your team, what's expected, and what's off limits.

So, that's the day. A French lab named its biggest model after a pun, a machine produced more maths than mathematicians can read, and sixty percent of us are working from a brief that says "do the AI thing." Which, to be fair, is also how the model was probably trained.

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.