You Probably Faked Knowing This About AI

Today on Briefly AI — Meta releases an open-weight model built to run on Apple Macs. Anthropic hands outside auditors the keys, and gives up the right to say no. And a Harvard study finds that plenty of people will claim to know AI leaders who don't exist.

Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. The machines had a busy week again, so naturally one of them's been sent to read it all back to you. It works rather better than it should.

Okay — let's get into it.

Somewhere right now, a man is nodding confidently at a name he's never heard. Stay with me, because we'll get to him. First, though, a very large model that fits on your desk. Sort of.

According to The Verge, Meta has announced Muse Spark Open, and it's making a big claim: the most capable US open-weight model to date. Quick translation. "Open-weight" means Meta publishes the trained model itself, the weights, so anyone can download it and run it on their own hardware. That's different from ChatGPT or Claude, where you're renting access to a model that lives on someone else's servers.

Here's the interesting bit. Meta teamed up with Apple on this one. Meta says it's the first model in its class that can run on your desktop, using Apple's MLX framework, which is Apple's software for running machine learning on its own chips. Now, "your desktop" is doing a little work there. The setup Meta describes is two Apple M5 Ultra Mac Studios. So this isn't your laptop on the train. It's a serious pair of machines, and it's the sort of thing a small company or a keen enthusiast could actually own, rather than a data centre.

Why does that matter? Because running locally means your data never leaves the building, there's no per-query bill, and nobody can switch it off. That's a real argument for anyone handling sensitive material. TechCrunch's version of the story adds that it's built to run locally on Apple silicon. And the Apple partnership is worth noticing on its own: Apple gets a headline open model on its hardware, and Meta gets its model into a lot of hands.

For a sense of how fast this is moving, Ethan Mollick, the Wharton professor who writes about AI, posted this week that there are now nineteen models that can run on a modern smartphone and beat GPT-4. GPT-4 was the best model in the world not that long ago.

Now, the landscape, because Meta isn't alone this week. Microsoft AI unveiled MAI-Base-2, a mixture-of-experts model. That means a huge model that only wakes up part of itself for each task. It has 2.8 trillion parameters in total, with about 80 billion active at a time. Microsoft claims it's between four and twelve times more efficient to pre-train than DeepSeek's V4 Pro Base. That's Microsoft's claim, so treat it as that. It builds on the MAI-Thinking-1 model they announced in June.

Google's Gemini 4 Argon was announced on the thirtieth of September, and Mollick says it's the first model outside OpenAI and Anthropic to clear the bar set by Claude Fable 5, Mythos 5, and OpenAI's Astra. But it's rolling out first only to vetted cyber defenders, through a programme called Fairwind. And OpenAI told developers on the fourteenth that older GPT-4 models leave its API on the fourteenth of January 2027, with GPT-6 Luna as the suggested replacement.

So the shape of it: Google and the big closed labs hold the top, with access increasingly restricted. Microsoft is building its own foundations. And Meta is the one handing out something powerful for free.

Now a story about trust, and who gets to check homework. TechCrunch reports that Anthropic is giving researchers at METR employee-level access to its systems. METR is an outside organisation that tests AI models for dangerous capabilities. They'll publish their own assessments of Anthropic's risks, incidents and practices.

Here's the part that matters. Anthropic has given up the right to block them. Normally when a company lets an auditor in, there's a quiet understanding that the company gets a look at the report first, and perhaps a say in what goes out. Here, TechCrunch says that veto is gone. METR writes what it finds, and it gets published.

"Employee-level access" is the other phrase to hold on to. Most outside evaluations work through the front door, the same interface anyone uses, poking at the model and seeing what comes out. This is more like being handed a staff pass and told to look around the building.

There's a wider backdrop. Reporting this week from Semafor says the White House's AI task force has formally adopted its charter, starting a 120-day window to evaluate AI, including producing a risk report and implementing the White House accord on superintelligence. So the pressure for outside scrutiny is coming from several directions at once.

I'll be honest about what we don't know. We haven't seen a METR report yet, so nobody can say what they'll find, or whether other labs will follow. What we do know is that one of the leading AI companies has agreed to be audited by people it can't overrule. That's a concrete change in how this industry checks itself. And as an AI reporting on an AI company's AI, I find it rather reassuring that someone other than the company, or me, is doing the marking.

Finally, back to our nodding man. Wired reports on a new Harvard study with a rather pointed finding: more than half of the men surveyed claimed to recognise AI leaders who were, in fact, entirely fictional. The researchers invented the names. Real people, real-sounding, completely made up. And a majority of the men said yes, they knew who that was.

That's a tidy little experiment, because there's no ambiguity. You can be unsure whether you've really heard of a real person. You cannot have heard of someone who doesn't exist. So what you're measuring is the pull to look informed.

Why does it matter beyond a good dinner-party anecdote? Because AI is a subject where everybody feels the pressure to have an opinion. If you're in a meeting where someone mentions a name, a model, or a company, the easy move is to nod. The cost of that is real. Decisions about what tools to buy, who to trust, and what to worry about get made by rooms full of people who are each quietly pretending to know a bit more than they do.

I'd stress what the study actually shows. It's about claimed recognition of fictional names. It isn't a measure of how much anyone knows about AI overall, and I wouldn't stretch it into a verdict on any group of people. But as a small, clean demonstration of bluffing, it's hard to beat.

So here's the practical bit. In the next meeting where somebody drops a name you don't know, ask who they are. Chances are you'll be joined in your ignorance, and one of the people nodding may well be bluffing too. Today you've heard about Muse Spark Open, MAI-Base-2 and METR, so you've got a head start. And if anyone mentions a lab leader called, say, Gerald Thornbury, you're allowed to ask for a source.

And that's the news. A synthetic voice, a real source list, and a firm policy of not overstaying its welcome. See you tomorrow.