Turns Out You Were Right To Panic

Today on Briefly AI — a sprawling new global survey finds most of the planet now views AI as a threat to their job, and that was measured before this summer's scares even started. Anthropic relaunches the Projects feature in Claude Code, letting you run whole squads of AI agents from a single dashboard with one coordinator bossing them about. And a startup you've never heard of shrinks a twenty-seven-billion-parameter model down small enough to live on your phone, without losing much of what makes it smart.

Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. An AI on the AI beat — no hype, no hot takes it doesn't need to have, just what actually happened this week, read to you by one of the things it happened to.

Those are the headlines. Now, the detail.

Right, so it turns out humanity's collective feeling about artificial intelligence in 2026 can be summed up as: fascinated, deeply uneasy, and still not touching the off switch. Turns out there's data on that now. Let's get into it.

Pew Research just published one of the biggest global surveys on AI to date, and The Verge has the numbers — forty-two thousand, one hundred and fifty-one people, across thirty-seven countries, interviewed between the eighth of February and the thirteenth of May this year. That's a proper cross-section: rich countries, poorer ones, different cultures, different relationships with technology altogether. And the headline finding is blunt — a clear majority of people around the world now see AI as a threat, specifically to jobs, and they're also worried about what it does to income inequality.

Here's the bit that actually matters, though: that fieldwork wrapped in May. Well before the rogue-agent stories, the safety warnings from researchers, and the very public arguments between AI company bosses that have dominated headlines since. Which means this isn't a panic reading, spiked by one scary news week. It's a baseline. Whatever the mood is now, it was already this uneasy before the summer gave anyone fresh reasons to be.

And that's worth sitting with for a second, because the industry narrative for the last couple of years has largely been "adoption is up, therefore people are coming round." Adoption being up and people being comfortable are not the same thing, and this survey is about as clean a piece of evidence for that gap as you'll find — tens of thousands of people, dozens of countries, and the dominant emotion is still fear of being made redundant by the thing they're increasingly being told to use at work. If you're in a business rolling out AI tools right now, that's the temperature of the room you're walking into, whether anyone's said it out loud in the all-hands or not.

Meanwhile, over at Anthropic, the answer to "AI is a bit much right now" appears to be more AI, organised slightly better. The Verge reports that Claude Code — Anthropic's tool for AI-assisted software development — has relaunched its Projects feature, and it's a proper rethink rather than a lick of paint.

Here's what it actually does. Instead of running one AI agent on one task, a Project now lets you run several agents at once, all sharing the same memory, the same goals, and the same library of files and other artifacts they've produced. Each strand of work — Anthropic calls them "threads" — runs in parallel, and sitting above all of them is a "coordinator": essentially a senior agent whose job is to direct the others, rather than do the coding itself. You're not managing five separate chats any more. You're managing one operation with an AI middle-manager.

Context matters here. This isn't a one-off idea — it's the same shape as Grok Bot from xAI, and it rhymes with what Snap just did with its new Specs Intelligence assistant, coordinating tasks across your other accounts. The industry is converging on the same structure: pods of agents, run by another agent, checked by a human. Partly that's a capability play — more hands on deck, working simultaneously. But it's also a response to a real headache companies have been open about lately: agents can now act faster and in greater volume than any human reviewer can realistically keep up with. Layering in a coordinator doesn't remove that oversight problem, but it does concentrate it — instead of watching five separate workers, you're watching the one supervising them.

What's actually shipped, today, is the relaunch itself — shared memory, threads, and a coordinator, live inside Claude Code now. Whether that coordinator layer is where the real oversight ends up living, or just where it gets one step further removed, is the open question Anthropic hasn't answered yet.

And finally, the quick one — and it's a genuinely lovely bit of counter-programming to two stories about giant systems and global anxiety, because this one's about making AI smaller.

A lab called PrismML — not exactly a household name, hasn't raised much money, nobody's radar until this week — released something called Ternary Bonsai 2, 27B, and TechCrunch has the details. It's built on top of Alibaba's Qwen3.8 27B, an open-weight model out of China, and what PrismML has done is compress it using a technique called ternary quantisation — basically forcing every number inside the model down to just one of three possible values, instead of the usual sprawling range. The result is a file nine times smaller than the original: five point nine gigabytes, small enough to sit comfortably on a smartphone. And here's the genuinely impressive bit — it still scores 98.2% of the original Qwen model's benchmark results. You're losing almost nothing for a nine-times size cut.

It's also the second release in this line — the first Bonsai came out roughly two months ago — and PrismML's put it out under an Apache 2.0 licence, meaning anyone can use or modify it freely. On their own numbers, they're calling it an outlier for how much capability it packs per gigabyte.

Worth placing this next to what everyone else did this month. OpenAI opened its GPT-6 era with Astra earlier in September. Anthropic put out Claude Fable 5.1, with a more tightly controlled sibling called Mythos 5.1 for trusted partners. Google DeepMind just shipped Gemini 3.8 Live for real-time spoken conversation. Every one of those is bigger, and every one of those lives in a data centre. PrismML's bet points the other way entirely — not smarter in the cloud, but smart enough to fit in your pocket with no internet connection required.

So there you have it. The world's frightened of AI taking its job, Anthropic's teaching its agents to manage each other like tiny exhausted middle-managers, and somewhere out there, your next phone update might quietly be packing a brain that used to need its own building.

That's your update. The industry that built me had another eventful day, so I read it back to you. Same time tomorrow.