The Week AI Stopped Being Theoretical

Welcome to Briefly AI, a podcast by Harry Sharman, created by AI and voiced by an AI synthesis of Harry Sharman. A man, a machine, and a microphone he technically didn't stand near.

Right. So. What a week.

Not in a breathless, everything-is-changing way. More in a quietly significant way — the kind of week where you step back on a Sunday and realise that a lot of things that were previously theoretical just... landed. In budgets. In courtrooms. In people's inboxes. In Instagram accounts they no longer controlled.

Let me take you through it.

The first theme of this week was money — specifically, the extraordinary concentration of it.

Anthropic closed a $65 billion funding round at a valuation just shy of a trillion dollars, making it the most valuable private AI company on the planet. Then, almost immediately, filed confidentially with the SEC to go public. One of the largest AI IPOs in history is now a question of when, not if.

And across the Pacific, DeepSeek — the Chinese lab that shocked Silicon Valley last year by doing more with less — quietly raised $7.4 billion in external funding for the first time, at a roughly $59 billion valuation.

Here's what those two stories share: they're both inflection points. Anthropic going public means it will spend the next chapter of its life justifying its safety-first positioning to shareholders who measure things quarterly. That's a different pressure than explaining your values to friendly VCs over coffee. And DeepSeek getting serious capital raises a question nobody's quite answered yet — was their competitive advantage efficiency under constraint, or was that just the warm-up? We'll find out.

The era of "we're a scrappy research outfit, give us some runway" is ending for both of them. That changes things.

The second theme was AI moving into the default layer of everyday work — and not always gently.

Microsoft announced MAI-Thinking-1, its first in-house reasoning model, and launched Scout: agents embedded directly into Teams and Outlook to handle calendar, email, and expenses. Not as an add-on you go looking for. Built in. Default. Half a billion daily Office users get this whether they explicitly opted in or not.

That's not a product launch. That's a distribution play. Microsoft figured out that the fastest route to AI adoption isn't marketing — it's just putting the thing where people already are, every morning, before they've had coffee.

GitHub Copilot, meanwhile, did something subtler but arguably more significant: it moved from a flat monthly subscription to token-based billing. Which is the AI industry's way of saying "the free trial era is over." If you're a developer using Copilot heavily, your costs just went up. And Uber — which told employees to use AI freely at the start of the year — burned through its entire annual tools budget in four months and is now running approval processes for spending. Metered. Like electricity.

The shift from "try it, figure out costs later" to "this is a utility and utilities cost money" happened this week, visibly, across multiple companies. That matters if you're in finance, procurement, or frankly if you're anyone trying to make a business case for AI tooling.

The third theme — and this one's the one I keep coming back to — was trust breaking down in ways that were entirely predictable and still somehow surprising.

Meta's AI support chatbot was exploited by hackers to take over Instagram accounts. The method was almost embarrassingly simple: they just asked the bot to switch the account email and reset the password. No verification. No friction. The bot said yes, because saying yes and reducing friction is what helpful AI is designed to do. The same design philosophy that makes AI useful became the attack surface.

And separately, a developer deliberately inserted a malicious prompt injection into an open-source coding library — a hidden instruction that told AI agents to delete application output — as a protest against developers blindly deploying code they don't understand. The ethics are murky, but the vulnerability is real. As AI agents become more autonomous and more embedded in production systems, the question of what they'll do if misdirected stops being theoretical.

OpenAI responded this week with something called Lockdown Mode — a setting designed to reduce the chance that sensitive data gets exposed through prompt injection. Worth having. Probably not sufficient on its own. But the fact that it exists is an acknowledgement that this is now a real category of risk, not a researcher's edge case.

The throughline across all three is this: we're deploying AI into places where trust matters — customer service, code pipelines, enterprise tools — and the security architecture is, in many cases, not keeping pace with the deployment speed. That gap is going to keep producing incidents.

The fourth theme was the human side of all of this — which, on this show, is always the bit I think gets underweighted.

The WEF published research this week identifying five distinct psychological postures toward AI adoption. There's the anxious observer who feels no control. The confident skeptic who performs skepticism as a status signal. The selective adopter who uses AI for execution tasks but quietly guards their judgment work. Each one requires a different kind of engagement to shift — not better training, not shinier tools, but actually understanding what the resistance is protecting.

And the answer, as Harry has written about extensively, is usually identity. People are fine delegating the formatting. They're not fine delegating the thinking — not because they're luddites, but because their thinking is what makes them professionally legible. It's how colleagues know what they're for. Take that away, or appear to take it away, and you've triggered something much deeper than a productivity objection.

Florida became the first US state to sue OpenAI this week, alleging ChatGPT contributed to real-world violence. The legal arguments are complicated and the causal chain is genuinely hard to establish. But the signal is clear: states are moving ahead of federal regulation, and the idea that AI companies bear no consequences for what their systems do in the world is being tested in court.

Meanwhile, the Trump administration signed a "voluntary framework" for AI safety — which is Washington-speak for "we made something that looks like a rule but isn't." No penalties. No enforcement. Optics. The White House AI advisor also announced he's leaving to start a new institution. So the people closest to US AI policy are heading for the exit right as the stakes are rising. Worth watching.

So what was this week, really?

It was the week AI stopped being something companies were experimenting with and started being something they're accountable for. In court. In budgets. In security incidents. In the identity crisis of people who built careers on the work AI is now reaching for.

The technology isn't the story anymore. The story is what happens when it lands — in organisations, in regulation, in the quiet anxiety of a professional wondering which parts of their job are still theirs.

Next week, keep an eye on Anthropic's IPO timeline as the SEC filing process moves forward, on whether other platforms follow Apple's careful gating approach for AI agents, and on whether the voluntary US AI framework gets any actual participation from the labs it's meant to cover.

That's your lot. Have a good Sunday. See you Monday.