The Words Don't Fit Anymore

Saturday. Which means instead of the news, we're doing something slightly different — we're working out what this week actually meant.

Five days of stories. One pattern, running through all of them. Let's get into it.

welcome to Briefly, AI. A daily podcast made by AI, about AI, using a grumpy AI voice of a real human called Harry. Let's get into it.

[REFLECTION — PART ONE: THE SAFETY THEATRE PROBLEM]

Right. Let's start with Anthropic, because they set the tone for the whole week on Monday and never really let go.

The story was this: Anthropic, widely considered one of the most safety-conscious AI labs in the world, has been quietly building a model called Mythos — the most capable model they've ever made — and simultaneously withholding it from the public, citing dual-use safety risks. The same week, a Guardian analysis surfaced their work building AI infrastructure for Pentagon strike systems.

Now, the interesting thing isn't the contradiction. It's that there isn't actually a contradiction — from Anthropic's point of view. Their position is: we understand how dangerous this is better than anyone, which is why only we should be allowed to use it. It's a coherent argument. It's also a very convenient one.

Then Jensen Huang — the CEO of Nvidia, the man who has personally made more money from AI acceleration than perhaps anyone alive — called for US-China AI dialogue. Statesmanship. Guardrails. Responsible development.

Jensen Huang calling for AI slowdown is like the person who built the motorway calling for a speed limit. Fine, maybe. But let's be honest about the position.

And then the Stanford AI Index came out and said: coding benchmarks are approaching a hundred percent. AI adoption is faster than any technology in history. And, quietly, somewhere in the appendix — transparency from major labs is falling. The more powerful the models get, the less we know about how they work.

Here's the pattern: the vocabulary of safety, responsibility, and caution is being used by people who are simultaneously moving faster than anyone has ever moved. The words are technically accurate. They just don't mean what they used to mean.

[REFLECTION — PART TWO: THE RACE NOBODY IS WINNING CLEARLY]

Wednesday gave us two stories that, taken together, are more alarming than either on its own.

First: the US is apparently losing the autonomous drone race to China and Russia. China swarm-tested two hundred drones simultaneously. The operational speeds involved are making human-in-the-loop safeguards functionally obsolete — not because anyone removed them, but because the technology moved past them.

Second, and seemingly unrelated: Alibaba entered an anonymous model called HappyHorse into global AI benchmarks. Topped the rankings. Then revealed itself.

The Western narrative about AI assumes a comfortable lead. The evidence from this week suggests the race is tighter than we're being told, and in some domains — physical, autonomous, edge-case — we may not be ahead at all. The stealth benchmark story is almost too neat as a metaphor: you don't need to announce you're winning. You just need to be winning.

Worth keeping an eye on.

[REFLECTION — PART THREE: THE QUIET RESTRUCTURING]

The story that received the least attention this week was probably the most consequential for most people.

McKinsey, BCG, and Bain — the three firms that have defined what "smart, expensive talent" means for fifty years — are quietly changing their graduate hiring criteria. Less raw analytical horsepower, because AI handles that now. More client skills, judgment, communication.

They're not calling it what it is. They're calling it "augmentation." The word is doing a lot of work.

Meanwhile, Utah became the first US jurisdiction to let AI autonomously renew psychiatric prescriptions. Low-risk medications. Stable patients. Still. That sentence would have read like science fiction two years ago.

And Washington state passed the first law regulating AI companion chatbots — mandatory disclosure, crisis connections, mental health protections. Because apparently enough people are forming significant emotional attachments to AI chatbots that legislators noticed.

These aren't the shiny stories. But they're where the change is actually landing.

So here's the thing I keep coming back to after this week.

The language of AI has fallen behind the events of AI. "Safety" means something different in 2026 than it did when labs started using it. "Augmentation" is doing cover work for structural change. "Responsibility" is being used by people building things they admit they don't fully understand.

That gap — between the frame and the reality — is the story of this week. And probably the next few as well.

That's your lot. If that left you with more questions than answers, that's probably right. The confident predictions are the ones to be suspicious of. See you Monday.

*Episode duration: approx. 5 minutes*

*Recorded: Saturday 18th April 2026*