There's a familiar pattern playing out this week. Someone writes something bold about AI. Half the internet shares it breathlessly. The other half dismantles it point by point. Nobody changes their mind.

Except this time the numbers are hard to ignore.

Matt Shumer's "Something Big Is Happening" — an open letter to his non-tech friends and family about where AI is right now — hit forty million views on X. Fortune picked it up. Inc. picked it up. Every AI newsletter in existence picked it up. Shumer's core claim: AI has crossed a line, the tools can now do things that would've seemed absurd a year ago, and most people haven't noticed. He compared the moment to early COVID — the "this seems overblown" phase of something that's about to change everything.

And the counter-arguments were actually good too.

Will Manidis wrote "Tool Shaped Objects" — probably the sharpest response I've read. His argument: the essay itself was a demonstration of the problem. Forty million people consumed and shared a piece about AI that was, in his view, made by AI. The sharing was the product. He compared the whole thing to FarmVille. Harsh, but not wrong.

Gary Marcus did what Gary Marcus does — catalogued every citation Shumer omitted, every benchmark he cherry-picked, every inconvenient study he ignored. The METR benchmark Shumer cited? It uses a 50% accuracy threshold. And it only measures coding. You wouldn't know that from the essay.

Ethan Mollick noted that AI remains "quite jagged." John Coogan reframed the whole thing: AI isn't COVID. It's not one exponential. It's a series of S-curves — each one real, but none of them the overnight catastrophe Shumer implies.

All fair points. All well-made.

And all missing the thing that actually matters.

Forget the projections

Here's what I keep coming back to. Both camps — the evangelists and the sceptics — are arguing about the future. Shumer says 50% of white-collar jobs will be automated in one to five years. Marcus says the models aren't reliable enough to justify that claim. One side says the sky is falling. The other says it's fine.

Nobody's looking out the window.

Because what's outside — right now, today, this week — is already pretty extraordinary.

I use AI every day. Not as a toy. Not to generate LinkedIn posts. I use it to draft strategy documents, debug code, build working software, analyse data sets, and handle the kind of operational sludge that used to eat half my week. It's not perfect. Sometimes it hallucinates. Sometimes it confidently gives me something wrong. Marcus is right about that — the reliability isn't there yet for everything.

But it doesn't need to be perfect. It needs to be useful. And it is.

That's the part that keeps getting lost in the argument.

The thought experiment nobody's running

Imagine every AI lab shut down tomorrow. No new models. No improvements. Everything freezes exactly where it is right now.

Would the tools we already have still be worth using?

Obviously. Overwhelmingly.

Claude, ChatGPT, Copilot — as they exist today — can make most knowledge workers meaningfully faster at a significant chunk of what they do. Not at everything. Not flawlessly. But at enough to matter.

If this is all we got for the next two or three years, the world would still be better for it. Not because AI is magic. Because it's useful. Daily, tangibly, boringly useful.

That's not a prediction. That's a Tuesday.

The compounding problem

Here's the part that's easy to miss. While the pundits argue about timelines and benchmarks, some people are just getting on with it. Using the tools. Building muscle memory. Figuring out where AI helps and where it gets in the way.

They're not waiting for the perfect model or the mature version. They're developing intuition — the kind you only get from doing the thing.

And every week they do it, they get a little better. Their prompts improve. Their workflows adapt. They learn the failure modes. It compounds.

The people waiting for the debate to settle? They'll be starting from zero. And the gap won't be theoretical — it'll show up in the meeting where someone says "I did this in an hour instead of three days" and everyone else is still doing it the old way.

So what do I actually think?

Shumer's right that something significant is happening. He's wrong about the COVID analogy — Coogan's S-curves framing is closer to reality. Marcus is right that the essay oversells things. He's wrong to imply the tools aren't useful yet. Mollick's "jagged" is probably the most honest single word anyone's used.

But none of them are saying the most practical thing: you don't need to know where AI is going to benefit from where it already is.

So what do you actually do with this?

Read Shumer's essay if you haven't — it's worth your time regardless of whether you agree with every claim. Read the counter-arguments too. Marcus and Manidis both make sharp points.

But then close the tabs and open the tools.

Sign up for Claude or ChatGPT if you haven't already. Not the free version — the paid one. Twenty quid a month. Take something you're working on this week — a report, a proposal, a data set, a piece of code, a strategy document — and work through it with AI. Not as a replacement for thinking, but as a partner in it.

Don't worry about whether AI will be conscious by 2030 or whether it'll replace your job by 2028. Worry about whether you're using what's available right now.

Because something big has already happened. You just might not have noticed.