I talked to a guy who values watches for a living the other day.

His whole job is figuring out whether a watch is real or fake and what it's worth, and being the annoying tech-obsessed person I am, I asked him if AI couldn't just do that for him now.

He told me AI gets it right 96% of the time.

In my head I was thinking “oh so you’re cooked then...”

Then he said "that's TERRIBLE," and I realized I had it backwards. His entire business runs on accuracy, so if you tell someone their watch is real four times out of a hundred when it isn't, you've just torched your reputation and possibly someone's bank account.

He and his team run at 99%, and they're fighting tooth and nail for every 0.01% past that.

That conversation summed up where AI actually stands in mid-2026 better than any benchmark chart could.

Every new model that's come out lately feels like a new iPhone launch: a small bump that excites the people already paying close attention and changes nothing for the 99% of people just trying to get through their day.

The most common argument that AI is reshaping the job market comes from vibes.

People point to how hard it feels for new grads to find engineering jobs…

I wanted something harder to argue with, so I pulled the actual employment numbers for computer systems design and related services, straight from the FRED data:

That line peaked in March 2023 and has dropped by less than 5% in the 3 years since, even with two years of "AI is coming for your job" headlines sitting on top of it.

Compare that to the dot-com crash, when the same category of jobs fell 18% in 2.5 years.

And remember, coding and software engineering tasks are supposed to be the single thing AI is best at right now.

So that whole "we're handing out free food to the homeless" joke is funnier as a meme than it is really a trend.

Here's where I think this goes from here.

When a company is trying to convince public market investors that AI is the thing that justifies its valuation, every announcement starts doing double duty as a pitch deck slide, and I think we'll see that ramp up rather than cool off.

I got an email from Anthropic last week about age and identity verification, and I can't help but wonder if that's connected to the Fable block and whether Claude is making small concessions to the US government to get back in its good graces.

Either way, the regulatory leash is getting shorter, not longer.

At least for the frontier models, people actually want to use.

That's already pushing more people toward cheaper alternatives like DeepSeek, with Microsoft leading the charge on adopting them.

The "AI gets cheaper every year" story is true at the bottom of the market and increasingly false at the top of it