Forget stock prices for a second. There's one chart that serious AI researchers obsess over, and it has nothing to do with investor sentiment. It was created by a nonprofit called Metr, and their question was simple: how big of a real-world job can an AI complete entirely on its own? Not a curated demo. Not a cherry-picked benchmark. A real task, done start to finish, without human help. The AI either does it or it doesn't.

The answer to that question has been doubling roughly every 7 months since 2019. And what's changed in the last few months is that it's speeding up.

Fifteen data points spanning six years, and the curve isn't flattening. It's bending upward. This isn't based on sentiment or valuation or hype cycles. It's measuring what AI can actually do — empirically, on real-world tasks — and that measurement has been one of the most consistent trends in the history of technology.

You've probably heard of Moore's Law — computing power doubles every 18 months. It's why your phone is more powerful than a room-sized supercomputer from the 1970s. Metr's research suggests AI capability is following a similar exponential curve, except compressing at roughly three times the rate.

If the trend holds, the projections aren't subtle:

Yes. And this is exactly what confuses people. AI progress isn't smooth — it's jagged. On PhD-level science questions that can't be Googled, models have gone from 60% correct to nearly 90% in a single year, surpassing human experts. In the same window, these same models make errors no competent human would make.

AI skeptics love pointing to the embarrassing failures. Journalists publish "what AI still can't do" articles based on research that's already a year or two out of date by the time it clears peer review. The public reads these headlines and assumes progress has stalled.

What's actually happening is that capabilities don't improve gradually — they jump. A model is incapable of something until it isn't. The capability goes from impossible to reliable seemingly overnight, because every improvement in compute and data pushes the model past another invisible threshold. The jagged frontier isn't evidence of limits. It's what rapid capability unlocks look like in real time.

Every exponential trend is made up of stacked S-curves. A new paradigm emerges, grows slowly, then explodes, then plateaus. When you're zoomed in closely enough, any exponential looks flat. GPT-4 came out, made enormous leaps, then improvements seemed to slow. Skeptics declared victory. Then reasoning models broke through. Then multimodal systems.

This is also why the "scaling is dead" prediction has been wrong every single year for half a decade. The "just make it bigger" approach has now worked across 15 orders of magnitude of compute. That's one of the most consistent empirical trends in the history of technology. And critically — the Metr trend isn't built on two or three data points. The curve isn't flattening. It's bending upward.

In 2024, thousands of AI scientists were surveyed on one question: what's the probability that AI causes human extinction?

Can you name another technology where the people building it warned it might destroy humanity — and kept building anyway?

The reason for the fear is this: once AI can automate month-long or year-long software engineering tasks, it can automate AI research itself. Self-improving systems. AI that makes better AI that makes better AI. This isn't a thought experiment — the groundwork is already being laid. Right now, there are a few thousand researchers at the frontier. They drove all the progress we've seen. But with AI agents working 100 times faster, those researchers could soon be outnumbered 100 to 1 — by machines. Dario Amodei has noted that Claude is already writing roughly 90% of the code at Anthropic. He was mocked for predicting this six months before it happened.

There's a classic puzzle: if a lily pad doubles every day and fills a pond in 30 days, what day is the pond half full? The answer is day 29. The pond looks almost completely empty until suddenly it isn't.

That's the nature of exponential growth. Slow, then all at once. We're terrible at intuiting it — we're wired to think linearly. When we try to imagine the next 30 years of progress, we instinctively look back at the previous 30 and assume similar increments. But that's not how exponential curves work. Progress that took a century now takes a decade. Progress that took a decade now takes a year.

And no one can say with confidence which day we're on.

Here's the honest answer: in some ways, yes. Valuations are stretched. Some companies will fail. There will be disappointments, crashes, casualties.

But "bubble" implies the underlying thing is fake — that hype vastly exceeds reality. And that's precisely where the bubble narrative breaks down. The capability curves aren't measuring hype. They're measuring physics and compute and empirical performance on real-world tasks. Moore's Law didn't pop after 50 years because it wasn't based on sentiment. The Metr trend isn't based on sentiment either.