Solar Flare, Not Tsunami
Takeoff is happening faster than even the bulls think, and most people won't notice it for long enough to matter. That gap is the biggest arbitrage of our careers.
Nov 16, 2026
All year, in every room I gave The New Software Factory talk, I put up this slide and asked people to place a bet.
Blue, orange, or red. Which line are you betting your company on?
Blue was METR's original trend: the length of task an AI can finish on its own doubles every seven months. Orange was the 2024–2025 pace. Red was fit only to the newest models. I'd push the room toward red, then click to the next slide and let it sink in.
Red, on a linear axis.
Doubling every 88.6 days. That would alarm people, and it was supposed to. I was the bull in the room. The company you build for blue looks nothing like the one you build for red, and I wanted them to feel it.
I was wrong. Not about the direction. About the speed.
And I think most people, including most bulls, are about to be wrong in the same way, and then wrong in a second, stranger way: they'll expect to see it coming. They won't. Takeoff is happening, it's faster than you think, and it will go unnoticed by most people for long enough to matter.
That gap between what the models can do and what anyone notices is the biggest arbitrage of our careers. Let me make the case.
Faster than the bulls
Here is my slide again, redrawn by Benjamin Todd with one more line on it.
Teal: 7 months. Gold: 3.3 months. Magenta, since Opus 4.5: 1.7 months.
Since Opus 4.5, the line is doubling every 1.7 months. My "scary" pace worked out to about 17x a year. This one is over 100x a year.
The honest caveat: the magenta diamonds are estimates, not measurements. METR doesn't have direct access to the newest internal models, so Nikola Jurkovic translated Anthropic's internal benchmark scores into time horizons, and Ben converted those to the 80% bar. Nikola calls it "pretty weak evidence." So don't tattoo "38 hours" on your arm.
Paul Graham saw the chart and put the whole question in one line:
If this is real, this is stupendously fast growth.
My answer was short: it's more correct than not. I stand by that. The exact diamonds will move. The shape won't. Because you don't need the diamonds. The scores came from a three-month independent assessment METR ran inside Anthropic, and its plain-language findings are enough on their own. Anthropic's progress is running at about 1.9x its old rate, "roughly two years of historical progress over the last year." They guess ~80% of that speed-up is AI itself, and they're "confident that the acceleration was not due to an acceleration in the scaling of training compute." Most important: "We did not identify any convincing bottlenecks."
The same week, someone who'd just left Anthropic posted the slide that helped convince him to go.
This slide from the Anthropic IPO roadshow was a big factor in my decision to leave. It’s an insane plot and I’m surprised they made it public. It shows the compute efficiency of pretraining over time. The y-axis is a **log scale**: they are seeing continued superexponential Show more
Look at the y-axis. It's a log scale. A straight line there would already be exponential. This one bends up.
Okay, but those are lab-internal numbers about coding and AI research. What about normal work?
Half of remote work
A year ago, Scale AI and the Center for AI Safety launched the Remote Labor Index: 240 real freelance projects, bought off sites like Upwork, each with the brief, the files, and the deliverable a human professional actually shipped. The median job took a person 11.5 hours and cost about $200. The AI gets the same brief and the same files, and judges decide whether a paying client would accept its work.
At launch, the best agent passed 2.5% of them.
Fable 5.2, released a week ago, passes 41%. That's double Opus 5.5 from a month earlier. Ethan Mollick noticed it's not just on the steep trend, it's above it:
Remote Labor Index: 41.3% against 27.1% on trend, about three months ahead of the line.
Do the napkin math with me. If the public model clears 41%, and METR says Anthropic's internal models are ahead of what ships, then it's likely that somewhere today a model clears half of these jobs. Not "can help with." Delivers work a client would pay for.
Yes, there are caveats. These are solo jobs, with no client calls and no live systems, cut off from "any broader business context," as Epoch puts it, and there's "far more remote labor than represented in RLI." Cost per job matters too. But a year ago this was 2.5%. The caveats aren't getting bigger. The number is.
Coding is already solved
Let me say the quiet part from my own work: traditional coding is solved. So is traditional code review.
Not ML research. Not building novel agent loops. Those are still hard. But the coding most companies actually employ people to do? Take a frontier model, put it in a strict enough codebase, give it enough tokens, and run adversarial loops of different models checking each other, and no human is needed in the loop. A human is still needed on the loop. That's the current state of affairs, not a forecast.
If that's true, it should be the biggest story in the economy. Instead, we've seen almost no consequences yet, apart from a lot more weekend side projects. Jevons paradox hasn't shown up. Hiring hasn't snapped. Most engineering orgs still look like they did in 2024.
We've already seen this pattern once. Around December 2025, developer sentiment flipped almost overnight. People came back from the holidays AI-pilled. The story everyone told was that Opus 4.5 changed everything.
I don't buy it. We were there before Opus 4.5. What changed was that most developers finally had two quiet weeks to sit with the tools and get comfortable. The capability had been there for months. The noticing was the bottleneck.
Now run that forward. Today a single shot gets you something like a 30-hour task. At 1.7-month doublings, that's 60 hours by Christmas. Run seven threads in parallel and have an agent do the planning, and you can lay out the equivalent of six months of engineering work in an afternoon.
Then what?
I don't know. Almost nobody does, because almost nobody has tried. That's the point.
Why nobody notices
I wrote about this last year in Seeing UV. To you, a bird's feathers are plain brown. To another bird, they glow with ultraviolet patterns. A dog's nose resolves a world of smell we don't have words for. The difference is real. We just don't have the senses for it.
That's what's happening with these models. Fable 5.2 should be a different animal from Opus 5.5. Ask the people around you whether they've noticed. Most haven't. Same chat box, same answers to the same questions, maybe a little sharper on the hard stuff.
That's because a 30-hour capability can't show itself in a ten-minute conversation. The difference between Opus and Fable isn't in better answers to your questions. It's in work you would never have thought to hand it. And the faster the curve climbs, the further it runs past what a normal person can perceive. We're going to be talking about alien intelligence, and like dog smell, most of us won't be able to describe it.
There's an older version of this story, too.
When electricity arrived in factories in the 1890s, productivity barely moved for roughly thirty years. Factory owners swapped the giant steam engine for a giant electric motor and kept everything else the same: the same drive shafts, the same layout, the same workflow. The gains only came when a new generation rebuilt the factory floor around small motors at every machine. The technology wasn't the bottleneck. Reorganization was.
Our version moves at 1.7-month doublings instead of over decades. The capability is the dynamo. Most companies are bolting it onto the drive shaft.
Solar flare, not tsunami
The usual takeoff story is a tsunami: mass job extinction overnight, everyone hit at once, a wave you can see from the beach.
I think it lands like a solar flare instead. Enormous energy hits the planet and almost nobody feels it. The world carries on as it was. Then strange, isolated things happen: a satellite dies, a grid flickers, a compass points the wrong way.
That's what the next few years look like to me. No single headline. Just a growing number of isolated cases nobody can quite explain:
- The labs, running the loop that builds the next model, faster each turn.
- A handful of companies that absorb each new model the week it ships and turn it into an absurd business. Everyone else chases them and calls it luck.
- Individuals whose output, running many threads of 60-hour agents, looks like a department's.
- The wrong hands. Someone will absorb it fast, and a cyber crew doesn't need anyone's change management. Expect outages that never quite get explained.
Why so narrow? Because capturing this isn't one filter. It's a stack of them.
First, you need a real use for near-limitless intelligence. That's anywhere you used to pay high wages and would have hired more people if you could have afforded them: quant desks, software teams, research labs. Not everyone has that need. Of those who do, most won't see it, for all the reasons above. Of those who see it, not everyone is positioned to capture it. Of those who are, many will get it wrong on the first try. And the few who get it right have to keep absorbing every new release while the curve doubles under them.
Whoever gets through all of that, even if they're "just" building at the application layer, gets their own takeoff. Everyone else sees a world that looks about the same.
That's why it's an arbitrage. The gap between what's possible and what's noticed is wide, and it's widening with every doubling.
So what do you do on Monday?
Here's the uncomfortable part: knowing the exact doubling time doesn't change much. Whether it's 45 days, 60, or 90, you can't use capability you haven't absorbed. Your limit isn't the model. It's how fast your company can take in the change.
From the talk. It matters more now.
So the job is to absorb change as fast as humanly possible. Concretely:
- Move humans from in the loop to on the loop. If a person has to touch every step, your speed is capped at that person's speed. Have them supervise whole jobs, not approve each keystroke.
- Re-test every release on your hardest real work. You can't feel the gains by chatting, so measure them. Keep a set of real tasks the last model failed, and run it the week every new model ships.
- Rebuild, don't bolt on. Don't swap the steam engine for a dynamo and keep the drive shaft. Redesign the workflow around what agents can do now, then do it again in two months.
Past that, I genuinely don't know. Six months of work planned in an afternoon, 60-hour tasks per shot, half of remote jobs deliverable by a model. What does a company look like on the other side of that?
I don't think anyone knows yet. But I'm fairly sure of one thing: there won't be a siren. Nobody will announce that it's here. You'll find out when someone in your market gets suddenly, unreasonably good, and you can't see why.