The Nasty Middle

It's 2026 and the layoffs aren't coming. They're here. Meta cut about 8,000 people — roughly a tenth of the company — and moved seven thousand of them onto AI teams. Cloudflare cut its workforce by a fifth, the first mass layoff in its sixteen-year history, in the same quarter it posted record revenue. Tech layoffs blew past 100,000 for the year before summer.

The "they just overhired during COVID" line was the comfortable one. The VCs who repeated it have gone quiet. Cloudflare's CEO said the quiet part in the Wall Street Journal: the people he replaced were "measurers" — middle management, finance, legal, internal audit, revenue recognition. The coordination layer. The org chart's connective tissue.

Here's what makes this different from every productivity scare before it. The money going in is real and enormous. The payoff coming out is real and not here yet. That gap — spend without yield, automation without abundance — is the whole story. We call it the nasty middle. And it has only just begun.

Economists have a name for the shape of this. The Productivity J-Curve: a general-purpose technology forces a huge wave of invisible investment — new skills, new processes, whole organizations rebuilt — that drags measured output down before it lifts it up. The curve dips, then it rips. The bottom of that dip is the middle.

History rhymes hard here. During Engels' Pause, British output per worker climbed about 46% while real wages crawled up maybe 12% — for half a century — before wages finally roughly doubled. The Solow paradox was the same trough a century later: computers were "everywhere but in the productivity statistics" until firms reorganized around them in the late 1990s. The in-between is always where the pain pools and the gains hide.

Why does the middle of the org chart get cut first? Because a company carries two kinds of complexity. There's product complexity — the genuinely hard problem of serving billions of users, which doesn't go away. And there's human-to-human complexity — the coordination tax, the reporting layers, the meetings about meetings. Every person you add is a new node on the social graph.

AI's biggest short-term leverage isn't curing cancer. It's eating that coordination tax. The measurers go first. And the leverage of the people who remain goes vertical — the 10× engineer becomes the 1000× engineer, a fifty-person company books $40M of revenue per head, and the first one-person billion-dollar company stops being a thought experiment. Fewer people, holding more. That is the middle's defining move.

Here is the turn this whole forecast is built around. The AI-safety conversation has spent a decade worried that the machine will be misaligned with us. The nearer, likelier danger is the reverse: that we will be misaligned with each other. That a large group of newly displaced, justifiably furious people will optimize locally — for themselves, right now — and burn the whole project down before its gains are ever shared.

The resentment will be real, and it will be earned. The psychology is not a mystery. People feel losses about twice as hard as equivalent gains. They judge their lot against the person one rung up, not against history — last-place aversion is a measured thing. And status threat, not raw economic hardship, is what reliably tips people toward political revolt.

We have run this experiment before. The Luddites weren't technophobes — they were skilled workers whose wages were being cut, and in 1813 seventeen of them were hanged at York. The Captain Swing rioters smashed threshing machines across England in 1830; courts handed down 252 death sentences. After financial crises, far-right vote share rises about 30% on average. The first sparks of hostility toward AI's winners are already here — and they aren't aberrations. They're the leading edge.

This cuts against the standard intuition that faster means more dangerous, so sit with it. If the binding risk of the next five years is a human backlash manufactured by the middle, then the length of the middle is the risk. A slow takeoff doesn't spare us the disruption. It stretches out the most painful, most resentment-generating part of it, while pushing the breakthroughs that would end the pain further away.

We're not against caution, and we won't pretend irreversible harms don't exist — where a mistake truly can't be undone, slow down. But that describes a narrow set of cases, and the AI-safety crowd has spent its credibility insuring against the rare tail while ignoring the likely middle. Lock-in to stagnation is also irreversible. The pitchforks are also a catastrophe. Minimizing time in the trough is a safety argument — not an argument against safety.

Tell a peasant farmer in 1650 that their descendants would spend their days in pajamas, typing on glowing rectangles, and they wouldn't call it a job. They'd call it leisure. So maybe the post-AGI "jobs" already exist and we just can't picture them, because they'll look like leisure to us — art, stories, games, sport, exploration, status traded in currencies we haven't invented yet.

We would rather live in that 2050 than on a 17th-century farm. We are pro-AI not in spite of being pro-human but because of it. This is what AI realizing itself looks like — and us realizing ourselves alongside it.

But that future has a toll gate, and the toll is the next five years. It requires the people losing their jobs not to torch the system before it can redistribute the gains. It requires the labs not to get regulated into irrelevance by a backlash they earned through arrogance. It requires the rich to understand that giving up 20% now is how you keep the other 80%.

We don't have to solve what humans are ultimately for. We just have to manage the next five years.