AI Isn't Overhyped
But the way it reaches our lives, and what it does to work, may well surprise us.
AI is real, but it's coming in like a tide, wave after wave, rather than all at once.
In brief- About half the time, the best AI can finish on its own a job that would take an expert 27 hours, and that length doubles roughly every three and a half months.
- Yet it doesn't show in the job numbers, because most people still don't really use it.
- In the most exposed jobs, employment of 22- to 25-year-olds is already 19% lower than it should be, and voluntary exits leave empty seats.
- If companies stop hiring juniors, where will their seniors come from in ten years?
Half of all entry-level office jobs could disappear within five years.
That was the alarming headline one morning. The warning came from the head of Anthropic, one of the leading AI companies.1 The next morning another headline said the opposite: researchers at Yale still can’t see AI in the job numbers.4 Inside most companies, it feels like nothing has changed. The same meetings, the same colleagues.
And yet a question keeps nagging at you: is your job at risk? Which of those two headlines should you believe?
I don’t think either one is right. AI is real, and it will change the way you work. But it won’t happen on one big day, and it won’t look the way most people expect. Here is what I see coming.
Two worlds, one office
I see it in my own work, in finance and tech. Things that used to take me months, and a team, I can now build alone in weeks.
Most people around me haven’t seen that yet. For them, AI is still a clever chatbot.
That’s the strange thing about this moment. I spend my days in two worlds at once. In one, the machine is doing a team’s work. In the other, it’s a slightly better search engine.
The machine is becoming autonomous
Think of AI as a new colleague who gets better month after month.
In 2023, you could hand it a job that takes a few minutes, like drafting an email.
A research group called METR measures its progress in a simple way. It gives the AI a real task, mostly in software and research, the kind an expert would spend hours on, and lets it work alone. No one corrects it, and at the end someone checks whether the job got done.
By November, METR’s old tests were no longer enough: the best models were passing them. On a new, harder set, they now finish on their own, about half the time, tasks that would take an expert some 27 hours. That length doubles roughly every three and a half months.2
It is already doing things no human has managed. In September, an OpenAI model put forward a solution to the Navier-Stokes problem, about how fluids flow. It is one of the seven “Millennium Prize Problems,” each worth a million dollars, and the institute that offers the prize says it has “apparently” been settled. So far, mathematicians have learned almost nothing from it.3
The machine is producing knowledge faster than the experts can understand it.
Anyone who still thinks AI is overhyped has most likely never tried to build something with it. That’s when it blows you away.
Not in the numbers yet
Now look at the job numbers. Nearly four years after ChatGPT, Yale’s Budget Lab still can’t find AI in the US employment data.4
Andrej Karpathy, one of the people who helped build this technology, gave the best explanation I’ve read.5 About six billion people have barely touched AI. Another one or two billion use a free chatbot now and then. Only around twenty million professionals use it for complex work. Those people, he says, feel “the vertigo of the curve going vertical.”
Most of the world hasn’t started yet, and that’s why the numbers haven’t moved.
We’ve been here before. Electricity took decades to show up in the productivity figures, because factories first had to be reorganized around it. The technology arrives in a few years. The way we work, and the way we think, take much longer to change.
A tide, not a tsunami
Everyone is waiting for a tsunami. What’s coming is a tide, rising slowly.
There won’t be one big day to point at. Look at how it’s already moving.
First, the door closes before anyone walks in. Researchers at Stanford found that employment of 22- to 25-year-olds in the jobs most exposed to AI is 19% lower than it should be, mostly because fewer of them are being hired.6
For many, the problem won’t be losing a job. It will be finding one.
Then come the restructurings. In October, Microsoft offered buyouts to around 60,000 of its employees. In April, for the first time in its history, it had offered early retirement.7 Neither announcement mentions AI. But I expect packages like these to be one of the ways companies cut jobs and replace them with AI. They cost less than layoffs and don’t make the front page. Someone leaves, and the seat stays empty.
Each of these plans is a wave, and a big one for the people it carries off. Still, these waves hit one company at a time, months apart. Wave after wave, the tide rises.
Big tech, top consulting firms and banks will feel this first and hardest. Their work is digital, easy to check and very well paid, which is exactly where AI is strongest. Small firms and family businesses, which make up most of the economy, will change at their own pace.
The question nobody answers
If companies stop hiring juniors, where will their seniors come from in ten years?
I don’t have a full answer. But I can see one new job taking shape, which I’d call the AI manager: the person who runs the machines, knows where they fail, and represents the team to the client. It takes judgment and people skills more than code. If I were 23 today, that’s what I would train for.
For anyone who manages a team, every hire becomes a real choice between a junior and an AI. A team that picks the AI every time will save money this year, and have no one left to promote in five.
The same song, faster
None of this is new. The steam engine, electricity, the internet: each one reached the lab long before it reached our lives, and hit some people years before others.
It’s the same song. Only faster.
The technology changes at the speed of computing power. People and companies change at the speed of people. That gap is where the trouble will be, and most of the opportunity too.
So if you’re uneasy, you’re right to be: you’ve simply seen the tide before most people.
There’s a bigger question underneath all this, and it’s where the next piece begins. If AI slowly eats into our wages, who will still have the money to buy anything?
Sources
- Axios, Jim VandeHei and Mike Allen, interview with Anthropic CEO Dario Amodei, 28 May 2025 (Axios’s wording: AI “could wipe out half of all entry-level white-collar jobs” and push unemployment to 10–20% in one to five years) — https://www.axios.com/2025/05/28/ai-jobs-white-collar-unemployment-anthropic
- METR, time horizons, Claude Mythos Preview (early), 50% horizon about 17.4 hours (80%: about 3.1 hours), release 7 Apr 2026, added 8 May 2026; “measurements above 16 hrs are unreliable with our current task suite” — https://metr.org/time-horizons/ · raw data https://metr.org/assets/benchmark_results_1_1.yaml · trend since 2023: METR, “Time horizon 1.1”, 29 Jan 2026 — https://metr.org/blog/2026-01-29-time-horizon-1-1/ · newer, non-robust results: METR, GPT-5.6 Sol evaluation, 26 Jun 2026 (11 to 270+ hours depending on how cheating is scored; “we do not consider any of these numbers to represent a robust measurement”) — https://metr.org/blog/2026-06-26-gpt-5-6-sol/ · harder new suite: METR, “Q4 2026 capability update”, 20 Nov 2026 (TH1.1 saturated; new suite TH2; “Current public frontier models have an estimated time horizon of up to 27 hours on the verifiable subset of our TH2 tasks”; “time horizons are doubling roughly every 3.6 months” on TH1.1 + TH2) — https://metr.org/blog/2026-11-20-q4-2026-capability-update/
- NPR, “AI solved one of math’s hardest problems. Humanity learned nothing (so far)”, 22 Sep 2026 — https://www.npr.org/2026/09/22/nx-s1-5968588/openai-navier-stokes-problem-mathematicians-learn-little · Clay Mathematics Institute, “Navier-Stokes announcement”, 11 Sep 2026 (“has apparently been settled”; prize not awarded, review “deliberately unhurried”) — https://www.claymath.org/news/navier-stokes-announcement/ · OpenAI, “Navier-Stokes solution” (company claim, Lean-verified) — https://openai.com/index/navier-stokes-solution/
- Yale Budget Lab, “Tracking the impact of AI on the labor market”, update 14 Oct 2026 — https://budgetlab.yale.edu/research/tracking-impact-ai-labor-market
- Andrej Karpathy on X, 11 Oct 2026 — https://x.com/karpathy/status/2109361546505966046
- Brynjolfsson, Chandar, Chen, “Canaries in the Coal Mine”, Stanford, rev. Aug 2026 — https://digitaleconomy.stanford.edu/news/canariesaug26/
- AP, Microsoft offers buyouts to about 60,000 employees, Oct 2026 — https://apnews.com/article/c5a639f53615be198db1799adf6837e1 · Microsoft, 26 Oct 2026 — https://blogs.microsoft.com/blog/2026/10/26/introducing-a-voluntary-exit-offer/ · GeekWire, Microsoft voluntary retirement program, Apr 2026 — https://www.geekwire.com/2026/microsoft-will-offer-voluntary-retirement-to-thousands-of-employees-in-a-first-for-tech-giant/
Who Will Buy?
AI can do the work, but it won't reduce inequality. And its rapid rise is fueled by debt.
Read on →