Two AI realities - The painful path to a glorious future

I had two browser tabs open last Tuesday. In one, Demis Hassabis described a future maybe 10 or 15 years out where the major diseases are curable, fusion works, and we start looking at the galaxy seriously. In the other, Geoffrey Hinton explained that the same technology will permanently displace a very large number of workers, and that nobody has a real plan for what happens to them. Both men helped build modern AI. Both have Nobel Prizes. And their two futures are difficult to hold in your head at the same time. I spent a weekend reading everything these people have said in the past year. The headlines have it wrong. The two camps are not arguing about whether AI is powerful - they agree on that. They are not really arguing about the destination either. They are arguing about a stretch of time in the middle that neither side is sure about.

The optimistic view

Elon Musk, Sam Altman, Jensen Huang, Demis Hassabis. Their claim fits in one sentence: once intelligence and energy both get very cheap, most forms of scarcity ease, and work stops being something you must do to eat.

AI and robots will be able to do everything, resulting in universal high income, and work will be optional. — Elon Musk

False, there will be revolution first. — Dr. Michael Burry, the investor from The Big Short

Hassabis is the more careful version and, to me, the more persuasive one. He talks about radical abundance, and he is not speculating from the sidelines. His protein-folding work won a Nobel and is now pointed at drug discovery at Isomorphic Labs. When he says a ten-year drug pipeline could compress toward weeks, he is extrapolating from a machine he already built that already worked. Most abundance talk is a forecast. His is partly a report. He puts the scale at ten times the Industrial Revolution, ten times faster.

Huang makes the narrowest and most testable claim: AI removes tasks, not jobs. A job is a bundle of tasks organized around a purpose. Strip out some tasks and the purpose survives.

You will not lose your job to AI. You will lose it to somebody who uses AI. — Jensen Huang

Altman adds the point that stayed with me longest, because it is not an economic argument at all. He separates two things this debate keeps mashing together: having to work, and wanting to. Then he runs the test backwards. Show our jobs to someone from a thousand years ago, someone whose entire life was food, shelter, survival, and they would not recognize any of it as real work. They would see people with plenty of food and plenty of shelter playing an elaborate game to entertain themselves. He is right, and it is uncomfortable. By that standard my job is a game. Yours probably is too. I sit in a room and rearrange abstractions.

So he points the telescope forward and hopes the people a hundred years out are doing things that strike us as trifling - because that would mean exactly what our jobs would mean to a medieval farmer. The survival problem got solved and everyone moved on. Which quietly inverts the scoreboard: a future full of work we would recognize as necessary is the bad outcome, not the good one.

The transition, he says, will be messy, and in places pretty painful.

The cautious view

Hinton's concern is not philosophical. It is a math not adding up argument, and it is the strongest thing said by anyone in this debate.

The 4 largest hyperscalers have guided to roughly $600–715 billion of capital expenditure in 2026, up from around $400 billion in 2025 - capital intensity of 45–57% of revenue, a ratio you see at utilities and steel mills, not software firms. Hinton asks the only question that matters about a number that size:

Where does the return come from?

Infrastructure at that scale, depreciating over five or six years, has to throw off something like a third of its cost annually just to justify itself. A few hundred billion dollars a year in new revenue, every year, growing. The global software market is not that big. Advertising is not that big. There is exactly one pool of money on earth deep enough, and in the US alone it runs to roughly $12 trillion a year. It is called wages. That is why he says these companies are, whatever their public framing, betting on AI replacing a lot of workers. The bull case for the capex and the bear case for employment are the same case seen from different ends. You cannot fund one without the other.

Yuval Noah Harari raised the adjacent problem years earlier. His concern was never unemployment as such - it was the arrival of a group who are not exploited but simply unnecessary. When production is automated, who owns the means of production gets replaced by who owns the algorithms. Income can be transferred. Standing is harder. Which puts him and Altman in a more interesting fight than either would accept: same subject, meaning rather than money, opposite read on human nature. Altman thinks people invent new games worth playing. Harari thinks a lot of people get handed a check and told the game is over.

Dario Amodei is the complicated one. In May 2025 he told Axios that AI could eliminate half of all entry-level white-collar jobs within one to five years - the most falsifiable claim anyone in a leadership position had made. Since then his framing has moved. And in July, Anthropic's own head of economics published an essay arguing AI has caused no material rise in US unemployment. A chief economist, publicly contradicting his own CEO. Altman softened over the same months, saying he does not expect a jobs apocalypse and that he is delighted to have been wrong.

I do not think that is dishonesty. It is what happens when a warning becomes a sales objection. Nobody buys the tool that ate the buyer's team.

What happened to the Radiologists

Almost everything here is a forecast. There is one closed test case which I find very intriguing. In 2016 Hinton said it was completely obvious that within 5 years deep learning would read medical images better than radiologists, and that people should stop training radiologists now. A decade later, radiology is one of the hardest specialties in American medicine to staff, pay has climbed past half a million, and vacancy rates are at record highs. The AI part came true - nearly every scan now passes through an algorithm. The capability arrived on schedule and the employment effect ran backwards, because radiologists consult, decide, intervene and carry liability. Image-reading was the compressible slice.

But there is a second-order effect nobody uses. Applications to radiology fell after the prediction. Students heard a Nobel laureate say the field was over and went elsewhere, and the residency pipeline takes years to respond. Some meaningful share of today's shortage was caused by the forecast. Predictions about labor markets are not neutral observations. They are inputs. Every man quoted in this article is moving the thing he is describing - steering capital, steering hiring plans, steering what an eighteen-year-old decides to study. That makes the confident voices on both sides less trustworthy, not more.

Which is roughly where the current data leaves you too. The Dallas Fed found unemployment among workers aged 20–30 in AI-exposed occupations up nearly 3 percentage points since early 2025, while Yale's Budget Lab found no meaningful aggregate change. The damage is real, concentrated at the entry level, and landing on people who are twenty-three. That is exactly what year two of Amodei's scenario looks like. It is also exactly what a normal, survivable adjustment looks like.

The good ending and the brutal middle

Strip out the theater and almost everyone agrees the far end looks strange and materially richer than today. Even Harari assumes enormous productive capacity — he just doubts it gets shared. The argument is about the middle.

Economic history has a name for this phenomena. Between roughly 1780 and 1840, British output per worker rose substantially while real wages barely moved. Robert Allen called it Engels' Pause. The Industrial Revolution ended extraordinarily well - antibiotics, weekends, life expectancy. It also spent two or three generations getting there, and if you were inside that pause, the fact that your great-grandchildren would be richer did nothing for you. The good ending and the brutal middle are both true, and they belong to different people. And if Hassabis is right that this runs ten times faster, the compression presumably applies to the pause as well.

That is what "pretty painful" is standing in for, and it is doing enormous work in that sentence. Painful for whom, for how long, and who is accountable? The abundance case treats the middle as friction on the way somewhere good. The cautious case treats the middle as the actual event.

🫤 Dileep's Skeptical Takeaway

I am watching two things. Whether AI revenue starts arriving from genuinely new activity rather than headcount savings. And whether the bottom rung of white-collar careers which are clearly in deep trouble right now - just ask the graduating class of 2026 - finds meaningful employment in some new form. No amount of abundance at the top 1% fixes a generation that never got a first job. There may be a glorious outcome for humanity in 10 years, but I believe that means the next 10 years will be quite painful for those without a lifetime of savings, a well paying secure job and assets.

Sources:

• Musk on X, July 2026 — AI and robots do everything, resulting in universal high income, work optional
• Musk on X, April 2026 — earlier universal high income post
• Gulf News — US–Saudi Investment Forum remarks on work becoming optional and money becoming irrelevant
• Fortune, June 2026 — trillionaire status alongside the money-is-irrelevant claim
• Business Insider, July 2026 — the four-word reply to Musk
• Hassabis, Substack, July 2026 — his own framing, in his own words
• Fortune / Yahoo, February 2026 — new golden era of discovery
• The Guardian, August 2025 — ten times the Industrial Revolution, ten times faster
• The medieval-farmer passage and the "pretty painful" line come from a thread by @realBigBrainAI collecting Altman's remarks
• Reuters, May 2026 — no jobs apocalypse, Commonwealth Bank of Australia event
• Business Insider, May 2026 — delighted to have been wrong
• BGR — you will lose your job to somebody who uses AI
• Fortune, July 2026 — tasks, not jobs
• Yahoo — dismissing Amodei's entry-level jobs prediction
• Fortune, December 2025 — Georgetown discussion with Sen. Bernie Sanders
• Yahoo News — betting on AI replacing a lot of workers
• Ideas.TED.com — the rise of the useless class, excerpted from Homo Deus
• TheStreet — the original May 2025 Axios warning, revisited against 2026 data
• Fortune, May 2026 — Jevons Paradox onstage with Jamie Dimon
• Peter McCrory on X, July 2026 — Anthropic's head of economics, in full
• Yahoo Finance — coverage of the internal disagreement
• The New Republic — a radiologist on the prediction that was supposed to end his career
• Fortune, May 2026 — a decade on, pay and demand both up
• Joshua Gans — separating Hinton's prediction from his prescription
• Marit Health — how the forecast helped cause the shortage
• Futurum Group — 2026 hyperscaler capex estimates
• OfficeChai — the $715 billion figure and 2025 comparison
• Introl — Goldman's $1.15 trillion 2025–27 projection and capital intensity ratios
• Dallas Fed, January 2026 — young workers in AI-exposed occupations
• Yale Budget Lab — ongoing AI labor market tracking
• Cryptopolitan — NY Fed graduate unemployment, PwC Global AI Jobs Barometer, Uber support cuts