Based on Dr. Roman Yampolskiy's conversation on The Diary of a CEO with Steven Bartlett. Yampolskiy is a computer scientist who has worked on AI safety for fifteen years — he coined the term — and his forecasts are among the most extreme in the field. Treat what follows as one expert's stark, contested prediction, not settled fact. But it's worth sitting with, because his logic is uncomfortably clean.

Ask most people which jobs will survive artificial intelligence and you'll get a comforting list — the trades, the caring professions, anything "creative," anything human. Ask Roman Yampolskiy, who has spent fifteen years on AI safety and coined the term, and the comfort evaporates. His answer, on The Diary of a CEO, is that the list is almost empty — and that the few jobs left standing survive for a reason nobody wants to hear.

A caveat before the bad news: Yampolskiy sits at the far, darkest end of the forecasting spectrum. His timelines are aggressive and hotly disputed, and plenty of serious people think he's too pessimistic. But his argument is clean enough that it's worth following all the way down, because the escape hatch it removes is one most of us are quietly counting on.

His timeline: cognitive work by 2027, physical work by 2030

Yampolskiy's near-term map is specific. By around 2027, he expects artificial general intelligence — a "drop-in employee" offering cognitive labor for the price of a $20 subscription, at which point, he says, "it makes no sense to hire humans for most jobs." Anything done on a computer goes first.

Then the part that closes the last door. By 2030, he expects humanoid robots — Tesla's and others', improving "at light speed" — to have enough dexterity to do physical work too. Plumbing, he notes pointedly, "felt like the last bastion of human employment." In his 2030, the artificial plumber exists. Intelligence plus a capable body, always connected to the network, leaves — his words — "not much" for human beings to do. He's not describing 10% unemployment, the kind that terrifies economists. He's describing 99%.

Why "learn to code" was the last bad advice

Every prior wave of automation came with a reassurance: this job is going away, so retrain for that one. Yampolskiy's core point is that this time the reassurance is a category error. Fire was a tool. The wheel was a tool. Every invention in history was a tool for a human to wield. But intelligence itself is the "meta-invention" — you're not building a better tool, you're building a new worker, "the last invention we ever have to make." And once the worker is general, there is no adjacent job to flee to, because the machine can be pointed at that one too.

He watched it happen in real time with his own field. First: artists can't make money, learn to code. Then: AI can code, become a prompt engineer — get a four-year degree in it. Then: AI writes better prompts than humans, so that's gone too; now the "safe" job is designing AI agents, which he guarantees is gone within a year or two. Each supposedly future-proof role had a shelf life measured in months. "There is no plan B," he says. "You cannot retrain."

The only jobs left — and the uncomfortable reason they survive

So what does remain? Here's the twist that reframes the whole "which jobs are safe" genre. According to Yampolskiy, the surviving jobs are not the ones AI can't do. By his account there are almost none of those. They're the ones we'll irrationally insist a human do anyway. Five kinds of work fit that description — and notice that every one is a preference, not a protection:

1. The human kept for tradition. His example is Warren Buffett, who would never switch to an AI accountant — not because the AI is worse, but because he's Warren Buffett and he likes his human accountant. A real market exists here, but it's the market of people wealthy or old-fashioned enough to pay extra for the old way.

2. The handmade premium. We already pay more for "man-made in the US" over "mass-produced in China," Yampolskiy notes — and he's blunt that there's "no practical reason for it. It's almost a fetish." The artisan survives as a luxury good, not a livelihood for millions.

3. The human you want because they're human. Care, comfort, companionship, counsel — work where the whole point is that a person is present. The machine can perform the service; some of us will still pay for the humanity behind it.

4. The authentic performer. We may keep watching humans play sport, make music, and tell their own stories for the same reason we still run races cars could win — because the human doing it is the point.

5. The face to hold accountable. A human kept in the loop to answer for a decision, to be the name on the outcome, to give a person something to trust or blame. Useful, but a sliver of the jobs it replaces.

Add them up and you don't get an economy. You get a boutique — a "tiny subset of the market," in his phrase. These aren't five safe harbors. They're five reminders that the only thing standing between a job and automation, in his world, is a choice we make about what we value.

The part he's oddly calm about

The strangest thing about the interview is Yampolskiy's serenity. He sleeps well, he says, deploying the same psychological machinery that lets all of us function while knowing we and everyone we love will die: we don't dwell on catastrophes we can't personally prevent. He rejects the easy exits — you can't "just unplug" a distributed system smarter than you that has already made backups and predicted your move; you can't retrain your way around a general worker; enhancing our own biology won't keep silicon-speed intelligence in a fair fight. His faint hope is that the people racing to build it might change course if they truly believed the downside applied to them too — that no amount of money is useful to a dead man — and choose narrow, problem-specific AI (cure a disease, make billions, stay in control) over general superintelligence.

You don't have to buy his 99% or his 2030 to take the useful thing from this. Strip away the doom and a practical instruction remains: stop asking which job title is safe, because that framing assumes a moat that may not exist. The real questions are the ones he ends on — not "what will I retrain as," but "what do we want to keep human, and why." Those are decisions about values, and unlike the technology, they're still ours to make.

Long talk short: Yampolskiy's real message isn't a list of safe jobs — it's that the usual escape hatch is gone. For every past wave of automation the advice was 'retrain for the next job.' But if the machine is a general worker, there is no next job to retrain into. The only work that survives is work we choose to keep human — and that's a decision about values, not a moat technology can't cross. Whether or not his timeline is right, that's the question worth preparing for now.

Further reading

• AI: Unexplainable, Unpredictable, Uncontrollable — Roman Yampolskiy
• Superintelligence — Nick Bostrom
• A World Without Work — Daniel Susskind

Affiliate links — a small commission supports the blog at no cost to you.