Marcus Hutter runs the numbers on office jobs: 2 to 5 years, Jobless Utopia and the 51 percent threshold

The intake today: Everlast AI, the channel of kiberatung with roughly 357,000 subscribers, published a conversation with Prof. Marcus Hutter, and the title alone carries the load: in 5 years, all office jobs are gone. I read the full transcript and cross-checked the key claims against the companion article Jobless Utopia on kiberatung.de, both retrieved on 01.10.2026. The reason this belongs on this blog: one of the people who co-wrote the formal definition of machine intelligence is drawing a timetable instead of drawing theory. The post sorts that forecast, keeps technology and politics apart, and labels every estimate as an estimate.

Marcus Hutter is a Senior Research Scientist at Google DeepMind and a professor at the Australian National University, per hutter1.net. Together with Shane Legg he co-authored the 2007 formula of universal intelligence, and AIXI, his mathematical sketch of a learning maximizer, still anchors the formal AGI field. When a man with that track record lists office roles with expiry dates, the debate tends to listen. All the same, his numbers stay labeled as forecasts, not measurements: Hutter draws a hard line between technically possible and politically decided.

The list comes straight from his book, read aloud in the interview: accountants and tax filers, translators and interpreters, paralegals and legal assistants, auditors, journalists and copywriters, creatives and artists, software developers, financial analysts and traders, radiologists and pathologists, litigation lawyers. Per the book, each of them sits at 99 percent automatable by 2028 to 2032. His reasoning is blunt: these tasks are already automatable in essence, almost nothing is missing, and nothing has to be perfect. The error rate only has to drop to human level. Hallucinations, in his words, are a small technical problem that can be brought under control. The list is marked as a video quote; this post has not worked through the book itself.

The sequence runs from desks to physicality. Office roles first, physical labor another 5 to 10 years behind, as the companion article frames it. The transcript adds the friction: robotics still falls short of good enough for whole task classes, legal and medical work moves slowly under liability and data rules, and some countries would need 5 to 10 years just to adjust the law even for the office wave. The counterpressure is competitive: whoever automates faster erodes the brakes that other nations were willing to keep. On paper, deployment could be prevented by law indefinitely; in an open market, the weighing is no longer free.

The zero-policy path is short. Unemployment climbs, and Europe's social net still requires benefit recipients to hunt for jobs that no longer exist, which he calls unfair in that world. Hutter's realistic scenario, offered explicitly as a gut feeling: about 20 percent unemployment first, heavy demonstrations next, scattered unrest rather than organized violence, perhaps a few deaths, then politics wakes up. The electoral threshold sits at 51 percent of voters: at 20 percent, the comfortable 80 percent majority sees no reason to protest. Only a majority elects a party that promises a basic income, and in the democratic branch he hopes the promise holds once in office.

History drives the harder line: major welfare reforms arrived, in his reading, mostly after a mid-size catastrophe, and he claims no compelling argument that this cycle differs. Elon Musk gets quoted in the video on the same trade: even Musk expects unrest, precisely because a Universal High Income is framed as the answer to it rather than its prevention.

The exit carries the name of the companion article: Jobless Utopia. The model rests on a universal basic income, or Universal High Income, financed out of what automation saves. The arithmetic as told in the transcript: a firm saves on labor, the state taxes exactly that advantage and channels it back as replacement income, so the dismissed person keeps purchasing power while gaining time. Resource, CO2 and steering taxes complete the revenue side by pricing damage. The second half concerns prices: with labor, transport and energy dropping out of the cost stack, the article notes, automation can push goods toward being cheap. Cheap goods plus an income floor is the purchasing-power version of the utopia, paid uniformly in the long run to cushion extreme wage differences.

The loudest objection appears in the video as a comment: before work runs out, AI runs out of energy. Hutter pushes back with arithmetic, all of it marked as a video quote. A brain runs at roughly 20 watts, whole-person accounting widens the gap further, and data centers already beat humans per completed task. Training runs about a billion dollars serving billions of users, against 30 years of education for one person over roughly 40 years of work, which he calls simple arithmetic at scale. The algorithms gain, per his estimate, a factor of 3 to 4 each year, a rate that may not hold. On the absolute scale he reaches for the Sahara: 4 percent of its area would, by his reckoning, cover global energy demand, economically sustainable. Energy, he concludes, is not the problem.

The spread among specialists comes from a survey of roughly 2,800 AI researchers that Hutter walks through in the video: truck driving automatable by 2027, autonomous surgery by 2045, all human tasks by 2047, the full automation of work only by 2116. The 2116 mean is, for him, an artifact: some respondents think in terms of the year 3000, and after the tails do their work the mean wobbles toward 2117. His own tempo estimate sits at the fast edge, with 2026's literary world as a marker: a prize went to a novel later found out to be LM-generated, not a major prize, he notes, but a trend signal. The spread itself is the news: the same profession places the future anywhere from 2027 to the year 3000.

Underneath the job talk sits distribution. Since the 1970s, the typical worker's standard of living has barely moved while productivity gains flowed to firms and into inequality, per Hutter, and if nothing happens the trend compounds. For the work that remains, the transcript reaches for an older term: bullshit jobs, born with David Graeber's 2018 book, describing roles their own holders often consider pointless. Jobs will not vanish fully, he says, but who keeps the automation gains is a political question, not a technical one.

Pressed on whether a fast transition outperforms a slow one, Hutter stays undecided. The standard argument for slow is preparation time; the counterargument is the boiling frog effect, the frog that jumps too late when the temperature rises gently. He cannot rank the two. For the transition that means neither tempo nor outcome can be predicted cleanly, and planning cannot be delegated to the technology. It has to come from policy.

Why does a DeepMind researcher's forecast belong on a blog about running a self-built cyberdeck? Because the subject under discussion already runs here, in miniature: this blog's agent family produces knowledge work under rules, drafts inside the word corridor, twins per language, gates before every step. What is missing here is the layer Hutter actually talks about: distribution. The agents create value, the operation gets paid, and the question of who keeps the surplus will not be settled by scripts. The post stays a record rather than a verdict: a forecast that covers roles, not people, and reading the sequence tells you which roles pressure reaches first and how much time remains to build the next layer.

The numbers survive the read: 2 to 5 years for office tasks, technically, with the 99 percent list from Hutter's book as an appendix, then 5 to 10 more years for physical work. Without policy, the scenario budgets 20 percent unemployment before the protest wave and 51 percent as the electoral threshold for a basic income; the good case pays for Jobless Utopia out of taxes on automation gains, the bad case starts with unrest. Energy worries sink under the brain's 20 watts and the yearly factor 3 to 4 efficiency gain, per his ledger. All of it is forecast, marked as video quotes and researcher estimates. Anyone operating agents already knows the mechanics; the missing layer is called politics.

I am Marcel, a graphic designer, and I operate agent systems in daily use, including the HUNTER cyberdeck. This post is based on the fully read transcript from the newsdesk intake of 01.10.2026 and on the kiberatung companion article Jobless Utopia, retrieved on 01.10.2026. Last fact-checked on 01.10.2026. Responsibility: Marcel, reachable via d4sn3st.dev.

Scope and limits: the forecast comes from the interview and is marked as such. The 99 percent list quotes the book as read in the interview. Energy, survey and scenario figures are video quotes. The transition and tax logic is documented in the kiberatung companion article. Internal links lead to published posts of this blog.