This week I attended Exponential Day which was organised by the lovely team at ACCIO Catalunya, in the World Trace of Barcelona. It was a full house for economist Daniel Susskind’s opening keynote, as he spent a good 40 minutes chipping away at everyone’s comfort zone: that the white-collar and knowledge-work jobs built on judgement, creativity and empathy are safe from artificial intelligence. Susskind, a research professor in economics at King’s College London, author of A World Without Work, and a member of the UK government’s Expert Panel on AI and the Future of Work, has spent the better part of a decade studying the matter, and definitely thinks no one should rest assured.
He started by taking us back a couple of centuries, when a third of Americans worked in agriculture; today about one percent do, while US farms produce far more than they used to. Machines tend to take over routine work first, which is why blue-collar jobs have long been considered the ones most exposed to technological disruption. White-collar work was supposed to be safer, protected by the fact that it is hard to write into code: it involves judgement, a feel for other people, knowing what to do when no rule applies.
The plot twist is that generative AI turned out to be actually quite capable of these non-routine tasks that were supposed to be out of technology’s reach. AI’s range is wide, how far it will go is still genuinely unclear, and while judgement is needed most when we face uncertainty; if AI does one thing well, it is reducing uncertainty, which reduces the need for judgement that goes with it.
Susskind’s most useful tip is that we need to change our metrics. A job is a bundle of tasks; some are routine, some are not; and automation replaces one task at a time, not one job at a time. Counting whole jobs at risk misses what actually happens: the tasks inside a job change, even when the job itself evolves or survives.
I did not come to Exponential Day for reassurance, and I definitely did not get any. The question Susskind closed with is something I have been losing sleep over at home; we keep asking the same thing, and really liking the answer: is the educational system preparing our kids for the job market, whatever that phrase will even mean in a couple of years.
This issue of The Compute is my attempt to address this question (even if partially) through the one part of it we can actually measure: the work itself.
Help wanted, generalist abstain
The demand for specialists is climbing, as recruiters have dropped the generalist search and are looking instead for people who can build, wire together and stress-test AI systems (turning a model into a working product), or who pair those skills with deep know-how in a regulated field such as health, finance or security. The talent gap is clear: a senior AI role now takes three to four months to fill, while a generic software role takes about 25 days. For AI engineers, demand runs at more than three times supply. There are no companies pulling back jobs, headcounts are approved, but the people to fill them are not available, so AI projects run late for lack of talent, not budgets.
Two things follow. First, pay for these specialists is set to rise, because naturally, when far fewer people can do the work, employers have to offer more to win them. Second, the bottom of the ladder is shrinking at the same time. New graduates are now only about seven per cent of hiring at large technology firms, so fewer juniors are coming into the workforce to be trained, and the talent pool that could grow into senior roles later is narrowing.
Recalculating
Here is something I learned a couple of years ago that I find very useful in evaluating technologies and products: a point solution makes an existing professional better at the job, while a systems solution redefines the job itself. Take GPS. It was built for the military, became a device on the car dashboard and then a map application in every pocket. That practically negated a London cabbie’s need for the “Knowledge” (the two to four years of study once needed to memorise the capital’s 25,000 streets), but it was only the point solution; the cabbie still drove, the job just got easier. Uber was the systems solution, and the real disruption, it allowed almost anyone to become a taxi driver, and the structure of the trade changed rather than the skill of its workforce.
Software development is at the same intersection. Claude Code, Co-Pilot and Codex can finish a developer’s line of code; this is the point solution, and it works. Vibe coding is the systems version: you describe the software you want in plain English and the AI builds it. Lovable, a Swedish company whose product does exactly that, reached about $400m in annual recurring revenue in roughly 18 months, largely by letting people who cannot code develop working websites.
The same transformation is arriving sector by sector, and it reaches the formal professions first. Contracts get drafted, slide decks built, spreadsheets modelled and campaigns produced by people who would not have done any of that a couple of years ago, instead resorting to lawyers, consultants, analysts and marketeers. For any business that charged for something because it was hard to make: difficulty and complexity were the moat, and that moat is dissolving. It has started reaching the creative industries too, where AI tools can now write and perform a passable song, or develop a social media reel, from a single prompt. The one thing none of them brings to the table is taste, the know-how and intuition to decide which of a dozen (or thousand) generated options is worth keeping. That is the work that does not automate, and with any luck it stays that way for a long while. These tools can make anyone a producer. They just cannot make everyone worth listening to.
Course correction
As tasks move rather than whole jobs, Susskind corrects the forecast from mass unemployment to mass redeployment, and that correction has to run through what schools and universities teach. The scale of the current spending shows how deeply convincing this shift is. In 2025 the largest technology firms spent more on artificial intelligence than the United States spent, in today’s money, on the entire Apollo programme over roughly a decade, and the 2026 figure is expected to roughly double. Capital on that scale is a bet that machines will do even more of the work, and the labour side of the equation has to keep up.
Susskind’s prescription is about learning: what we teach should divide between skills that compete with the machines and skills that use them; how we teach should stop resembling a classroom from 200 years ago; and when we teach should run across a lifetime rather than just the start of one. Behind the reform is a strong possibility he does not dismiss: artificial general intelligence (AGI, the point at which AI outperforms people at most economically valuable work), something that all three frontier labs expect within five to 10 years. Susskind does not call AGI inevitable, but he does think too few people take it seriously. His next book, out this autumn, puts it in a parent’s voice: What Should My Children Do?
I started Wednesday with a version of that question already in mind: my own kids are being prepared for a job market that is rearranging itself while they are still in school, and neither their teachers nor I can say with confidence what it will look like. The only place to start is not the job they might apply for, but the tasks that will still be theirs to do.
Note to readers
A note before we close.
Follow the Token returns in its 60-minute form on Thursday July 16, at 17:00 Gulf Standard Time, organised and hosted by our good friends and extended family at TiE Dubai and open to entrepreneurs, leaders and investors.
Registration is open: https://events.tie.org/Dubai/follow-the-token
Wirelessly yours,
Ziad Matar
Co-founder, Tokenando
Editor-in-chief, The Compute
After Hours
The World Cup is on, but the tournament is not only between national teams. When football stops, good television has been running its own competition. We have been watching Legends, a six-episode series inspired by the undercover UK Customs officers that Margaret Thatcher had unleashed to stop the heroin trade in 1990, with Steve Coogan acting well outside his Alan Partridge repertoire. Next on the watchlist are the new seasons of Silo and The Bear.
The soundtrack this week started on Saturday night, way before the news broke, with an unplanned late-night dive into Bonnie Tyler’s “Total Eclipse of the Heart”, which stayed there all week. Tyler died on Wednesday night in Portugal, aged 75. She was still recording last year, collaborating with David Guetta on “Together” with that raspy voice entirely intact. What began as a playlist, turned without warning, into a farewell. Rest in peace, bright eyes.
References
Susskind, government role: UK Government Expert Panel on AI and the Future of Work — https://www.gresham.ac.uk/about-us/news/professor-daniel-susskind-advises-governments-new-ai-and-future-work-unit
Susskind, biography and books (A World Without Work; What Should My Children Do?, forthcoming autumn 2026) — https://www.danielsusskind.com/
US agricultural share of the workforce (majority in the early 1800s to under two per cent today) — https://agecon.unl.edu/structure-us-agriculture/ ; https://www.gilderlehrman.org/history-resources/teacher-resources/statistics-trends-american-farming
London taxi “Knowledge” (25,000 streets, six-mile radius of Charing Cross, two to four years) — Transport for London: https://content.tfl.gov.uk/introduction-to-knowledge-booklet-march-2022.pdf
2025 AI infrastructure spending versus the Apollo programme, and 2026 near-doubling — Stanford AI Index 2025; TrendForce; Futurum. [Canonical links to finalise]
Labour market: specialist demand, time-to-fill, graduate hiring share — Bloomberg (July 9), with SignalHire, Randstad and iCIMS. [Primary links in the July 9 deep-dive bundle]
Lovable, about $400m ARR in roughly 18 months — Sifted (July 8). [Primary link in the July 8 deep-dive bundle]
Brilliant! Love reading
you!