Open the calculator app on your phone and you're holding something like 100 million times the transistors that the best computers had in the early 1970s. That's one of the largest jumps in raw capability any technology has handed an ordinary person, in any span of decades. And yet you are not 100 million times more productive than an economist working in 1972. Two or three times, maybe, on a good day. That gap, and what it implies about what AI can and can't do to the economy, is the subject of a recent essay by Stanford's Chad Jones, written for the Journal of Economic Perspectives and expanded on in a talk he gave this spring alongside MIT's Daron Acemoglu. It's the clearest thing I've read this year on where AI's economic upside actually comes from, and why it might arrive on a very different clock than its risks.
Start with a chart Jones calls his favorite in all of economics. Plot US income per person over the last 150 years on a scale where equal distances mean equal percentage changes, and the line is almost eerily straight: about 2% growth a year, give or take, since 1870. What makes that strange is everything that happened along the way. Electric lighting, the internal combustion engine, antibiotics, the transistor, the personal computer, the internet, every one of them a genuine transformation of daily life. And the growth rate barely moved.
So which future does AI belong to? Jones lays out two deliberately extreme scenarios, since neither is quite right but both teach you something. In the first, AI keeps compounding on itself: it makes software engineers dramatically more productive, then automates AI research itself, then uses that to design better robots, and eventually handles nearly every task a human can do, cognitive or physical. Dario Amodei's phrase for the middle of that story, a "country of geniuses in a data center," is doing a lot of work in Silicon Valley right now, and there's already real evidence behind the early steps. A benchmark that measures how long an AI agent can work on a task unsupervised before it fails has been doubling roughly every six months since 2020; since late 2024's reasoning models arrived, that doubling time has reportedly quickened to under four months.
The second extreme takes that 2% line at face value and says AI is just the next entry on the list: transformative the way electricity was transformative, but not growth-rate-changing. The logic is that within any given technology, the easy ideas get found first, so each field's contribution naturally fades unless something new arrives to take its place. On this account, every "transformative" technology in that graph didn't accelerate growth so much as prevent a slowdown, and each one took decades to fully arrive. Factories had to be physically rebuilt around small electric motors instead of one giant steam-driven shaft before electricity's productivity gains showed up, and, famously, you could see the computer age everywhere except in the productivity statistics for years after computers actually turned up everywhere.
Jones's real contribution is a way of thinking that sits between these two stories, and it's a genuinely useful idea even outside economics: weak links. A chain is only as strong as its weakest link, and an economy, he argues, works the same way. Shipping a new smartphone means design, chip fabrication to impossibly tight tolerances, a global supply chain, timely delivery, marketing, retail, and getting any one of those links wrong sinks the value of the whole chain, the way one twenty-five-dollar rubber O-ring brought down the Challenger. Automating some links beautifully doesn't fix a chain if the rest are still slow.
This is where the phone in your pocket comes back in. Cheap computing solved one link spectacularly well: matrix inversion, once a job for a room of human "computers," is now instant. But deciding what matrix to invert, what question to ask, what to test, that's still a human doing a human's job at close to a human's pace. A hundred million times more of one input, bottlenecked by everything else that hasn't changed, buys you maybe two or three times the output. Jones turns that intuition into an actual formula: economic output behaves like the harmonic mean of the tasks that feed into it, a kind of average that's dominated by its smallest term, which is exactly why making one input infinite still leaves total output finite.
None of this rules out an eventual growth explosion, and Jones is careful to say so. Automating tasks produces new ideas, new ideas enable automating more tasks, and that loop can genuinely run away with itself once nearly everything has been automated. Run a simplified version of the model forward, assuming the automatable share of the economy grows by one percentage point a year, and the shape is unmistakable: for the first 50 to 60 years the growth rate barely moves off its historical 2%, and only in the final quarter of the run does the curve bend sharply upward, just before the math sends it to infinity as full automation is approached. In the fuller, calibrated simulations Jones ran with Acemoglu, even a fairly aggressive setting only gets the US to about 2.3% growth by 2050, up from 2% today, and it can take a century or more before the flywheel fully takes hold.
Two concrete examples anchor how slow "slow" can be. Self-driving cars looked like an easy problem back in 2004: turn the wheel, work the pedals, that's really the whole job. None of the fifteen vehicles in that year's DARPA Grand Challenge finished the course; the best one covered eight of the 142 miles. Twenty-two years later, Waymo has logged over 170 million driverless miles, and it remains, functionally, a San Francisco phenomenon. On the labor side: Geoffrey Hinton predicted in 2016 that medical schools should stop training radiologists, since AI would replace them within five years. There are more radiologists today than there were in 2016, and they're paid more, because reading a scan was never the entire job. Automating three-quarters of a role's tasks can raise the wage of whoever still does the remaining quarter. Automating all of them, closer to what may be coming for narrower jobs like driving for a living, doesn't.
Here's the turn that makes this worth reading past the growth charts. Jones argues weak links cut both ways: they slow the benefits of automation, but they don't slow the risks, they might even speed them up. A chain that takes decades to strengthen link by link can still be destroyed in an instant by damaging just one of them. An AI capable enough to be a genuinely superhuman software engineer doesn't need to automate the whole economy to do serious harm; it only needs to find the one link worth attacking. In the talk, Acemoglu described a next-generation model that reportedly surfaced thousands of previously unknown bugs in decades-old, heavily used software, and asked the obvious next question: what happens once a capability like that reaches anyone with bad intentions, not just the lab that built it. The more speculative version of the worry is stranger: we're growing a new kind of intelligence we don't fully understand, and historically, encounters between more- and less-advanced civilizations haven't gone well for the less advanced side. Berkeley's Stuart Russell put the concern in a single line during a research seminar: "How do we retain power over entities more powerful than us, forever?"
Jones has separately tried to put numbers on how much risk is worth accepting for how much upside, and the results are uncomfortable in an interesting way. Take a simple model of how people value additional income, and give AI a genuinely enormous payoff: 10% annual growth, doubling living standards every seven years, in exchange for a one-time chance of wiping out humanity. Under one common assumption about how much extra income is worth to people who already have a lot, the acceptable risk is startlingly high, better than a one-in-three chance of extinction. Under a more realistic assumption about how fast that value fades, it falls to 2.5%. But add one more realistic ingredient, that an AI capable of that kind of growth would probably also drive major medical breakthroughs, and the calculation flips again: if it cuts mortality in half, people become willing to accept up to a one-in-four chance of extinction, because the tradeoff is no longer consumption against death, it's one cause of death against another.
For a benchmark that's easier to feel in your gut: in 2020, the US effectively spent around 4% of GDP to reduce a roughly 0.3% individual mortality risk from Covid, by shutting down large parts of the economy. Standard government estimates of the value of a statistical life imply Americans would be willing to pay more than the equivalent of their entire per-capita GDP just to cut mortality risk by a single percentage point. Weighed against even a rough estimate of how effective safety research might be, Jones's calculations suggest current AI safety investment is too low by a factor of thirty or more, and that spending on the order of $100 billion a year, about a third of one percent of GDP, would clear a standard cost-benefit test easily.
So where does that leave the two extremes I opened with? Probably both right, at different timescales. The weak-link view says the big economic gains from AI are likely real but genuinely slow, plausibly decades away, the same decades it took factories to fully absorb electricity or offices to fully absorb the internet. It also says the risks don't have to wait for that. A single weak link, hacked, can do damage on a timescale of weeks, long before the flywheel of automation has spun up the rest of the economy. Big gains arriving late, and big risks arriving early: that asymmetry is the single most useful idea in the whole essay, and it argues for spending the slow years actually preparing, for labor-market disruption and for the sharper, nearer-term risks, rather than waiting to find out which extreme turns out to be right. The chain only needs one weak link to fail. It needs nearly all of them fixed to fly.