The Two-Thirds Quote

The Vanishing Invoice

Mara: A freelance translator, writing anonymously to the newsletter Blood in the Machine, put it in five words: "The work has dried up." For six months they'd watched translators they knew on LinkedIn say the same thing. Another translator in that same collection went from fifty or sixty hours a week to essentially zero. Not one request in June 2025. Then a regular client had a contract. The translator quoted about two-thirds of their usual price. The client said that was still too expensive.

Eli: That client was making a bet: that a machine could do enough of the job to make a human's price look too high. Nobody yet knows whether tax rules and safety nets can adapt before bets like that add up across a whole economy.

Mara: And that's the backdrop for the question a listener sent in: what happens when AI becomes better than humans at almost everything? Today's angle is the trap of premature automation.

Eli: Start with the machines, because they're strange. Researchers at METR describe AI capabilities as jagged: uneven against humans, with the length of work they can handle varying enormously from one task to the next. Then there's a real-world check. Scale AI built the Remote Labor Index, a set of actual freelance projects that take days. As of March 2026, the best system tested, Claude Opus 4.6 Cowork, matched or beat the human standard on about four percent of them. What that index really captures is still argued over. But four percent is the score.

Mara: So not a rising tide. Spikes of brilliance, with gaps in between.

Eli: And the gaps show up at work. Anthropic estimated that, in theory, AI could touch 94 percent of computer and math tasks. When it looked at how people actually used Claude, it was about a third.

Mara: And yet the layoffs came anyway. More than 100,000 tech workers lost their jobs in 2025. AI was named a main driver in over half of those cases. Then, in February 2026, the payments company Block cut nearly 4,000 people. That was almost half its workforce. Jack Dorsey said AI had made many of those roles unnecessary, and he went further: "Within the next year, the majority of companies will reach the same conclusion."

Eli: Four percent of long projects done to standard. Half a company gone.

Mara: That's the translator's two-thirds quote, at company scale. Translators describe clients demanding steep discounts and switching to machine output that still needs a human to fix it. Eli, companies are cutting people for software that, on the long jobs, mostly can't do the work yet.

The Broken Rung

Eli: The job numbers ought to be on fire, then. They're not. Stanford's Digital Economy Lab, going through 2026 data, put it flatly: "We do not see widespread, economy-wide job displacement associated with AI." The most and least exposed workers look broadly alike. Except for one group. A Stanford preprint, a study released before peer review, found workers aged 22 to 25 in the most exposed jobs had seen a 16 percent relative drop in employment since ChatGPT arrived. By June 2026, that shortfall against less-exposed peers had reached 19 percent. Mostly through hiring that never happened.

Mara: So the young aren't being pushed out. They're just never let in.

Eli: There's an old precedent people reach for. In the late 1800s, shoemaking moved from hand to machine. By hand, a medium-grade laced shoe took 72 separate tasks. By machine, 173, with the work cut so fine each worker did about one. Some hand tasks vanished. New ones appeared: keeping steam engines running, supervising workers, packing goods for distant markets. And Yale economists found the shift led to a vast expansion of the manufacturing workforce, in absolute numbers and as a share of the nation's.

Mara: Eli, that precedent fits one reading, held by economists like Neil Thompson: AI mostly complements people and spurs new kinds of work. The other reading is the one Jack Dorsey's warning belongs to: AI will automate faster than retraining and new industries can absorb the people it displaces.

Eli: Whether a shoe factory says anything about a translation desk is still open.

Mara: Our translator found a clue of their own. An existing client, quiet for a while, sends a large document. Refer to our website for terminology, they say. So the translator reads the site. It had all the telltale signs of AI, they said: acronyms and terms translated inconsistently, which they called a big sign, the usual weird constructions, and some of it "just plain nonsense."

Eli: So the cheaper option was already sitting on the client's own website, by the translator's reading.

Mara: The writer behind Blood in the Machine argues that "an AI system does not have to be super-powerful for management to use it to degrade, deskill, and kill jobs." And the translator was being asked to match terminology that, by their reading, a machine had already garbled.

The Demand Trap and the Physical Wall

Eli: Two economists, Brett Hemenway Falk and Gerry Tsoukalas, built a model of exactly that puzzle. Picture one firm replacing workers with AI. It keeps the whole saving. But those workers were also somebody's customers. When they stop spending, the lost sales spread across every competitor. Falk and Tsoukalas call that a demand externality: a cost a firm causes that mostly lands on someone else. So each firm automates anyway, and none can escape on its own, they argue. And the more fragmented the industry, the tighter the trap.

Mara: For the client, just software cheaper than a rival's translator.

Eli: And the cost piles up somewhere. Research published through the National Bureau of Economic Research argues that even when new jobs push back hard, automation raises output per worker more than wages. Labor's slice of national income shrinks. How strong that effect is remains contested.

Mara: Which is how a tool that's still weak on long projects can empty desks, and still get the translator's quote turned down. What nobody knows yet is how far that race can run before the machines hit their own limits.

Eli: Some limits are already visible. Physical tasks are about half the work in the US economy. A 2026 Anthropic study found today's robots can technically handle about three quarters of those tasks. But by its estimate, which is contested, robots are cost-competitive with people for just 0.3 percent of work. Most need tightly structured surroundings. Then there's power. TechRadar reported that electric-grid connection is the primary challenge affecting AI data centers, though that claim is contested.

Mara: Same shape as the office numbers. Anthropic also lists work still out of AI's reach: pruning trees, running farm machinery, representing a client in court.

Eli: The jobs a machine can't cheaply take sit somewhere between, while the walls hold the machines back.

The Resilient Core and the Fiscal Dilemma

Mara: That middle ground has people in it. Boston Consulting Group looked at current AI and found that 57 percent of jobs depend heavily on physical presence, hands-on work or sustained human interaction. Those, in its reading, are the least likely to be disrupted. A talent chief at the bank HSBC named what she sees as the ground worth holding: "It's the human aspects: empathy, flexibility, adaptability, resilience, relationship-building. That's where I think the world of work will be contested in the future."

Eli: There's an institutional floor as well. The analysis site JobsVsAI reports that wherever lives, public infrastructure or legal compliance are at stake, institutions require a human who can be held accountable.

Mara: That protects some chairs. It doesn't refill a translator's inbox.

Eli: Then the money. A 2023 congressional report, cited by Brookings, found about three quarters of US federal tax revenue comes from labor. So payrolls carry the government too. Falk and Tsoukalas, the economists behind that trap, propose a Pigouvian automation tax. It's a fee charged each time a firm replaces a worker with AI, the way a polluter can be charged for smoke that lands on the neighbors. They argue it's the only fix that makes each firm pay for the demand it destroys.

Mara: Bernie Sanders proposes something bigger. A one-time tax, paid in stock, equal to half of any company pulling in large AI revenue. Those shares would go into a new American AI Sovereign Wealth Fund. Sanders is among those arguing that redistribution funded by taxes on capital can protect welfare and keep consumer demand alive.

Eli: The other reading starts with the bill. A tax policy group priced a poverty-line basic income at 30,000 dollars a year for a family of four. Paid to everyone, that's about 8.5 trillion dollars. And where Sanders's tax is collected once, that bill comes every year. IMF chief Kristalina Georgieva is among those arguing that cutting income loose from work risks severe deficits, dulls the push to innovate, and can't replace the purpose a job gives people.

Mara: Eli, while it's argued over, the translator still has a career to rebuild.

The Long Horizon

Eli: Those 22-to-25-year-olds at least have a whole career ahead to adjust in. Not everyone does.

Mara: Our translator is still 15 to 20 years from retirement. They wrote that they've had to put so much energy into reskilling, for teaching and editing. Those jobs don't pay as well, they said, but they still require in-depth, professional skill. Another translator in that collection said her income fell 60 percent in 2024. For 2025, she expected it to land 80 percent below where it sat from 2020 to 2023.

Eli: Even on the long jobs the software mostly still fails?

Mara: It never had to beat the translator.

Eli: So, the listener's question. On the evidence, AI isn't better than humans at almost everything, and the experts surveyed think that's decades off. A 2026 forecasting survey called LEAP set a bar for artificial general intelligence. It's a system anyone can buy that beats the top tenth of professionals on at least 90 percent of useful non-physical tasks, in every mostly desk-based occupation, at no more than five times a person's cost.

Mara: Better than nearly every professional, at nearly every desk task, and affordable.

Eli: The median expert gave an 80 percent chance that a majority of panelists agree it exists before 2100. If they do, the middle guess is 2050. Contested forecasts, all of them.

Mara: One respondent spelled out the doubt: if it hasn't come by 2040, the likeliest reason is that today's approaches hit fundamental ceilings that scaling alone can't overcome. And the two-thirds quote didn't wait for any of that.

Eli: Three signals will show which way it breaks. Buyer feedback collected by Base10 Research, on whether multi-task robotic reliability exceeds 95 percent across different kinds of physical work. The monthly payroll count for 22-to-25-year-olds, and whether that hiring gap spreads into general unemployment. And whether frontier models break through on multi-day work, or hit a plateau before 2040.

Mara: A concept to read up on: jagged capabilities.

Eli: Follow the economy with us: how should companies balance automation and jobs?

Mara: That's where we'll leave it. I'm Mara.

Eli: And I'm Eli. Thanks for listening. Until next time on Angle.