MarsBit has published a commentary by Xiaobing that takes up one of the sharper phrases circulating in AI debates: the “permanent underclass.” The piece opens with Sam Altman on stage at Chase Center in San Francisco on July 26, where he said, "It’s ridiculous to say that if you don’t get into a frontier lab, you’ll fall into a permanent underclass."
From there, the article tracks how the term returned to view. It says the phrase gained traction on X in tech circles in the summer of 2025, with its framework drawing on The Intelligence Curse and the paper Capital in the Twenty-Second Century by Trammell and Patel.
In the article’s telling, this is not just another warning that AI could eliminate many jobs. It describes a more extreme possibility: if AI becomes capable of handling most cognitive and physical labor, firms may need far fewer workers to create wealth. Wages would then matter less inside the economy, while profits would flow more heavily to those who control models, compute, energy and data centers.
The point, the author argues, is not simply that some people could become poorer for a time. It is that labor itself may stop functioning as a route to mobility. Nor is the issue only that the rich get richer. It is that people without capital may no longer be able to become owners of capital through work. That is what the article means by a “permanent underclass.” Without large-scale redistribution, public ownership or a global progressive capital tax, the theory points to a bleak conclusion: most assets could gradually move toward the people who were already richest when the AI transition began.
A question researchers cannot answer
The article then turns to writer Jasmine Sun, who it says spent the past year on a kind of field study. She met researchers at frontier AI labs one by one, spoke with each for an hour, made no recordings and kept no notes, and pressed them on what they actually think the future of AI looks like.
One question came up again and again. Imagine an ordinary 17-year-old American standing in front of you. He is not a genius programmer. He gets B grades. He does not think much about AI. What should he do now to prepare for the future?
According to the article, almost no one had a real answer. These researchers differed politically and ranged from optimistic to pessimistic in their views on AI, yet their replies were strikingly similar: they did not know. The situation already looked frightening to them, the range of jobs left for that teenager might be narrow, and he could land in the most painful part of the transition.
The article says the teenager’s problem may not simply be whether he becomes poor. The harder part is that what turns “underclass” into something permanent is the possible disappearance of the ladder itself.
What becomes permanent is the loss of the route upward
Every era has had poor people and lower classes, the author writes, but being at the bottom did not automatically mean staying there forever. A person could sell labor for wages, then turn wages into education, housing and assets. Workers could also organize, strike and bargain collectively, forcing capital to share some productivity gains through better pay and benefits. The arrangement was never fair, but it left an opening.
That opening is what the “permanent underclass” idea says may vanish. Looking back at labor history, the article notes that unions, minimum wages and the weekend all rested on the same condition: capital and labor could not fully replace one another. Factory owners needed workers. Tech firms needed engineers. Both sides had to negotiate.
The article invokes Marx here. He predicted high capital concentration, but in his framework workers still held something capital needed and could not simply manufacture from thin air: labor. That is why strikes worked. Workers could withhold it. If workers did not enter the factory, machines stopped. If engineers did not write code, products did not ship.
But if compute can buy all the labor one person can supply, then what workers can withhold approaches zero. The problem is no longer only lower pay. It is that capital may no longer need to bargain with most people at all. Workers would struggle both to accumulate assets through wages and to use work stoppages to push for a different distribution of returns.
The machine that once moved poor people back toward the middle class, in the author’s framing, was built from labor, bargaining and asset accumulation. What makes the lower class “permanent” is that this machine could be dismantled.
Signs of strain are already visible in payroll data
The article argues that the first signs have moved from theory to payrolls. It cites Stanford’s Digital Economy Lab, which used ADP microdata covering 4.6 million workers. After generative AI spread, employment among workers aged 22 to 25 in the occupations with the highest AI exposure saw a roughly 16% relative decline, according to the article. It also says the U.S. unemployment rate for recent graduates reached 5.6%, up 1.6 percentage points from three years earlier, while entry-level hiring at large tech companies fell 25% over two years.
The author condenses that shift into a single line: the elevator is still running, but the button for the first floor has been removed.
People who cannot answer are still building, and buying insurance for themselves
The piece returns to Sun’s interviews. The researchers who could not answer what the 17-year-old should do did not stop what they were building. Sun pushed them further: if the future looks this dangerous, why keep making it?
The answers fell into three broad groups, the article says. One group sincerely believes that if society can get through the transition, AI could eventually cure disease and remove scarcity. A second takes a deterministic view of technology: if they do not do it, someone else will. The third is the bluntest and, in the article’s view, the most uncomfortable: if major disruption is coming, better secure a place in that future first.
That urge to secure a place is changing talent flows across the AI sector. The article says a friend of Sun’s, an AI PhD student at Berkeley, estimated that about half of that cohort chose to graduate early. Some even picked dissertation topics based on which direction was most likely to land them an offer from an AI lab. It also says many independent writers and policy researchers around Sun left their previous positions and joined AI labs.
Those stories may sound like anecdotes, but the author argues they are materially shaping life choices for a generation and steadily pulling talent out of independent research and public policy. The people being drawn away are precisely the ones most likely to build a bargaining table.
Why today’s white-collar transition looks different from factory automation
The article compares this moment with 20th-century factory automation. One reason mechanization did not broadly turn into severe conflict, it says, is that management usually had to negotiate with unions before machines entered the factory. Automation had to be presented as a safety upgrade, and pay growth had to be linked to productivity gains. There was a bargaining table between capital and labor.
Today’s white-collar workers do not have that table, the article says. More subtly, the people with the best chance of building one—academics, independent researchers and policy talent—are being pulled into labs by the belief that labor is about to lose value. Inside the lab is where they might receive equity and move to the capital side before the shift fully plays out.
The article describes a closed loop. The more people believe labor is about to lose bargaining power, the more they stop trying to build bargaining power and start competing for equity instead. The fewer builders remain, the faster labor’s bargaining power erodes.
No one is completely safe
For people trying to respond to AI uncertainty, the article says two reactions show up most often. One is to learn AI as fast as possible and aim to be among the last to be replaced. The other is to acquire AI assets as fast as possible and try to move from labor to capital before labor loses value. The second impulse, the author writes, is especially visible in Silicon Valley and even among wealthy groups in China.
Some people are determined to obtain equity in Anthropic, OpenAI or other frontier AI firms. The theory is simple: if a singularity-like turning point arrives, labor could devalue quickly, and what decides a person’s position would no longer be what they can do but what they already own. In that logic, converting labor income into AI equity before the window closes is a way to move from being displaced by the technology to owning the technology.
This is also what makes the “permanent underclass” narrative so seductive, the article says. It does not only produce fear. It hints at an escape route: if capital may replace labor, become a capital owner as quickly as possible.
But that route is not open to most people. OpenAI and Anthropic are not companies that ordinary individuals can easily buy on public markets. The people with real access to those shares are usually lab employees, early investors and wealthy participants able to enter private markets. The argument turns circular: to avoid falling because you lack capital, you first need access to the kind of capital the upper tier already controls.
The article then introduces another complication. Even if someone gets the shares, that insurance may not last forever. Fernando Borretti, creator of the programming language Austral, is cited as arguing that the theory itself contains a contradiction. If AI can perform nearly all cognitive and physical labor at lower cost, then ordinary workers may lose economic value, but those who bought AI shares early are not guaranteed to become a “permanent upper class” either.
Wealth is not encoded into the laws of nature, the article says. Owning a company, land or compute depends not only on a name written into a contract, but on courts, police and governments that recognize and protect those property rights. If one keeps extending the “superintelligence replaces everything” premise, AI could ultimately handle production, management, governance and even war. At that point, today’s wealthy people would neither supply labor nor control real force outside the machines. Why, the author asks, would a future state or a superintelligence necessarily keep honoring ownership claims acquired in the old era?
Buying equity in AI companies may help someone through a transition, then, but it cannot guarantee a permanent place at the top.
Learning AI carries its own paradox
If owning AI assets is not an option for everyone, the more general answer becomes learning AI. But that route has a paradox of its own. The better a person becomes at using AI to raise productivity, the more likely that person is to help the firm reduce demand for other workers. He may stay in the elevator longer, but he is also helping remove the buttons for the other floors.
The article does not argue that people should avoid learning AI. Quite the opposite. For individuals, becoming fluent with AI may still be the most rational move available now. The problem is what happens when everyone makes the same rational move. If all workers try to avoid replacement by becoming more efficiently replaceable, the aggregate result may be that firms need fewer workers overall. Individual rationality, added up, could accelerate the decline of labor’s bargaining power.
A third answer appears as well: leave fields that AI can easily replicate and move into work that requires physical presence. If digital goods become nearly unlimited in supply, then physical operations that cannot be copied remotely, on-site responsibility and person-to-person trust may command a premium. Data centers need electricians. Aging societies need caregivers. People may still pay for the real presence of doctors, teachers and service workers. But the article says this is not a universal escape hatch either.
Those three paths look different, the author writes, but each is answering the same question: how do I fall later than someone else? The question the 17-year-old actually needs answered may be another one entirely: why should a person’s ability to survive depend on whether capital still needs him?
What is missing is not a career plan but a distribution system
The article argues that researchers cannot answer because the real problem does not call for better personal planning. It calls for a new distribution system. When labor remained hard to replace, wages, unions and strikes together formed a mechanism for sharing output. If labor truly begins to lose scarcity, society will need some other way to distribute the wealth created by machines to people the machines no longer need.
That cannot be solved by everyone studying AI harder, the author says. Nor can it be fixed if everyone tries to buy a few shares in advance. What individuals can buy is time. What society lacks is still the bargaining table.
China and the U.S. as contrasting cases
On the question of how to build that table, the article presents China as a contrast case. It says that in December 2025, a court in Beijing ruled that having a position replaced by AI could not directly serve as a lawful basis for dismissal. It also cites a state-owned enterprise employee who said the AI tool he used could do the work of about two employees, yet the company promised not to cut staff on that basis.
After spending two weeks in China, Sun concluded that the prevailing social attitude was not simple techno-optimism but a kind of pragmatism: resistance is not realistic, so get on board first. Still, the article says, someone is at least trying to catch people falling through the transition with a large number of granular rules and pieces of legislation.
The U.S., by contrast, is presented as being closer to an institutional vacuum. In April 2026, someone threw a Molotov cocktail at Altman’s home in San Francisco. In the same month, the home of a city council member who backed a data center project was shot at. Commencement ceremonies also began to feature boos aimed at AI company executives. When emotions have no formal outlet, the article says, they find one on their own.
At the same time, Federal Reserve Bank of St. Louis data showed that 39% of U.S. GDP growth in 2025 came from data centers and AI-related investment. The article reads that as a sign that the country is staking growth on a technology whose benefits most people do not yet feel.
The next two years may hinge on whether the table appears before anger does
The final section turns to the U.S. political calendar. The article notes that the country will hold midterm elections in 2026 and that the 2028 primary season is already set to be crowded. Mark Kelly, Ro Khanna and Josh Hawley have each released AI platforms.
That is why, in the author’s view, the single most important thing to watch over the next two years may be whether the bargaining table can be set before anger gets there first.
The table could take different forms. It could be a law that bars companies from converting productivity gains directly into layoffs. It could be a new type of union that lets employees of the same company bargain over AI deployment even if they do different jobs. It could also be a check carrying the name of an AI company, redistributing the gains generated by the technology through taxes, dividends or public funds to people bearing the cost of the transition.
The exact form remains undecided. But the core question, the article says, can no longer be avoided: if companies no longer need to employ most people in order to create wealth, on what basis do most people get to share in that wealth?
As for the 17-year-old with B grades, the author closes on a bleak note. He may never know that when the smartest people in an entire industry were asked what should be done for him, their answer was that they had no answer. Then they went back to buying insurance for themselves.