Jasmine Sun's reported essay on the permanent underclass exposed the first half of the problem: knowing without stopping. The people closest to the technology can describe the danger, explain the democratic risk, publish the policy paper, and continue building.
Her follow-up exposes the second half. After the capability leaves the lab, it does not carry one fixed labor meaning into the world. The context that receives it translates it.
Put the same image model in front of two designers. The senior designer is still responsible for taste, judgment, client context, and the first shape of a brand. The model gives her more room to try. The junior designer turns finished guidelines into disposable assets. The model gives the company a way around her.
The same system widens one job and compresses another.
Put the same coding assistant inside three software firms.
The first has more demand than it can satisfy. Every engineer made faster opens more product surface, more experiments, more things worth hiring people to pursue.
The second has flat growth and investors waiting for margin discipline. Every engineer made faster becomes evidence that fewer engineers are needed.
The third has revenue, cashflow, too many projects, too much process, and no obvious next market. It cuts, not because the future is over, but because the organization has become too noisy to search. The savings might become buybacks and layoffs. They might also become the budget for the next thing that cannot yet be described in a board memo.
Again, the tool did not change. The receiving context did.
Automation is context. Augmentation is context. Margin is context too.
The comforting version of the AI labor debate draws the moral line around the product. Tool AI augments. Agent AI automates. Keep the human in the loop and work survives; remove the human and work disappears.
That distinction matters, but it is too shallow to decide the outcome.
A lab can design affordances that make collaboration easier. It can prefer copilots to agents, workflows to replacements, human review to one-click execution. Those choices bias the result. They do not control absorption. Once the tool enters an organization, the organization asks its own question: does this let us do more, does this let us spend less, or does this let us search again?
If the firm has more profitable work than people, AI becomes scope. The human remains scarce because judgment, authority, client trust, taste, and responsibility remain bottlenecks. If the firm has more payroll than growth and no credible frontier, AI becomes extraction. The same human-in-the-loop posture becomes a transition state on the way to fewer humans.
But cost discipline is not automatically extraction. A company can cut because it is dying, because owners want cash, because managers want an earnings beat, because the market has saturated, or because the company has lost the operating clarity required to do hard new things. Those are different contexts. The layoff notice may look the same. The strategic meaning is not the same.
That is why "we build tools, not job cuts" is not an answer. It can be true at the product layer and false at the labor layer. A tool becomes a job cut when it enters a context whose current problem is labor cost. It becomes search capital when it enters a context whose current problem is organizational noise between the firm and its next frontier.
The product did not become evil in the second firm. The second firm had a different constraint. The third firm has a different constraint again.
The press can see a cost cut. It struggles to see what the cut is for.
A headcount reduction has a date, number, memo, and stock-market reaction. The market that a firm may discover two years later has none of those. It has no customers yet, no category, no budget line, no job titles, no trade association, no constituency. If it already had those things, it would not be the next market. It would be the current one.
The CEO with cashflow to spare is not holding a pile of obvious future demand. She is holding the right to search. Cashflow buys time, talent, experiments, acquisitions, tooling, failed prototypes, and the organizational patience to test a direction before the market has learned to ask for it. But search is not the same as spending. A bloated organization can burn cash without learning. A focused organization can spend less and search harder.
Meta is the public case because the sequence is visible. In 2022, Brad Gerstner urged the company to get fit and focused. In 2023, Mark Zuckerberg made the Year of Efficiency official, framing efficiency as a way to make Meta a stronger technology company and improve financial performance so it could execute its long-term vision. By 2025, Meta had shifted the center of its frontier bet toward AI, invested $14.3 billion in Scale AI, and recruited Alexandr Wang, Nat Friedman, and Daniel Gross into the superintelligence effort.
This does not prove Meta is right. The AI bet may work, fail, or become the metaverse mistake with better timing. The point is categorical. The efficiency campaign was not simply the opposite of innovation. It was a way to regain strategic optionality after a period when the company's spending, story, and confidence had drifted apart.
Scale is a smaller version of the same lesson. Wang publicly described Scale breaking even as a function of building a more sustainable AI business than the previous generation of AI companies, and later projected profitability while raising at a much larger valuation. In a venture world trained to treat burn as ambition, profitability was not anti-growth. At one point internally, he spoke of "cheapo mode" and subsequently the company developed prelabeling workflow automations to reduce human labor costs associated with Remotasks. This was a constraint that _fueled_ growth and enabled an expanded real business.
The older management version is capability discipline. Charles and Chase Koch describe Koch Industries as capability-bounded rather than industry-bounded: not "we are in oil, therefore everything oil," but "where can our demonstrated capabilities create more value than others?" That frame matters because spare cash without self-knowledge becomes empire-building. Spare cash with deep self-knowledge becomes search across adjacent opportunity.
Cost cutting is good when it removes noise between the firm and the capability it can actually compound. Bad when it removes the apprenticeship, judgment, trust, and operating memory that made the firm capable in the first place. Neutral when it transfers cash from payroll to shareholders. The accounting line cannot tell you which happened.
The context decides.
At the firm level, productivity becomes expansion, extraction, or search.
Expansion is the clearest case. Demand outruns capacity. AI lowers the cost of serving demand. The firm opens backlog, experiments, product surface, and hiring around the judgment layer.
Extraction is the ugly case. Demand is flat. The firm has no credible next market, no capability-bounded search thesis, and a capital market that rewards margin. AI lowers the cost of producing the current output. The firm returns the savings through layoffs, buybacks, or earnings repair.
Search is the hard case. Demand for the current product may be mature, but the firm has cashflow, capability, and a plausible frontier. AI and cost discipline reduce noise so the firm can run more strategic experiments without losing itself. The immediate labor effect can still be painful. The moral and economic question is whether the freed capacity becomes a new path or only a harvested margin.
At the household level, productivity becomes either more human service or cheaper digital substitution. The human-premium argument is right at the top of the distribution: rich households can buy more tutors, therapists, trainers, chefs, assistants, and event-makers because machine abundance makes human attention feel more valuable. Normal households may move in the opposite direction. If the human was already expensive, the digital substitute wins by being available, patient, and good enough.
The human premium is real, but not automatically a mass-employment plan. Without broad purchasing power, it becomes status labor above a floor of digital substitutes.
At the state level, productivity becomes either employment policy or political instability. Sun's China contrast matters because it changes the loss function. An American executive under capital-market pressure can treat AI layoffs as discipline. A state that prices idle young people as political risk may treat employment as public-order infrastructure. That does not mean the state has solved AI labor disruption. It means the state may rationally preserve inefficient work because instability is more expensive than inefficiency.
At the cultural level, productivity becomes either a race to self-amplify or a politics of protection. Sun describes a Chinese "save yourself" register around AI: learn the tools now because refusal only hurts you. The American permanent-underclass register sounds different because it is voiced by a class used to being the automaters, not the automated. One culture says the system will not stop for you. The other still believes public argument might make the system answer.
Neither register is simply optimism or pessimism. Each is a survival strategy fitted to a different context.
The optimistic labor answers keep assuming demand appears on schedule.
The Jevons answer says cheaper production increases demand, so jobs return somewhere else. Often, cheaper production does increase demand. But there is no managerial shortcut for creating the new demand. A firm cannot order a market into existence because it saved money. It can only make search cheaper, faster, more reversible, and more disciplined.
This is the CEO problem that labor commentary usually skips. The executive does not start with the next market fully visible. She starts with cashflow, customers, talent, infrastructure, brand, technical capacity, habits, politics, and accumulated commitments. Some of those are assets. Some are anchors. The job is to know the firm deeply enough to tell the difference before the outside world can.
That is why cost discipline often precedes innovation. Not because starving a company magically makes it creative. Starvation usually makes it stupid. The useful version is different: remove the projects and processes that were absorbing attention without producing learning, preserve the capabilities that make the firm special, and convert the freed margin into tests whose outcomes can teach the firm where demand might exist.
New markets are usually impossible to describe in advance because they are partly produced by the act of searching. AWS did not begin as a cleanly pre-existing "cloud infrastructure market" waiting in a drawer. It emerged from Amazon's operating capability, customer pain, pricing granularity, and reversible trial structure. The market became obvious after the product made it legible.
AI may create many such markets. It may also create none for a particular firm. The difference will not be found by asking whether the model is a tool or an agent. It will be found by asking whether the receiving firm can turn cheap capability into disciplined search before competitors, incumbents, or its own bureaucracy collapse the window.
Jevons needs a path from cheaper output to human bottlenecks. If cheaper software creates more demand for software, and the new software demand is also handled mostly by agents, output can rise while the human share falls. The demand curve is not enough. You need to know who satisfies the demand.
Search capital needs a real searcher. Cutting costs inside a firm without capability-bounded judgment does not create innovation. It creates a cash pile, a buyback, or a management team pretending every adjacent market is adjacent. The capability-bounded test is sharper: what has the firm actually learned to do that creates more value than others can create?
Distribution matters too. A narrow capital-owning class cannot consume the same labor basket as a broad middle class. A few wealthy households can buy more relational service. They cannot each consume the work of millions of displaced office workers. If AI concentrates purchasing power faster than it expands broad demand, the service economy does not catch everyone. It tiers.
The human-premium answer has the same conditional structure. It works when people with money want visible human attention. It fails when the buyer would rather have convenience, privacy, consistency, or price. Scarce human attention does not automatically win. The context decides whether scarcity is a feature or a cost.
Sun's follow-up keeps rubbing the optimistic answers against different contexts. Tool AI, Jevons, cost discipline, and the human premium are not wrong. They are conditional. They need context before they can predict anything.
Before taking an AI labor claim seriously, ask six questions.
What context is the capability entering?
What does that context currently reward: scope, margin, search, price cuts, public stability, status, convenience, or survival?
If productivity becomes margin, what is the margin being converted into: extraction, survival, or search?
Does the firm have a capability-bounded thesis for where new demand might be discovered, or only a hope that savings will make innovation happen?
Who has purchasing power after the productivity gain is captured?
Which human bottleneck remains, and is it scarce enough to preserve work beyond the top of the distribution?
These questions make the claim falsifiable. If stagnant firms adopt AI and expand hiring at the affected layers without a new demand path, the extraction story is wrong. If efficiency campaigns regularly produce new markets rather than buybacks or empire-building, the search-capital story strengthens. If cheap digital substitutes create broad human-service demand rather than luxury segmentation, the household story is wrong. If China lets AI-attributed layoffs run as freely as American capital markets encourage them, the state-stability story is wrong. If workers in the "save yourself" culture produce less AI fluency than workers in the politics-of-protection culture, the cultural read is wrong.
The point is not to force every case into one answer. The point is to stop asking for one answer before identifying the context.
The permanent-underclass debate keeps looking for a single verdict because it is still asking whether AI replaces workers. The answer is yes, no, not yet, and sometimes it funds the search that creates the next work.
That sounds evasive until the mechanism is named. The producer does not get to decide what its capability is by naming its intended use. Tool, assistant, copilot, agent, model, platform, public good: those names matter less than the system that receives it.
The receiving system is where moral language becomes labor outcome.
The optimist cannot say "tools augment" without asking where the tool lands. The pessimist cannot say "AI replaces workers" without asking which workers, in which contexts, under which constraints. The efficiency defender cannot say "cost cuts fund innovation" without asking whether the firm has the self-knowledge, cashflow, and reversible trial structure required to search. The producer cannot say "we only build capability" because capability is built to be absorbed somewhere, and the likely absorptions are not mysterious.
The aggregate curve can be right and still miss the life. GDP can rise, software can get cheaper, output can increase, Meta can make the right AI bet, Scale can build a sustainable business, Koch can compound capability across industries, and a specific cohort can still lose the only context in which its labor had value. Averages hide contexts. Policy has to build contexts on purpose.
Do not ask whether AI is good for jobs. Ask which contexts turn AI into jobs, which turn it into cuts, which turn it into search, which turn it into public-order management, and which turn it into a luxury economy with digital substitutes underneath. Then argue over which contexts to build.
Automation is context. Augmentation is context. Efficiency is context. The model supplies capability. The receiving system decides what the capability becomes.
Source trail: Jasmine Sun, "Party in the Permanent Underclass"; Brad Gerstner's October 2022 Meta letter as reported by CNBC; Meta's March 2023 "Year of Efficiency" memo; Associated Press and CNBC reporting on Meta's 2025 Scale AI investment and Superintelligence Labs; Alexandr Wang interviews with The Business of Business and Fortune on Scale's break-even, sustainable growth, and profitability expectations; Charles and Chase Koch on the May 2026 All-In Podcast.