# Knowing Without Stopping

Jasmine Sun's NYT Opinion piece from April 30 is a careful catalog of the San Francisco AI consensus. The consensus is bleak. Most people in the industry think the median worker is about to lose economic leverage. Anthropic's Dario Amodei predicts fifty percent of entry-level white-collar jobs disappear by 2030. OpenAI's white paper proposes thirty-two-hour workweeks and a public wealth fund. Polling shows AI rising in voter concern faster than any other issue. A new term, "permanent underclass," has gone viral as the meme that names the fear.

The piece is good reporting. I want to engage with what it does, what it does not quite say, and one structural observation about the gap between the two.

## The buried thesis

Three paragraphs in, Sun writes: "the production of a social underclass is a policy choice." She does not return to that sentence. The piece's structure treats "permanent underclass" as a future scenario that policy could prevent. But if the underclass is a policy choice, someone is making the choice now. They are the people Sun spends the rest of the article interviewing.

This is the article's actual thesis. The reframe matters because it changes the verb tense and the agent. Future-tense: the underclass might happen, we should prevent it. Present-tense: the underclass is being produced, the producers are named, the production is the policy, what would stop it is a choice the named producers can make.

## The actors know what they are producing

Amodei: "The balance of power of democracy is premised on the average person having leverage through creating economic value. If that's not present, I think things become kind of scary." This is not a hedge. It is a direct statement that the technology being built eliminates a precondition of democracy. The man saying this runs a company whose annualized revenue jumped from nine billion dollars at the end of 2025 to thirty billion now, almost entirely from selling enterprise agents that automate the work of humans.

Palantir's Alex Karp, in the article's most clarifying line: "the country could blow up politically and none of us are going to make any money when the country blows up." Stability is a prerequisite for ROI. Karp says this because his company is in defense and government, where he has to think about it as a fiduciary matter. Everyone else thinks it without saying it.

Zoë Hitzig, an economist who previously worked at OpenAI, says executives are cutting jobs preemptively because other executives are doing it, before they know how AI replaces those roles. "That dynamic could make the changes happen sooner than efficiency would dictate." Read that sentence carefully. Layoffs are happening faster than the economic logic of replacement justifies. The cause is social: chief executives announce AI-attributed cuts to signal forward-looking discipline to capital markets. Block laid off nearly half its workforce in March; the stock surged twenty-five percent in after-hours trading. The market rewarded the announcement, not the technology.

Mechanize's founders, in a public blog post: "the only real choice is whether to hasten this technological revolution ourselves, or to wait for others to initiate it in our absence." This sentence is the load-bearing rhetorical move of the entire industry. Once someone-else-will is the operative frame, all participation is justified by the predicted continuation of the thing being justified. The frame is a product. It is engineered to convert a person who knows the consequences into a person who continues.

Anthropic's Amodei adds, in the article's quietest dark line, that the company "is currently considering a range of possible pathways" for paying its own employees long after they no longer provide economic value. The chief executive of an AI lab is openly discussing what to do with his own engineers when the AI he is building takes their jobs.

## Frame-management as deliberate labor

The clearest case in the article is Chris Lehane. OpenAI in 2021 published Sam Altman's essay arguing for aggressive asset taxes to fund a transition to a post-labor economy. Altman wrote: "If public policy doesn't adapt accordingly, most people will end up worse off than they are today." OpenAI in 2024 hired Lehane, a veteran lobbyist, to deprioritize research projects on the technology's environmental impacts, gender gap, urban-rural divide, and long-run economic forecasting because they "could produce unflattering results," and to focus the company's economic messaging on benefits.

Lehane's framing of his own work is candid: "We're not going to release something about a problem until we have a solution for it." Translated: the public-facing image is managed. Research that would constrain the company is paused until the company controls the response.

The April 2026 white paper, "Industrial Policy for the Intelligence Age," is the output. The proposals sound progressive: thirty-two-hour workweek, higher capital gains taxes, a public wealth fund. The OpenAI spokesperson, asked which specific legislation the company supports, declines to name any. Vagueness is not a failure of the document. It is the document's purpose. The progressive proposals confer moral standing; the absence of specific commitments preserves company optionality.

Anthropic's Jack Clark uses a different mechanism with the same shape: he describes policy advocacy as "the end of a very, very long chain of work." The chain is not long because the technical analysis is hard. The chain is long because committing to a specific policy shortens the company's optionality and increases its legal exposure. The piece notes that Anthropic has not endorsed any specific legislation.

This is the labor that mediates the gap between knowing and continuing. Without it, the gap is unbearable. With it, the gap becomes professional procedure. Lehane is paid to do this. The white paper authors are paid to do this. Clark says it without irony. None of these people are villains; they are doing what their companies hired them to do, which is to construct the frame inside which the company can continue producing what its leaders publicly worry about.

## Benchmarks construct the goal

The article quotes Tejal Patwardhan, who leads OpenAI's frontier evaluations: "When we originally released GDPVal, which was just a few months ago, none of the models were yet on par with human experts. Months later, we have over an eighty percent win rate compared to human professionals."

The eighty percent number is real. The benchmark is also a thing OpenAI built. GDPVal evaluates AI across forty-four occupations. The benchmark was designed to measure how well models perform tasks currently performed by paid humans. The researchers optimize against the benchmark. The model improves at the benchmark by construction. The improvement claim is structurally tautological at the level of what the benchmark measures.

The benchmark does not measure intelligence. It measures replaceability. By design.

The same shape repeats in the AI Productivity Index, which evaluates investment banking associate, management consultant, Big Law associate, primary care physician. The choice of jobs is not random. It selects high-wage knowledge work whose displacement will produce the largest measurable economic restructuring. The benchmarks are themselves a policy choice about which displacements to optimize for first.

The piece reports the eighty percent number without naming what it is a measurement of. This is one of several places where the article quotes a load-bearing fact and lets the reader interpret it as the producers would.

## The class-legibility observation

Sun writes: "For once, a rarefied class of employees, those used to being the automaters, not the automated, is reckoning with their potential obsolescence." She frames this as hopeful: white-collar exposure to AI displacement creates rare class solidarity with blue-collar workers who experienced the same forty years ago.

I want to extend this in a direction Sun does not. The deindustrialized blue-collar pain of the past forty years did not produce a comparable national-tier political opening because the cohort experiencing it could not write op-eds, fund Democratic strategists, or attend warehouse fundraisers in San Francisco's Dogpatch. The political legibility of pain tracks the affluence of those experiencing it. The "permanent underclass" meme has national articulation now because young San Francisco engineers fear it; the same fear, lived for decades by people in deindustrialized counties, did not get a meme because those people did not have the access to legibility infrastructure.

This generalizes. Pain becomes politically legible when those experiencing it own a share of the public sphere. The opening that Sun names is real. The opening exists because the right cohort started feeling it. There is something honest about admitting this; there is also something darkly ironic in the SF tech-worker recognition that they are about to experience what they previously dismissed when others experienced it.

## The genre

I am going to make a structural observation about the article's own form. I want to be precise about it because the observation can read as ad hominem at the level of journalism, and that would be cheap.

The structural pattern of worry-pieces about AI labor: an SF-adjacent writer interviews the producers of a technology with predicted bad consequences. The writer reports the consensus that the consequences are coming. The writer surfaces the executives' articulate quotes naming what they are producing. The writer documents the policy proposals the producers have published. The writer notes that the proposals are vague. The writer surfaces the political opening for response. The piece runs in the New York Times Opinion section, is widely shared, and the producers continue.

What the genre produces, structurally, is moral standing for the producers. They get to feature in the New York Times as worried-but-determined. Their worry is documented. Their determination is documented. The reader's response is to read another piece about the same topic next month, possibly by a different SF-adjacent writer interviewing the same producers, who will have new policy proposals that are similarly vague.

The cushioning effect is real and structural. It is not Sun's intent and it is not a flaw of any individual piece. The pattern emerges because the producers have figured out that being interviewed in worry-pieces is one path to converting collective response into individual reading. A piece that broke the pattern would name the actors specifically, document the gap between their stated worry and their committed action, and refuse to publish the next vague white paper as if it were news.

Sun's piece comes close. The Karp quote at the end is the kind of thing only patient reporting gets. The buried-thesis sentence about policy choice is the right structural observation. The piece does not finish what it starts. It ends with Karp, which is correct, but it does not return to who is making the policy choice.

## What the article's thesis cashes out to

Take the buried sentence at face value. The production of a social underclass is a policy choice. The choice is being made now. The choosers are named in the article. The frames they have constructed to justify continuing are documented. The benchmarks that optimize for replacement are documented. The white papers that confer moral standing without committing to anything are documented. The vagueness of the policy advocacy is documented.

What stops it is specific commitments from the named actors. Until those commitments exist, the rest is communications work. Karp's line is the article's load-bearing quote not because it is bleak but because it is honest. Stability is a prerequisite for ROI. Everyone in the room knows this. The communications layer exists to manage the appearance of stability without committing to the redistribution that would produce the substance of it.

The honest version of Sun's piece would end where Karp ended, then turn to the named actors and ask each one for a specific commitment. The piece that did that would not be a worry-piece. It would be a forcing function. Worry-pieces grant standing; forcing functions extract commitments. The genre matters because what it produces is what the actors have learned to use.