The Parlor Trick That's Burning Down the House

What we call AI is a subset of a subset of AI, and what it actually is—stripped of the tuxedo—is a parlor trick: elegant mathematics, trained on a heroic quantity of other people’s work, producing a very convincing impression of a mind. Let me be precise, because the precision is the whole disagreement. An LLM is a specific application of deep learning that uses the transformer’s attention mechanism to perform statistical inference over language tokens, scaled until certain emergent behaviors—reasoning, in-context learning—begin to flicker into view. The thing in your pocket autocompleting your emails and summarizing meetings that should never have been convened is not intelligence. It’s a next-token predictor that has read the entire internet and can tell you, with a perfectly straight probabilistic face, that the word after “the mitochondria is the powerhouse of the” is “cell.” That is not reasoning. That is a thermodynamically ruinous autocomplete, and we have built cathedrals of GPUs to worship it. It’s ELIZA again, with a better tailor and a worse electricity bill.

It’s ELIZA again, with a better tailor and a worse electricity bill.

The loudest people selling this are not the experts. Not the mathematicians. Not the engineers. Not the PhDs who spent decades appending confidence intervals and limitation sections to every claim they made, because they were raised right. There’s the occasional real technologist wheeled out at an investor day to lend the proceedings a patina of credibility—a hostage in a lanyard—but mostly it’s founders and financiers. The same people who brought you the metaverse, the crypto crash, the SPAC bubble, NFTs, and the theory that nap pods are a viable substitute for health insurance. They don’t understand the technology in any load-bearing sense. What they understand is its use as a narrative device—a lever, a cudgel, a permission slip. The AI story was never about building things. It’s about the elimination of headcount reframed as destiny. Nothing says “riding the rocket ship of history” quite like setting fire to your own staff and calling the smoke “momentum.”

Ask a question about the macro narrative—any question, not even a cynical one—and watch it get received as an act of aggression. How much are we spending on this? What’s the ROI? Is this a better use of capital than, say, conferences? Have we done any diligence on what the vendor does with our data? You raise one of these in a meeting and the silence that follows is a very specific silence: the silence of a room full of people who have staked their entire career on nobody asking it. Someone from the AI strategy team—and I’d like a quiet word with whoever decided the correct job title was “evangelist,” a word that ought to raise the hair on the neck of anyone who’s read a history book—shifts in their seat and says, “We’re really excited about the potential.” Which is not an answer. It’s a coping mechanism with a slide deck. And I’d know the tone anywhere, because I was raised in an actual evangelical setup, altar calls and all. I know a revival when I hear one. Question: why does the excitement always get louder the moment the evidence gets asked for?

I’ll admit my own bias: I have open contempt for the religiously devoted use of these tools, and I have sharper questions I can’t get within a mile of asking. Things like: how long until you replace all of us with a tool whose defining feature is that it never asks a follow-up and responds to every prompt like a salesman telling you that you look like a really intelligent guy? How long until the summarization model that can’t reliably distinguish a critical bug from a feature request is making staffing decisions? How long until the hallucination becomes the policy, the policy becomes the layoff, and the layoff is executed by a system that has never once been held accountable for anything—on the technicality that it isn’t a person, it’s a maths problem in a chat interface? None of these are rhetorical. That’s the trouble. Someone really should answer them.

Now factor in that the United States has, by a comfortable margin, the worst record on labor representation in the industrialized world, and that the political apparatus notionally responsible for the problem is occupied with other priorities. Ballrooms. Fountains. No-bid contracts to fix a reflecting pool for something north of the Apollo program’s budget, while the actual plumbing goes untouched. Grand ceremonial gestures, in other words, executed with great confidence and no plumbing. So you’ll forgive the absence of Pollyanna in my analysis. Real jobs—tech jobs—are being offshored in the tens of thousands. Trucking is automating. AI capex is driving very real cuts to very real American labor. And the policy response to all of this has been, to summarize the collected works: nothing. Not a memo. When, precisely, was labor going to come up?

Meanwhile the tech oligarchs are guests of honor at $50,000-a-plate galas and seven-figure crypto launches. So pardon my total absence of faith that anyone—any branch of government, any party, any executive board, anyone on the capital side of AI—regards labor with anything warmer than contempt. They cannot wait for it to crush and hollow out every white-collar job in the country, provided the crushing makes them one percent richer. And that’s the part worth sitting with: the destruction of millions of livelihoods is, on the arithmetic, worth it to them for a one-percent bump. That’s not hyperbole. It’s revealed preference—the economist’s polite term for what people show you they want by what they actually do, as opposed to what they put in the values statement. What do the earnings calls reveal, if not exactly that?

These tools were dropped on my team without anything resembling support. The expectation was that our output would, through some unspecified magic, increase exponentially. Nobody understood the tools. We were to build them into our product and simultaneously acquire deep expertise in them—no training, no resources, no investment—by osmosis, presumably, or prayer. The deadlines were, as ever, ambitious, arbitrary, and untethered from engineering reality. And the team assigned to do it wasn’t fully staffed until roughly one week before launch. We had a single quarter to ship a production-grade AI application built on a technology that none of us—that nobody, at that point—had ever used. One quarter. On a technology that hallucinates, degrades under load, and changes behavior without warning the moment the vendor pushes a model update. What could possibly require testing?

The pressure was enormous, and some of my team left over it, and I applaud every one of them for defending a boundary. They understood something the invisible man—the same one from the boardroom, the shadow that doesn’t know your name but knows your headcount cost—structurally cannot: that their finite, mortal, irreplaceable lives are not fuel for a hype cycle. So they walked. Some are thriving now at companies that still build real things. Some are still recovering. All of them, it turns out, read the situation correctly. Which of them, exactly, was wrong?

But one thing genuinely puzzles me: the experts are missing from the room. They’re present for the fun parts—building tooling, explaining hard concepts, geeking out over some classification technique hot off the press. The real experts—the mathematicians, the engineers, the people who actually understand the loss curves and the attention mechanism and the catastrophic-forgetting problem—are extraordinarily generous the moment you approach them with genuine curiosity. But on the strident claims—labor replacement, AGI in eighteen months, most-important-invention-since-fire—they go quiet. Cautious. Measured. Conspicuously unwilling to be reckless with the rhetoric. Odd, isn’t it, that the people who understand it best are the least excited to sell it?

It’s almost as though knowing things grants you two perspectives the confident amateur lacks. First, you can see the consequence of a sentence before you say it. You understand the cascade. You know that if you tell a room of CEOs their support staff will be obsolete in two years, those CEOs will begin firing people now—not in two years, not when the technology is ready, but immediately, because the narrative is the permission structure and the layoffs juice the stock whether the technology ever works or not. Second, expertise teaches you that even an infinitesimal chance of being wrong is a career-ending thing to be wrong about. The real expert knows the limits of the model, has met the edge cases, understands that the gap between a demo and a production system is not a gap of optimization but of fundamental capability—and that it may not close on the investors’ timeline, if it closes at all. So the real expert stays quiet, or speaks in careful probabilities, and the resulting vacuum is filled, reliably, by the loudest and least qualified voice available. Nature abhors a vacuum. Marketing adores one.

The executive class carries none of this liability. They can say whatever they like, free of consequence, because they are America’s darlings—exempt from the ordinary laws of cause and effect on the grounds that “sometimes visionaries are simply too genius for physics,” or whatever the PR department issues when the self-driving car is, for the—what is it now, the twentieth?—consecutive year, “about a year away.” For anyone else, a two-decade streak of confidently wrong predictions about your own product would be called misleading investors. For this class, it’s called vision. We have somehow granted a specific cohort of thin-skinned, incurious, perpetually aggrieved men a status wholly immune to accountability, and then we act surprised when they behave accordingly. Why would they stop? Nothing downstream of them ever sends them the bill.

Why, though? Shouldn’t accountability increase as you ascend? When I earned my captain’s license, the deal was this: any collision on the water is partly my fault, always, because there is always something I could have done to avoid it. That’s the standard. That is what responsibility means. As the tech lead on my team, when a decision produces a technical failure, a security incident, an outage—I’m first in line for it. We call it a blameless culture, which is a lovely phrase and partly theater; I am accountable, and I should be. I take seriously the duty to own the failures publicly and share the credit collectively. That’s not heroism. That’s the floor.

But the executive class? Drive the financials off a cliff, lay off a tenth of the staff to cover it, and that’s simply “business today.” When did we agree that this isn’t a catastrophic failure of budgeting, strategy, and direction—of the exact things executives are, definitionally, responsible for? What are they for, if not to prevent precisely this? I suspect the honest answer is that we never decided; we just stopped asking. The accountability gap widened so gradually, over so many decades, that no one clocked the moment it became a canyon. Now the people who crash the company leave by golden parachute, and the people who built it leave with a severance figure that reads like a typo. So: what, exactly, are they for?

As the hype gets screamed down from on high by this same unaccountable class—the same people who will absolutely use AI to justify trimming headcount to cover their own budget overruns—the people they cut will be the experts. And the experts were here all along. They were there when the Challenger launched, warning about the O-rings. They were there keeping the screwworm from marching north into the United States—a quiet, unglamorous, billions-in-averted-damage miracle of public health—right up until the program was reclassified as waste, fraud, or abuse; I would genuinely love to know which of the three a parasite-eradication effort was deemed to be. They were there urging vaccination, and you can measure the results in the measles numbers, assuming we haven’t also defunded the counting. Bird flu: back on the menu. Ebola: fingers crossed. A nastier Lyme variant turning up in New York, severe enough to raise the question of whether it deserves a new name. The experts are always there—standing just off to the side, speaking in careful, measured tones, being comprehensively out-shouted by people who mistake confidence for competence and volume for validity. When has the confident amateur ever once turned out to be the one who was right?

Hold on—let me take a sip of raw milk.

It always takes a catastrophe for people to concede they should have listened. To concede that the people who learned the thing not to make a buck but because it fascinated them—because they were, and I mean this as the highest available compliment, nerds—were the ones who actually knew. They became experts out of a conviction to give something back, not to extract something. And somehow we’ve villainized the curious and canonized the extractors. We built a system in which the people who understand the technology are drowned out by the people positioned to profit from misrepresenting it. We let the invisible man swap expertise for trendslop, accountability for impunity, curiosity for the cold arithmetic of cost reduction. How did the least curious people in the building end up in charge of the most complicated thing in it?

The parlor trick is genuinely impressive. I’m not denying it. The maths are elegant, the scale is staggering, the demos are lovely. But the trick is being sold as a deity, and the people selling it are not the people who built it. They’re the ones who saw it and thought: how do I use this to fire people? How do I use this to stop paying for labor? How do I use this to remove the last remaining friction between me and total extraction? Three questions. Notice which one isn’t among them: does it actually work?

The experts are right outside the door. They’ve been there the whole time. They aren’t hard to find—they’re on arXiv, on the research forums, in the quiet corners of the internet where curiosity still lives. We just don’t listen, presumably because they decline to take a hammer to their own jawline, say something racist for engagement, sell supplements, or preach prosperity gospel—whatever the algorithm’s incentive structure actually rewards. They speak in probabilities and caveats and “further research is needed,” which is not sexy, does not move the stock price, and does not justify a $500 billion capital expenditure on data centers. So we listen to the oligarchs instead, and let them burn the house down with a parlor trick—and, as a bonus, flatten a neighborhood with a data center nodded through by six council members who traded their constituents for a signed photo. Remind me which part of that was the innovation?

The fire is spreading. The experts are still talking. Maybe—just maybe—we could try listening before the roof comes down on all of us. Or we could take another sip of raw milk. Your call. Blink.