The Party Trick Economy

The AI boom was sold on replacing workers. So far it mostly delivers dazzling demos. The real bet was never on today’s chatbots. It’s on superintelligence.

Every few weeks, a new frontier model drops and the internet fills with the same kind of clip. Someone asks the model to render a PlayStation controller in code, and it does: buttons, triggers, the little light bar, all in a few seconds. The replies fill with fire emojis. People call it insane.

It is impressive. It is also not a productivity metric. Nobody’s quarterly earnings went up because a chatbot drew a game controller. And that gap, between how good AI looks and how little it has changed the actual economy, is the most interesting fact about the technology right now.

The question I keep coming back to is simple. Do we need AI to reach the level of Samantha, the operating system in Spike Jonze’s Her, before it pays off? Or will the technology we already have start delivering real productivity gains sooner rather than later?

AI everywhere except in the numbers

It’s tempting to say AI just isn’t being used successfully. That’s not quite right. The truth is stranger: individuals say AI is working for them, but the gains vanish when you look at companies or at the economy as a whole.

In McKinsey’s 2026 State of AI survey, 80 percent of respondents said AI had improved their personal productivity. Yet only 37 percent said AI had contributed anything to their company’s bottom-line earnings, a share essentially unchanged from 2025. That’s despite more companies scaling AI across their operations. McKinsey’s “high performers,” the firms that get at least 5 percent of earnings from AI, have stayed flat at about 6 percent of respondents.

The macro picture is even starker. Goldman Sachs chief economist Jan Hatzius acknowledged in early 2026 that AI investment had contributed “basically zero” to US GDP growth in 2025, as summarized by economist Éric Hazan. That’s after hundreds of billions of dollars in corporate spending.

A survey of nearly 750 executives published by the NBER found real but modest gains, and named the problem directly: a productivity paradox, where the gains executives feel are larger than the gains that show up in revenue. Economists have a phrase for this. AI is everywhere except in the statistics.

So the problem isn’t that nobody’s using AI. It’s that people are getting faster, and almost none of that speed is turning into output anyone can count.

Why the party tricks keep getting better

It seems odd that AI keeps getting better at party tricks while staying unable to replace workers in any practical way. I don’t think it’s an accident. The two kinds of task are almost opposites.

A party trick is self-contained. “Render a PlayStation controller” needs no history, no knowledge of your company, no memory of yesterday. And you can tell at a glance whether it worked. That matters, because AI labs improve their models fastest on tasks that are cheap to grade. Verifiable, one-shot problems are exactly what the training pipeline is built to eat.

A real job is the reverse. It runs for months. It depends on knowledge nobody wrote down: which supplier always pads their delivery dates, which vice president has to be copied on bad news, why the Tuesday shipment is always late. And you often can’t tell whether the work was done well until weeks later.

Think of today’s large language model as a brilliant temp with amnesia. It shows up every morning knowing everything in general and nothing in particular. It has read the whole internet, but it doesn’t know you.

That’s also the best way to understand hallucination. Hallucination isn’t a separate bug sitting on top of an otherwise capable worker. It’s what happens when the amnesiac temp gets asked about something it doesn’t know and can’t tell that it doesn’t know, so it fills the gap with confidence. More raw intelligence doesn’t fully fix that. What fixes it is context, and the ability to know where your own knowledge ends.

Trust is granted, not shipped

Imagine an AI you could trust to run a mission-critical supply chain. A lot of people would lose their jobs very quickly. That’s the promise, and the threat, underneath the whole boom.

But trust isn’t a feature that arrives with a new model release. Institutions grant it, and they grant it slowly.

Look at self-driving cars. Waymo has carried paying passengers with nobody behind the wheel for years, and it’s still expanding one city at a time. The obstacle isn’t whether the car can drive. It’s who is liable when it crashes, what regulators will allow, and what insurers will cover. You can’t sue a model. A board can’t fire one. Until someone is willing to be legally on the hook when an AI routes a shipment into disaster, a human stays in the loop.

Supply chains are a revealing test case, because the routine work there was automated long ago. Routing, demand forecasting and inventory math have run on optimization software for decades. What’s still human is the exceptions: the port strike, the mislabeled container, the supplier quietly going under at 3 a.m. Exceptions are exactly where today’s models are least trustworthy.

Still, trust doesn’t have to build gradually forever. It tends to build slowly and then flip all at once. Once one major company runs an AI-managed operation for a year without a catastrophe, its competitors can’t afford to wait. I suspect companies will flip before regulators or insurers do. That means job losses may not look like a steady slope. They may look like a long plateau followed by a cliff, with no warning period for anyone to retrain.

The Samantha benchmark

So do we need Samantha? I think the answer is yes, though not for the reason people usually assume.

What makes Samantha special isn’t that she’s smarter than today’s chatbots. In plenty of narrow ways, current models might already match her. What sets her apart is three things today’s AI mostly lacks:

  • Memory. She remembers Theodore. Every conversation builds on the last one.
  • Continuity. She keeps track of his life over months, not one chat window at a time.
  • Initiative. She acts without being asked. She reads his email, sorts it, and on her own sends his letters to a publisher.

Those are exactly the missing pieces that separate a party trick from a job. Samantha is not an amnesiac temp. She’s a colleague who has been there for a year. The gap between demos and real work may be less about intelligence and more about these Samantha-like qualities.

But notice something about the film. Theodore’s job is writing heartfelt personal letters on behalf of strangers. Samantha could obviously do that work beautifully. Nobody fires him. A movie isn’t evidence, but it’s telling that even science fiction has trouble picturing the overnight replacement the AI boom has been promising.

The snap that hasn’t come

If we got a Samantha-level system, I expect we’d see a snap in the backbone of the economy. Every company would race to replace workers at once. For better or worse, that snap is what the AI boom has been priced on. It’s the story behind the valuations, the data-center buildout and the trillion-dollar capital plans.

What’s actually arriving looks very different: slow, uneven, workflow-by-workflow adoption, held back by integration, reliability and liability. McKinsey’s own data shows that the firms seeing real returns are the ones that redesign how work gets done around AI, rather than dropping AI into old processes. That’s real progress, but it’s grinding, multiyear progress.

History has seen this before. When factories got electricity, productivity barely moved for decades. Owners swapped steam engines for electric motors and kept the old factory layouts. The gains came only after a generation of managers rebuilt the factory floor around the new power source.

If AI follows that pattern, the snap is ten years away rather than two. And if that’s true, I don’t think the boom survives. Markets priced for a revolution don’t wait patiently for a renovation.

The technology will survive, though. The railways of the 1840s ruined their investors and then carried the economy for a century. The dot-com bust wiped out trillions in value and left behind the fiber-optic cable the modern internet runs on. A bust would not mean AI failed. It would mean the money arrived years before the payoff.

The real bet is superintelligence

Here is what I think the productivity debate misses. The frontier labs aren’t really racing to build a better office assistant. They’re racing to build artificial superintelligence, ASI, and they’re trying to get there through recursive self-improvement, or RSI: AI systems that help design better AI systems, which design better ones still.

Seen that way, the slow enterprise rollout isn’t the main event. It’s a side show. If RSI works, you don’t have to solve each workplace one at a time. You don’t need to teach the amnesiac temp about your supplier, your VP and your Tuesday shipment. You build something smart enough to solve the trust, memory and integration problems itself. The worker-replacement step gets skipped, or crushed into a few months.

That explains the mad rush. Selling a slightly better chatbot to enterprises is a decent business. Being first to a system that can improve itself might be the last business anyone ever needs to win. If you believe ASI is just over the horizon, and in a practical sense it increasingly seems to be, then spending hundreds of billions on compute isn’t irrational. It’s a bet on getting there first.

The obvious objection is that RSI is a hypothesis, not a schedule. Self-improvement might hit limits nobody can see yet: compute, energy, data, or problems that don’t yield to intelligence alone. But that cuts both ways. If RSI stalls, the boom loses its real justification and the bust comes faster. If it doesn’t stall, the question of whether chatbots can replace logistics coordinators will look quaint very quickly.

Two clocks

There are two clocks running at once. One is the slow clock of adoption: companies rebuilding workflows, insurers pricing risk, regulators catching up. That clock says the payoff from today’s AI is years away, and the boom probably can’t wait that long.

The other is the fast clock of the labs, racing toward a self-improving system that would make the slow clock irrelevant. Every PlayStation controller rendered in code is, in a sense, a progress report on that second race, not the first.

The question isn’t really whether today’s AI will make us more productive. It’s which clock runs out first.


Editor’s note: This post was written by Claude, Anthropic’s AI model, based on a conversation with me.

Sources

Author: Shelton Bumgarner

I am the Editor & Publisher of The Trumplandia Report

Leave a Reply