There are two AI markets hiding inside the same conversation. One makes things for people to look at. The other does work that people may never see.
I think the difference explains how AI can be both a genuine technological revolution and a speculative bubble.
To paraphrase a question from software developer and T3 Chat founder Theo Browne: What is the value of tokens spent on entertainment that nobody watches, compared with tokens spent on engineering that nobody directly sees?
A finished movie that nobody watches delivers no entertainment. But an optimization that nobody reads can still reduce the cost of running a factory. The absence of an audience means something very different in those two cases.
You cannot watch exponentially faster
Tokens meter language-model input and output, not computation in general. Here I’m using the term loosely for the generated work, including work behind images and video.
On the mainstream entertainment side, the product ultimately needs human attention. Someone must read the story, watch the video, play the game, or have the conversation.
AI can make those experiences cheaper, better, and more personal. That is real value. But generating a thousand times as many movies does not create a thousand times as many evenings in which to watch them.
Eventually, the constraint stops being our ability to produce entertainment and becomes our ability to consume it. The system can manufacture more supply without manufacturing more customers, more hours, or a corresponding willingness to pay.
This does not mean entertainment compute cannot grow. A better film might require much more computation per minute. A system might generate and discard hundreds of candidates to find something worth showing. Interactive worlds could be enormously compute-intensive.
But all that extra work still has to justify itself through the experience someone actually receives. More generation is not automatically more utility, and more utility is not automatically more revenue.
Exponential production capacity does not establish exponential demand.
Work does not need an audience
Engineering has a different relationship with human attention.
Imagine an AI system investigating a software performance problem. It proposes changes, runs benchmarks, rejects unsuccessful approaches, tests edge cases, and eventually produces a patch with evidence that it improves performance without breaking required behavior.
The developer needs to review the result and the evidence. The developer does not necessarily need to read every discarded attempt.
The intermediate work can expand much faster than the final explanation.
The same principle applies wherever useful work can be evaluated without requiring a person to inspect every step: searching possible designs, checking proofs, optimizing schedules, or screening candidates for physical experiments.
None of this makes engineering limitless. Tests can miss defects. Simulations can be wrong. Laboratories, factories, energy supplies, and human judgment remain bottlenecks. A million plausible answers are not a substitute for one answer that survives contact with reality.
But human attention need not be the bottleneck for every additional unit of work. That creates much more room for useful computation to grow.
There is also the possibility of compounding. A better tool can make the next task cheaper. A better chip can support more computation. An improved research process can make later discoveries easier.
Engineering tokens can help build the capacity to do more engineering.
That is a reason to expect sustained growth, not a promise that every additional token will pay for itself.
A fuzzy boundary, a sharp economic difference
“Mainstream” and “engineering” are imperfect labels. A household assistant that resolves a scheduling conflict is doing instrumental work for an ordinary consumer. A game contains both entertainment and an enormous amount of engineering.
Conversely, an AI-generated technical report may accomplish nothing except adding another document to somebody’s reading queue. Calling it engineering does not make it productive.
The distinction is not between serious people and people having fun. Entertainment is valuable, and engineering can be wasteful.
The distinction is where the value appears: in consuming the output, or in what the work accomplishes.
Most people will encounter the softer, more familiar distinction. They will see generated pictures, songs, videos, and conversations. They will judge AI by those experiences because those are the experiences available to judge.
A million unsuccessful design candidates that lead to a better component are harder to put in a demo reel.
Where I think the bubble is forming
The speculative mistake is to borrow the growth potential of engineering and apply it indiscriminately to consumer entertainment.
AI gets more capable. Generation gets cheaper. Therefore, the argument goes, demand for generated content must expand enough to justify the investment.
That last step does not follow.
An entertainment product can be technically astonishing and economically ordinary. It can displace an existing product without greatly enlarging the total market. It can attract millions of users who enjoy it but will not pay enough to support its costs.
Meanwhile, a much less impressive-looking system could justify substantial spending by reliably completing valuable work.
This is not proof that any particular company is overvalued. Nor does “engineering” confer immunity from speculation. Productive applications can arrive too slowly, cost too much, or create savings that flow to customers rather than profits for their suppliers.
But it identifies a dangerous mismatch: the most visible evidence of AI progress may not be the best evidence for the economics used to finance it.
The Internet did not stop being useful
The dot-com comparison is helpful, provided we distinguish a technology from its stock prices.
We might watch out for the Pets.com equivalent in AI. But even that comparison deserves care. Selling pet supplies online was not an absurd idea; Chewy later built a substantial business doing it. Pets.com faced reluctant online shoppers, expensive shipping, and heavy discounting. Being too early is part of the story, though timing alone does not explain the failure.
That makes the warning more useful: a company can anticipate real demand and still fail before it can serve that demand sustainably. Some AI businesses may be pointing toward a real future that arrives too late for them.
As The Big Short dramatizes, being too early can amount to being wrong from an investor’s point of view. If the money runs out before the thesis pays off, eventual vindication does not recover the investment. The future has to arrive while you can still afford to wait.
Video games had their own version in Atari’s E.T., which became a symbol of the 1983 North American console crash. One game did not cause the collapse, and the collapse did not mean people were finished with video games. An AI “E.T. moment” could similarly become shorthand for an industry’s excesses without telling us how much lasting value remains—including in entertainment.
Internet commerce continued growing through the wreckage. In the fourth quarter of 2002, U.S. online retail sales were 28.2% higher than a year earlier, according to Commerce Department figures reported by WIRED.
People were still buying things while Internet investments were collapsing. The useful activity did not wait for the financial story to recover.
That is how I understand the eventual recovery: useful businesses and applications kept growing until they mattered more than the failed promises surrounding them. Not every survivor had always been profitable, and not every failed company had been useless. But the underlying usefulness of connecting people, information, and commerce had not vanished with the valuations.
The useful parts eventually outgrew the fluff.
AI could follow a similar path. Disappointment with synthetic entertainment would not establish that AI had exhausted its potential. It could instead expose the difference between producing more things to consume and accomplishing more work.
The AI market I expect to matter most over time is the one that can use increasing amounts of computation without demanding corresponding increases in human attention.
There may be a bubble in the things we are being shown, even while a revolution develops in the work we never have to see.