1,355 Days: What a Thirteen-Year-Old Will Never Do

On November 30, 2022, OpenAI put a text box on the internet. It was framed as a research preview. There was no launch event, no keynote, no ad campaign. Most of the people who would eventually reorganize their working lives around it did not hear about it for months.

That was 1,355 days ago.

Here is the exercise I want to run, and it takes about ten seconds. Take that number and lay it out ahead of us instead of behind us. Count 1,355 days forward from today and you land on May 2, 2030.

Now hold a specific person in mind. A kid who is thirteen right now — seventh or eighth grade, somewhere in America, phone in pocket, no particular opinion about any of this. On the far end of that span, they are seventeen. A high school senior. Old enough to have been shaped by everything in between and young enough that none of it will strike them as remarkable.

Everything that happened between the research preview and this morning is about to happen again, at length.

We are bad at this particular kind of estimate, and the reason is worth naming precisely. It is not that we underrate technology. We overrate it constantly — the flying cars, the paperless office, the year of the desktop Linux. What we get wrong is narrower: we underrate how much ordinary life reorganizes itself during a stretch of time we have already lived through once.

So it helps to look at the tape rather than argue about the future in the abstract.

Notice what is not on that timeline. No new interface. No robot. No general intelligence. Just one capability — generation — getting cheaper, faster, and more embedded, until roughly an eighth of the human species touched it in a given week.

The next span does not need a miracle either. Everything in the five predictions below is already running somewhere at small scale, with a known bottleneck. That is the whole method: find the thing that already works in three cities and ask what it looks like at three hundred.

Forecasting four years out is mostly arithmetic on things that already exist. The surprises come from the ordinary rituals that quietly stop being ordinary.

This is the prediction people push back on hardest, and it is also the one with the most evidence already on the ground.

Start with the fact that the rite of passage was already weakening before any of this. In 1983, 46% of American sixteen-year-olds had a driver’s license. Today it is about 26% — and that entire decline happened without a single robotaxi. It happened because of graduated licensing laws, insurance costs, two working parents, and a generation that discovered its friends were reachable without a car. Worth being precise: the slide bottomed out around 2014 and has ticked up a point or two since. The trend I am extending is not currently in motion. It is a floor that something new is about to push through.

Now add the supply side. Waymo went from about 50,000 paid rides a week in May 2024 to roughly 500,000 a week across ten U.S. cities as of March 2026 — a tenfold increase in under two years. Its co-CEO has stated a 2026 goal of a million weekly rides and service in more than twenty cities.

Then look at what that does to a business almost nobody thinks of as a technology business. GEICO’s annual advertising spend has run between roughly $0.8 billion and $2 billion depending on the year. Progressive now outspends it substantially — around $3.5 billion in 2024, and higher through 2025. In 2024, four auto insurers sat inside the fifteen largest advertising spenders in the United States, together putting more than $7.5 billion behind a single decision: which company a private individual picks to insure a car they personally drive.

A seventeen-year-old who does not own a car and does not drive is not in that market. There is no ad to show him. The most heavily advertised category of his parents’ adult life is simply absent from his.

Already true

Only about a quarter of sixteen-year-olds are licensed. Driverless rides grew roughly tenfold in under two years and now operate as paid commercial service in ten-plus metros.

What has to happen

Coverage has to go from a handful of Sun Belt metros to most suburban America, and per-mile cost has to land under the total cost of teenage car ownership — a bar set as much by insurance premiums for a sixteen-year-old as by the vehicle.

Where I am probably wrong

Geography. This kid lives in Phoenix, Austin, or Atlanta. In rural Wyoming he gets a license at fifteen and drives a truck, exactly as his father did. The honest version of the prediction is not that licensing ends — it is that it stops being universal and becomes optional, like knowing how to drive a manual transmission.

In May 2026, Nature Medicine published a randomized, blinded evaluation of a multimodal version of AMIE, Google’s diagnostic dialogue system. It was compared against 19 board-certified primary care physicians across 105 clinical scenarios — 210 simulated telehealth consultations in total, with trained patient-actors and 18 specialist physicians grading the transcripts. AMIE interpreted images, ECGs, and clinical documents inside live diagnostic conversation, and came out ahead of the physicians on most measures of diagnostic accuracy, communication, and empathy.

I want to be careful here, because this is exactly the kind of result that gets oversold, and I would rather do the deflating myself.

Those were simulated consultations, and the authors say plainly that real-world clinical validation has not happened. The physicians were working through a text interface most of them do not practice in, which the paper itself flags as a limitation. The underlying work had been circulating as a preprint since May 2025, so this is a year-old result finishing peer review rather than a bolt from the sky. And a model that performs well against a scripted patient is not a model that has been trusted with a scared one.

But the direction is not really in doubt anymore, and the interesting question has moved. It is no longer can it. It is who is permitted to act on it, and how long that permission takes.

The capability will arrive years before the permission does. So the kid will not wait for the permission. She will route around it.

That is the actual prediction, and it is less about medicine than about sequencing. Diagnosis is licensed, liability-bound, and reimbursed through machinery that moves on legislative time. None of that binds a seventeen-year-old with a phone. She will get a read in four minutes, and then she will spend three weeks getting an appointment to have a licensed human agree with it.

Already true

A blinded, randomized study across 105 scenarios and 19 physicians put the system ahead on most diagnostic and communication measures. Millions of people already describe symptoms to a chatbot before they call anyone.

What has to happen

Not much, technically. The load-bearing changes are malpractice law, licensure, and reimbursement — whether a payer will fund a visit that a model has already resolved.

Where I am probably wrong

Confident wrong answers scale as fast as right ones, and a single well-publicized death changes the regulatory climate overnight. The waiting room may well outlive the need for it — institutions rarely die on the schedule their obsolescence suggests.

In January 2026, an IBM Institute for Business Value study with the National Retail Federation found that 45% of consumers already turn to AI somewhere in their buying journey. A year earlier, Capgemini’s consumer research had found 58% saying they had replaced traditional search engines with generative tools for product recommendations — up from 25% in 2023. McKinsey puts the global agentic commerce opportunity at three to five trillion dollars by 2030. Standards work is underway to let agents transact directly against merchant systems.

Strip the vocabulary away and one thing is happening: the entity being persuaded is changing species.

Consumer advertising is a machine for winning a moment of human attention and converting it into preference. It is enormous — a substantial share of the internet exists to serve it. But it is aimed at a creature with limited attention, emotional associations, and a fear of choosing wrong. An agent has none of those. You cannot make an agent feel aspirational about a shoe.

• the jingle
• the celebrity endorsement
• the sponsored feed placement
• the loyalty program
• the limited-time offer
• the abandoned-cart email
• the influencer code
• the brand mascot

Each of those is a technique for moving a person. Each is inert against a procurement process.

Here is the part I would not celebrate. Advertising has survived every previous obituary, and it will survive this one by relocating. If the agent chooses, then the money goes to influencing the agent — ranking, catalog data, protocol compliance, whatever quiet placement fee turns out to be purchasable. That is not the end of persuasion. It is persuasion moved somewhere the kid cannot see it, cannot recognize it, and has no cultural immune system against it. A GEICO ad at least announces itself.

Already true

45% of consumers use AI somewhere in the buying journey (IBM/NRF, 2026); a majority report substituting generative tools for search when looking for products (Capgemini, 2025).

What has to happen

Agents need durable payment authority and enough trust to buy without confirmation. That is a fraud and liability problem more than a model problem.

Where I am probably wrong

People may simply enjoy shopping. Delegation is efficient and browsing is pleasurable, and pleasure has beaten efficiency in retail before. The likeliest outcome is a split: agents take the boring purchases, humans keep the fun ones, and advertising follows the fun.

This is the prediction I hold with the most concern and the least comfort.

Stanford’s Digital Economy Lab, working with ADP payroll records covering millions of workers, has been tracking one number since August 2025: the employment gap between the youngest workers in the most AI-exposed occupations and everyone older doing the same work. When Erik Brynjolfsson and his colleagues first published it, that relative decline for twenty-two to twenty-five-year-olds was 13%. Their February 2026 update put it at 16%. As of this month it is 19%, and it has widened at every revision.

Employment for experienced workers in those same occupations has held steady or grown over the identical period. Young software developers are the sharpest case: employment in that group is down roughly 20% from its late-2022 peak.

Students have noticed. A Gallup and Lumina Foundation survey of 3,801 college students found 47% had seriously considered changing majors because of what AI might do to the job market, and 16% already had. Unemployment among recent graduates sat at 5.6% in March 2026.

One honest complication on that last figure: economists at the New York Fed attribute roughly two-thirds of the rise in young-graduate unemployment since 2019 not to AI but to remote work — fewer junior roles because fewer people are in a building where a junior person can be useful. I do not think that undercuts the argument. I think it describes a second mechanism arriving at the same door.

The thing being removed is not jobs in the aggregate. It is the apprenticeship — the arrangement where a junior person is paid slightly more than they are worth, for a couple of years, in exchange for becoming senior. Every profession runs on it. None of them have another mechanism.

You cannot have senior people without a way to make them.

So by 2030 either that mechanism gets rebuilt into something else — simulation, credential, portfolio, some formalized version of the apprenticeship that firms fund because the alternative is having no successors — or a set of professions quietly stops replacing itself. I do not know which. I am fairly confident the kid will not be handed the version his parents got.

And in fairness: not everyone reads this data the way I just did. There is a real argument that the class of 2026 is struggling primarily because of interest rates, a hiring hangover, and a cohort that grew faster than the roles did. That case deserves a hearing, and the remote-work finding above is evidence for it. My response is only that the age gradient inside the same occupations is hard to explain that way, and that it has widened at every revision rather than reverting.

This is the softest of the five, and I would put the lowest confidence on it.

What exists today: 1X opened consumer preorders for its NEO humanoid on October 28, 2025, at $20,000 outright or $499 a month, with first U.S. home deliveries slated for 2026. The company has not published an autonomy figure; secondary estimates put it somewhere around 60–70% at launch. What 1X does describe openly is the mechanism — a set of tasks the robot handles on its own, and an “Expert Mode” in which a human operator at the company puts on a headset, takes remote control, completes the job, and lets the robot learn from the demonstration.

Read that last sentence again, because it contains both the bull case and the thing that should bother you.

The bull case is that teleoperation is a data-collection strategy. Every remote takeover is a labeled example of a physical task performed correctly in a real home. That is precisely the input the field has been starved of, and it is the mechanism by which the autonomy number climbs. Four years of that is a lot of demonstrations.

The thing that should bother you is that during those four years, a stranger can see inside your house. Not metaphorically. A person at a company, wearing a headset, looking at your kitchen while your kid does homework at the table.

A thirteen-year-old today will grow up with that as the baseline condition of a helpful machine, and will find our discomfort with it quaint.

Already true

A consumer humanoid is priced, ordered, and shipping to homes — not a concept video. A substantial share of its tasks currently need a remote human, by the manufacturer’s own design.

What has to happen

The teleoperation share has to fall toward zero and unit cost has to drop by roughly an order of magnitude. Both are manufacturing and data problems, which historically move faster than people expect and slower than founders promise.

Where I am probably wrong

Homes are the hardest environment in robotics — unstructured, cluttered, full of unforgiving surfaces and living things. This could still be a rich person’s novelty in 2030. The narrow-purpose machine that does exactly one chore well may beat the humanoid that does twelve poorly.

Let me be exact, because forecasts like this are usually written to be impressive rather than checkable.

1. I am not claiming all five land

   Some of these will look silly in 2030. That is the arrangement. A forecast with no failure modes is a marketing document. Every one above ships with the condition that would sink it, stated in advance and in my own words.

2. The direction is more reliable than the timing

   Nearly every miss in technology forecasting is a timing miss, not a direction miss. Autonomous vehicles were four years away for fifteen years, and then were carrying half a million paying passengers a week. Assume the arrow is right and the date is soft.

3. The real claim is about the count, not the contents

   The load-bearing prediction is not any single item. It is that the number of ordinary rituals which quietly stop being ordinary over 1,355 days is larger than anyone budgets for. In November 2022 nobody had “a third of the planet talks to a machine weekly” on their list either.

4. None of this says whether it is good

   A world with fewer teenage drivers has fewer teenage funerals, and also fewer sixteen-year-olds who have ever been alone and in charge of something. A world without entry-level work is a world without a certain kind of humiliation, and without a certain kind of education. I am not going to pretend these trade evenly.

The last thing, and it is the one I keep circling back to.

The kid is not going to notice. Not the DMV he never visits, not the insurance ads he never sees, not the apprenticeship he was never offered. You cannot perceive the absence of a ritual you were never handed. He will simply arrive at seventeen with a different set of defaults, and if you ask him about any of it he will look at you the way we look at someone describing a rotary phone.

That is what an era change feels like from the inside. It does not feel like anything.

We happen to be standing at the midpoint, with both spans visible at once. That is a genuinely unusual vantage point, and it will not last. It is worth using while we have it.