The Slow Part is the Interesting Part

Sam Altman has an agent on his computer that can do most of his day for him, and he still gets through his inbox by hand.

He raised it himself, unprompted and at length, in a conversation with David Senra published on 24 August. Twenty years of the same habits: clicking around, pasting between messaging apps, scanning email for whichever message will hurt least to open, keeping a to-do list. He has Codex. So does everyone. By his own account he knows there is a faster way and does it the old way anyway, and when he reaches for an explanation he lands on something more useful than self-criticism. "I think this is mostly a product failure."

That sentence is the most load-bearing thing in ninety minutes of conversation, and it points at where the next several years of value actually get built. Not in the models. In the twelve inches between a capable system and a person who hasn't changed how they work.

Altman is describing a gap he already tried to close and got wrong once. After GPT-4 arrived in 2023 he expected software businesses to be up for grabs almost immediately. They weren't. His diagnosis now is blunt: "the economy just has so much inertia." People keep buying from the same companies and using their tools the same way. His conclusion, applied to himself as much as anyone: "we've all been too ambitious on timelines."

What makes this worth reading closely is that he doesn't file it as a disappointment. He files it as good news. The transition, in his framing, will go "smoother and slower," and he says plainly that he is grateful for it. Human societies have some built-in resistance to rapid change, he argues, and that resistance is probably a feature.

What a missing product looks like

The comparison he reaches for is smartphones before the iPhone. He was an early adopter, had a Palm Treo in 2003 or 2004, and remembers that most of the technology was already sitting there. Multi-touch was missing. Mostly what was missing was the product idea that made a phone into a different object.

Read the present through that lens and the current moment looks less like a plateau and more like a staging area. The pieces exist. Models that hold a hundred thousand tokens of context, agents that plan and use tools, inference cheap enough to run continuously. What hasn't arrived is the thing that makes a person put down the old workflow without being persuaded.

Senra supplies the historical rhyme. Larry Ellison in the 1980s, telling his people this was never a software problem, that they could install the software and watch customers not use it. Netflix shipping DVDs while Senra drove past a Blockbuster still doing business on his way to school. Both are stories about a technology that had already won and a habit that hadn't yet noticed.

The AI Gilded Age has the same shape as the first one. Electricity was demonstrably better than steam decades before most factories were rewired, because rewiring a factory meant rethinking what a factory was for. The tracks get laid years before the traffic arrives. That interval is not dead time. It's when the thing that runs on the tracks gets designed.

The constraint Altman names next

Ask what closes the gap and he answers with something more specific than better models.

"I feel more limited at this point by the amount of useful context AI has on me."

He is explicit about what he wants and why he can't have it yet. He is not going to read every post on OpenAI's internal Slack. He is not going to read every account a customer has of where ChatGPT worked or failed. He could read more research papers than he does, and doesn't, because it costs mental energy he spends elsewhere. What he wants is a system that reads all of it and brings the relevant part to bear at the moment a decision is being made.

His assessment of where the field's attention has gone is a criticism of his own company as much as anyone's: the industry focused correctly on model intelligence, he says, and has not yet thought hard enough about what it means to give a model more context than any person could hold. He describes the field as "on the precipice" of a different way of working, on an axis where humans simply cannot compete. No one reads tens of thousands of pages in seconds and uses them accurately.

Senra, unprompted, produces the working prototype. He has kept every note and highlight from every book he has read since 2018, added the transcripts of his own episodes, and queries the lot while making a show. He describes asking what Bob Noyce said about a thing, or what Rockefeller did about a thing, from a book he read seven years ago and no longer remembers. He uses it every day.

That is the shape of the product that hasn't been built at scale. One person, hand-assembling a corpus, getting compounding returns from it. Multiply it by the number of people who have accumulated a decade of their own work in scattered files and the size of the opening is visible.

What it looks like when it lands

Take Altman's claims at face value for a moment and run them forward. This is scenario, not reporting.

By 2029, a person starting a company does not begin by hiring. They begin by pointing a system at everything they have ever written, every contract in the industry they can lawfully read, every filing their competitors have made, and asking what has already been tried. The founder-shaped work — the judgement about what is worth doing — stays with the person. The apprenticeship that used to precede it does not.

Altman's version of this is a prediction he states without hedging: "the greatest boom in people starting smaller businesses that we have ever seen." His reasoning is that starting a small company has historically required a quantity of privilege, luck and resources that screened most people out, and that the screen is coming down. He couples the prediction to an admission that the field, OpenAI included, has neither talked about this enough nor built enough products to accelerate it.

The strategy he describes is built for that world. OpenAI as a platform rather than a product company: one direct interface to a capable system, one API underneath it, sold at every point on the cost-performance curve. He wants a hundred million new businesses and eight billion people building on it, and says explicitly that OpenAI should not try to occupy every product category or compete with its own customers. Two of last year's casualties illustrate the discipline. Sora — a "good product and fun and cool," in his description — consumed compute that went to Codex instead, and was shut down in March 2026, according to reporting at the time. Atlas, which he calls the best web browser, lost a talent argument rather than a compute one; as TechTimes reported, its browsing capability migrated into ChatGPT and the shell was deprecated on 9 August.

Where the disagreement is

Tobi Lütke told Senra that 2026 would be the year every business was up for grabs, that someone would build the AI-native version of Shopify, and that it would be him — rebuilding it himself at night, from scratch, with current tools.

Altman disagrees, and says so with the timeline named. Not every business, first of all: he expects some categories to get harder to compete with as AI improves, because people will pay more for authentic non-technological experiences and care more about the things that are irreducibly human. And on the schedule, he thinks many software businesses genuinely are up for grabs, and that it takes longer than Lütke's calendar allows.

The interesting thing about that disagreement is that it barely changes what either man does next. Lütke is rebuilding at night. Altman is buying compute and hunting for the interface. Both are behaving as though the outcome is settled and only the date is open.

Why the slowness is the opportunity

There is a version of this story where the inertia is the problem and the industry's job is to overcome it. Altman's account suggests the opposite reading.

A fast transition rewards whoever ships first and punishes everyone still learning. A slow one rewards whoever is still standing when the interface finally arrives, and gives the people whose work is about to change time to change it on their own terms. It also gives the safety practice time to mature, which is why his FAA comparison is more than an analogy: robust accident reporting, clear-eyed postmortems, no handwaving over the awkward parts, and the findings shared with other labs rather than kept. That discipline is only buildable in an industry that has years rather than months.

He is candid about what he thinks the two real risks are. Loss of control, and power concentrating in one company, one model or one person. His stated position on the second is that people should be deeply in control of the future, and he is scathing about the alternative pitch he hears in his own field: a cure for cancer and material comfort, offered in exchange for autonomy and any say in what comes next. He calls it a terrible sales pitch. It is also the pitch that becomes plausible if the transition is fast enough that nobody has time to argue.

Ten years ago, roughly a dozen people walked into Greg Brockman's apartment on the first working day of 2016 to start an AGI lab, looked around the room, sent someone to find a whiteboard, and discovered that none of them knew what to do next. Altman remembers the energy in the room collapsing. What followed, in his phrase, was years of chaotic stumbling that eventually produced the scaling laws and the research path to GPT.

The whiteboard-and-no-idea phase is where the interface sits right now. Somewhere a person is building the thing that makes everyone stop doing their email by hand, and when it arrives it will look obvious, and the years it took to find will look like the point.