Crawl, Walk, Run, Fly

This is where ChatGPT began on November 30, 2022: a “research preview” that could “answer followup questions” inside a conversation and remembered nothing across them. OpenAI’s own description of that era, written three years later, is the best one. Using ChatGPT “has always meant coming to ask a question.”

Crawling is chatting with something that has read the library and never met you. Every session starts from zero. The machine brings a prediction of what people in general want; it has no prediction of you, because you have left it no record. You do the verb. You type every word, you supply every piece of context, and you carry the whole relationship in your own head between sessions.

Most people I meet are still here. I cannot cite a survey for that and have not seen one worth citing, so take it as an observation from a room in Charleston rather than a statistic.

Walking is the same conversation with a memory behind it. The dates matter, because this stage arrived quietly and most people missed the day it became the default.

OpenAI began testing memory on February 13, 2024 with a small share of free and Plus users. It reached all users on September 5, 2024. On April 10, 2025 it began referencing every past conversation rather than a list of saved facts. Anthropic introduced memory to the Claude app on September 11, 2025 for Team and Enterprise plans, extended it to Pro and Max on October 23, and on August 25, 2026 made it on by default and the same memory everywhere the product runs. Google switched on “personal context” for Gemini on August 13, 2025, on by default, and in January 2026 began a beta that reads a user’s Gmail, Photos and YouTube.

Notice what did not change. You still open a window and type. The machine still waits. It knows you, but the sentence is yours, and the machine’s only job is to finish it. The verb is still your verb; it just carries less weight.

What changed underneath is the record. Every conversation you leave behind is an artifact, and artifacts are the only thing that ever trains a prediction. The machine has no other road to knowing you. It cannot be told who you are in a way that sticks; it can only be shown, one exchange at a time, and the showing is slow. This is why the walking stage feels like nothing happened. On any one day, nothing did.

Running is when the sentence changes mood. Interrogative becomes imperative. “What is” becomes “do this.” And with that grammatical shift the verb changes hands: for the first time the machine is the one doing something, and you are the one pointing.

The record here is short and recent. Anthropic shipped computer use on October 22, 2024, Claude “looking at a screen, moving a cursor, clicking buttons, and typing text,” for developers only. OpenAI’s Operator followed on January 23, 2025: “you give it a task and it will execute it.” On July 17, 2025 OpenAI folded Operator into ChatGPT agent, which “thinks and acts.” Anthropic launched Claude Cowork on January 12, 2026 with the plainest description of the stage anyone has written: “Give it a goal, and it works across your files and tools. You come back to polished work for your review.” And four days before this post, on September 16, 2026, Cowork stopped being a separate product and became simply Claude, the delegating mode and the chatting mode in one window.

When delegation stops being a product and becomes a mode, the stage has arrived.

Running also feeds walking. A delegation is a richer artifact than a question. A question shows the machine what you wanted to know; a delegation shows it what you wanted done, how you wanted it done, and what you sent back for another pass. The record thickens faster. Which is the setup for the fourth rung.

Flying is the removal of the sentence. Nothing is typed. The machine’s prediction of you has gotten good enough that it acts on the prediction, and the act is what you would have asked for. The verb has not just changed hands; it has changed kind. You crawl, you walk, you run. You do not fly. You are flown.

Every large company said the word this year. Google, at I/O in May 2026: a Daily Brief that “works overnight for you,” “anticipating your needs and suggesting next steps,” and Gemini Spark, “your personal AI agent… taking action on your behalf and under your direction… 24/7 so you don’t need to keep your laptop open.” Anthropic, on June 23: an “ambient” mode in which Claude “will proactively keep you updated about whatever it thinks you might need to know” and follow up on threads that have gone quiet. Microsoft, at Build on June 2, coined a category name for it: “autopilots,” always-on agents that “work autonomously, with their own identity, and act on your behalf.” That is the word at the top of this essay, and it was not chosen for King’s sake. It was chosen because it is the accurate one.

And OpenAI, first of all of them, on September 25, 2025: “Now ChatGPT can start the conversation.” Pulse read your memory and chat history each night and delivered cards in the morning. It was, in OpenAI’s words, “the first step toward a more useful ChatGPT that proactively brings you what you need.”

Nine months later it was gone. The June 17, 2026 release notes: “Pulse is being sunset as proactive updates move into scheduled tasks.” Scheduled tasks are a fine feature. They are also the third stage. You tell it what to watch and when, and the machine runs on your instruction. That is pointing, not predicting.

So the honest description of the fourth rung in September 2026 is this: announced by everyone, held by no one at consumer scale, and the first serious attempt was withdrawn back to the rung below. I do not think this means flying is not coming. I think it means the stage is harder than the demo, for a reason the next section gives.

I have a framework for this, and the ladder is a clean instance of it. Reality is a ratio: what arrives, divided by what was expected.

And there is a rule with no exceptions: no entity can touch the denominator directly. The only road to a prediction runs through the record. Actions leave artifacts, artifacts become history, and history is what retrains prediction. One way, and slow.

Now put the machine on the other side of the table. A chatbot at the crawling stage has a prediction of everyone and a prediction of no one. It holds no artifacts of you. At the walking stage it has begun to collect them, and its prediction of you is being retrained the only way any prediction is ever retrained, by the record you leave. At the running stage the artifacts get richer, because a delegation records a want and not just a curiosity. And flying is the point where the machine’s prediction of you sits close enough to your own next action that it can move first.

Ask, then, whose Reality changes at the fourth rung. Name the Actualizer: it is you. What arrives is the thing the machine did. What you expected is what you would have asked for. When those match, the ratio reads one, and one, in this framework, is a particular feeling: still present, still conscious, but without the charge of unexpectedness. No AAH. No OO. Unitless stability.

That is the design goal of every flying product, stated in the framework’s terms: drive the surprise in your day toward zero. And that is also why the stage is hard, and why the first attempt was pulled. A machine that gets your next move right removes a surprise. A machine that gets it wrong adds one, and adds it with your name on it: a table booked that you did not want, a reply drafted in a tone you would not use. Below one the ratio is not neutral; it is the OO, and the nervous system prices a miss at roughly twice a hit. The product has to be right far more often than it is wrong just to feel like nothing. Pulse was not a bad idea. It was a machine sending you its predictions of you before its record of you was thick enough to make them, and the users felt every miss.

There is a second thing to say, and I will say it plainly rather than let it sit under the essay. Surprise is where reality is felt. A machine that predicts your next move and makes it is, by construction, a machine that removes surprise from your own life. Whether that is a gift depends entirely on which surprises it removes. The ones that were only friction, the rebooking, the reminder, the reply you would have written the same way, I will give up gladly. The others I would like to keep. The machine cannot tell the difference, because the difference is in the angle, the part perpendicular to feeling, and nobody registers that, including the machine.

Already true

Memory is on by default at all three major labs, with the dates above. Delegation is a mode rather than a product, as of September 16, 2026. Every major lab shipped or announced a proactive feature in 2026, and the first one at consumer scale was withdrawn after nine months.

What has to happen for the fourth rung to hold

A consumer product that acts unprompted survives twelve months without being pulled back into scheduled tasks. Its hit rate clears the bar loss aversion sets, roughly two hits per miss, so that the misses do not dominate the feeling. And the record has to be thick: flying is only reachable by people who have spent real time walking and running, which means the first users to hold the stage will be the ones who delegated most in 2026.

Where I am probably wrong

The verb test may be too clean, and fly may be too strong a word. It claims a change of kind, and it is possible the fourth stage arrives as a change of speed: a machine that runs ahead for one task and hands the verb back, which would make three stages and one feature, and would make the Army’s ladder the right one after all. Or the fourth rung arrives for the company rather than the person. Google’s Daily Brief is a flying product on Google’s side of the table and a walking product on mine, because I still read it and decide. If either holds, the ladder still describes the machine and stops describing you.

Which rung you are on is not a function of which product you pay for. All four are in the same window now. It is a function of what you type. If you type a question, you are crawling or walking. If you type an instruction, you are running. If you type nothing and something arrives, something is flying, and the only question left is who is at the controls.