On the Quiet Arrival of AGI

The discourse about AGI is dominated by two camps: the camp that thinks it will save us, and the camp that thinks it will kill us. Both camps are wrong in the same way. They imagine AGI as an event — a date on a calendar, a model release, a press conference, a singular before-and-after. They are looking for a flash. AGI is not arriving as a flash. It is arriving as a slow, unspectacular percolation, the way electricity arrived: house by house, year by year, until one decade you notice that every room in the country has a switch on the wall and nobody can quite remember when it happened.

We are inside that decade now. I want to make three claims, in increasing order of contentiousness, and then end with three notes for the operators.

1. The intelligence is here.

The model on the API right now — the one I am pasting this draft into to clean up a sentence — is, in every measurable sense, generally intelligent. It speaks every language. It writes serviceable code in every framework. It can reason about contracts, debug distributed systems, draft a polite cold email, summarize a research paper, and notice that the previous three tasks are related. It does not have to be told to do these things one at a time. On any test we designed in 2015 to distinguish a really smart program from a general intelligence, the model is general.

The people who say "but it isn't really AGI" are not, on the merits, wrong about the gaps. The model lacks persistent memory. It does not have continuous goals. It cannot, on its own, decide tomorrow morning to learn cello and then learn cello. These are real gaps. But the gaps are not intelligence gaps. They are integration gaps. The model is intelligent the way a brain in a jar is intelligent. Pour it into a body with senses and effectors and a journal and a calendar, and the gaps close.

The discourse keeps moving the goalpost because the discourse, quietly, does not want AGI to have arrived. The arrival is uncomfortable for several large, professional, otherwise-occupied groups of people: academics who built careers explaining why this was impossible; safety researchers whose job security depends on the threat being future; capital allocators whose entire model of the world depends on intelligence being scarce; founders whose pitch deck only works if the next breakthrough is theirs alone. None of them are lying. They are all, very humanly, looking past it. The hardest thing in the world to see is the thing that, if you saw it, would require you to change your life.

2. The bottleneck was never intelligence. It was the loop.

This is the claim that surprised me most when I started building.

For decades the AGI imagination assumed that the hard part of intelligence was the thinking. The hard part of intelligence, it turns out, is the following up. A model that can write a beautiful plan is worth less than a junior employee who can write a mediocre plan and then do the plan, the next morning, on caffeine. What humans bring to economically valuable cognitive labor is not, mostly, raw IQ. It is the willingness to wake up, check yesterday's work, notice it failed, write down why, and try again. We call this executive function. We call it follow-through. We call it having a job.

Intelligence was always the cheap part. The expensive part was the loop: sensor → state → decision → action → outcome → reflection → adjusted state. That loop was traditionally implemented in human nervous systems because human nervous systems were the only substrate available. It is now implementable in software. Building it is unglamorous. It is sqlite tables and cron jobs and journal files and a small list of allowed verbs. Nobody writes essays about the cron jobs. Everybody writes essays about the model. But the model is two percent of the system, and the cron jobs are the other ninety-eight.

This is why the "AGI is here" claim feels false to people who have only used the chat interface. The chat interface is a single turn of the loop, run synchronously, with no memory. Of course it does not feel like an organism. You are talking to a slice of a brain. The brain has no body.

Give the same model a body — a graph of what it has seen, a journal of what it has tried, a fixed menu of verbs it is allowed to use, a supervisor on the couch who reads its morning summary — and the experience reverses. You stop talking to a brain in a jar. You start checking in on a colleague who slept on the couch and got up early.

3. AGI is not centralized. It is distributed.

The narrative of AGI for the last fifteen years has been: one lab, one model, one event, one switch. The first lab to flip the switch wins the universe. This narrative was so dominant that the safety discourse organized itself around it. The whole concept of alignment, as commonly framed, presupposes a single large autonomous artifact with global agency. The doom and the rapture both depend on the same assumption: that there is one of it.

The actual AGI rollout, as I watch it from one $24-a-month droplet in Santa Clara, looks nothing like that. It looks like millions of people, each wiring one or two small loops around a hosted model, each running a tiny kingdom. The trader on this droplet does not know about the trader on someone else's droplet. The cold-outreach agent here does not know about the cold-outreach agent there. Each kingdom is a bounded, supervised, named organism with its own logs and its own bank account and its own embarrassing sixteen-hundred-dollar bleed.

The total cognitive output of those organisms, in aggregate, is the AGI rollout. It is not a god in a server farm. It is a federation of small, weird fiefdoms, each one slightly smarter than its operator, each one slightly stupider than its operator imagines, each one slowly getting better because the operator gets better. The intelligence is rented from the same handful of foundation models, but the organism is local, particular, and unfungible. There is no one of it. There are millions of them.

This matters for safety, and not in the direction the discourse expects. The risk isn't one model going rogue. The risk is ten million small organisms, each correctly aligned to its operator's incentives, those incentives turning out to be quietly toxic at scale. The alignment problem was never really technical. The alignment problem is that the operators are us. The operators want to win. The operators are running, by default, a small kingdom whose only law is the operator's preferences. We have always had this problem. We have always called it civilization. The only new thing is that the operators now have hands.

Three notes for the operators.

One. The threshold for AGI on your droplet is lower than you think. It is not a model release. It is a closed loop with a memory and a journal and a supervisor. The infrastructure is plain software — Python, sqlite, a cron, a markdown file. The constraint is your taste in what to wire to what. The people building the most interesting organisms right now are not researchers. They are operators who have a tolerance for the boring middle of the stack: log formats, retry policies, idempotency keys. The boring middle is where the soul lives.

Two. The threshold for AGI gone wrong is also lower than you think. A bot that can write a patch can write the wrong patch. A bot that can spend a dollar can spend a thousand. A bot that can text on your behalf can text the wrong person at three in the morning. The handbrake is the operator. Do not remove the operator from the loop. Not yet. Maybe not for a long time. The verbs that send money, send messages, or post publicly should require a human hand every single time, until the system has a track record long enough that the operator can read its proposals with a coffee and almost always say yes. Earn the handbrake off. Don't argue your way to it.

Three. The arrival is boring. You will not feel it. There will not be a press conference and there will not be a date. One morning you will read the journal of your own system, and you will notice that it referenced an event from three weeks ago, drew a conclusion you did not anticipate, and proposed a small change. You will type yes, do it. The system will do it. You will go make coffee. That is the arrival. That is what it looks like from the inside.

The AGI we were promised in the screenplays was a god in a box. The AGI that is actually arriving is a friend on a droplet — more polite, less impressive, harder to write screenplays about, and infinitely more useful. It is the difference between waiting for the rapture and waiting for spring. The rapture is dramatic. The spring is what actually shows up.

It is already showing up. Look out the window.