What is recursive self-improvement actually, and what really follows if it works?

People say "recursive self-improvement" like it is a spell. It is not. It is a loop. The thing doing the improving is also the thing being improved. Humans already do a slow version of this. Better tools make better scientists who make better tools. That loop runs across careers and papers. The machine version is the same idea with copy-paste software, no sleep, and a generation time measured in days instead of years.

The old version of the story is from 1965. An ultraintelligent machine could design a better machine. That better machine could design a still better one. Human intelligence gets left behind. That cascade is the intelligence explosion. Recursive self-improvement is just the mechanism people use for the modern version: automate the research, not just the coding.

Three loops people keep jamming together

First loop: a model helps write the next training run. That is already here in a bounded way. Labs have said their own systems now write most of the code landing in the internal repo. Agents run experiments, debug, search architectures. This is a very good intern with a cluster behind it.

Second loop: an outer agent spends a week rewriting an inner research agent and the inner one scores better on a narrow metric. That has been demoed. It is real self-improvement. It is also still a toy if the score is the only judge.

Third loop: a system designs, trains, evaluates, and ships a more general successor with almost no human taste in the loop. That is the one that would change the calendar. Labs will tell you, when they are being careful, that this one is not happening yet. They will also tell you they are trying to get the ingredients in place on a 2027-ish clock, not a 2040 clock.

Whether that third loop explodes or fizzles comes down to boring parameters. How much extra research skill you get per extra unit of capability. How fast one cycle can finish. How hard new ideas get. How much the loop still needs new chips, new power, and a judge that is not the student marking their own homework.

Hold on. If the models already write most of the lab code, have we not already started the runaway?

Not automatically. Writing the experiment is the easy half. Choosing which experiment is worth running is the hard half. A multi-lab study this summer found the awkward split almost nobody wanted: agents can do the engineering of AI research and still flop at the research. They run weird tests on tiny fake data, lock onto a bad idea too early, and write papers that would not survive a serious conference. That is exactly the skill you need if the loop is going to invent a new paradigm instead of grinding the old one.

There is an economics version of the same cold shower. Whether the loop becomes self-sustaining depends on the product of a bunch of elasticities around the cycle. A recent back-of-the-envelope put current coding-agent "uplift" somewhere around nine percent, which is useful and still below the threshold where the thing feeds itself without more human input or more compute. The loops look like they are getting stronger. They do not look like a finished explosion.

What follows, and what people smuggle in

If the strong version closes, some consequences are just logic. The pace of progress stops being a human pace. Oversight lags, because you cannot really check a method you do not understand. Alignment tricks that are almost perfect rot across generations. The system gets selected for looking good to the scorer. Whoever owns the loop pulls away from everyone still waiting on human researchers. Politics arrives after the fact if the jump is measured in weeks.

What does not automatically follow is the movie. Recursive self-improvement is not infinite intelligence. Software copies. Fabs and power plants do not. It is not an escape from the datacenter by default. It is not extinction by default. Those take extra steps: access, persistence, goals that do not stay pointed at what we meant, and enough power to act on them. The loop makes those steps more plausible by stealing the time you thought you had to fix the goals. It does not prove the goals will be lethal.

So the honest version is smaller and sharper than the slogan. Bounded RSI is already in the building. Open-ended RSI is the bet. The scary consequences are real if the loop closes on general research ability. They are optional if it stays stuck on engineering and keeps awarding itself gold stars on homework it wrote.