In 1993, Vernor Vinge wrote a paper. Its argument was that once intelligence crosses a certain threshold, that intelligence begins to design better intelligence, and the newly created intelligence designs the stage after that, in a continuing loop. He called this the technological singularity, and at the time the computer science community filed the paper away as interesting but speculative. Thirty-two years later, Anthropic released a report tracking which stage of that loop we have actually reached. AI systems, it reported, are now being put to real work on parts of the job of training the next generation of AI.
When the report landed on Hacker News, it drew 489 comments. It is rare for a single technical article to generate that volume of response. The direction of the comments was not uniform, either. The "this is a warning sign" camp and the "they're overstating where we are" camp used the very same words to point in opposite directions. The absence of a settled reading is itself a signal—a sign that this shift is still at a stage that is hard to name.
When a tool starts taking part in the work of building the next tool, what happens to the person who uses it? That is the question this piece begins from.
What It Actually Means for AI to Train the Next AI
The phrase recursive self-improvementRSI, Recursive Self-Improvement has to be used with technical precision. What Anthropic's report calls the current stage does not mean that AI builds its next version entirely on its own, fully autonomously. It means that, under the direction of human researchers, AI systems are contributing to the work—generating the data needed to train the next generation of models, setting up evaluation criteria, reviewing experimental designs.
Even so, once that contribution passes a certain level, the character of the improvement rate changes. Unlike work done directly by humans, work that AI assists is processed in parallel and runs 24 hours a day. If AI helps design the experiments that 50 researchers would have spent three months on, the number of experiments that can run in the same window grows dramatically. More experiments running means a faster path to finding the direction of improvement for the next model. This is where the most concrete change in the early stage of RSI is found.
From a user's standpoint, this change connects directly to how often the tools get replaced. GPT-3 was released in June 2020, and GPT-4 arrived in March 2023—an interval of roughly 32 months. After that came GPT-4o in May 2024, and o1 in September of the same year. Models with a different character from their predecessors began appearing on a four-to-five-month cycle. The Claude, Gemini, and Llama families raised their versions at a similar pace. It is hard to credit all of this to RSI. But the fact that the period of acceleration overlaps with the point when AI began assisting AI development is worth noting.
The Benchmark for Expertise Is Shifting
Over the past three years, the people praised for using AI tools well shared a few things in common: the ability to write precise prompts, a feel for which model suits which task, and the speed to edit outputs quickly. Of these, prompt precision is the very skill that will lose its value fastest.
The reason is simple. Today's precise prompts are, for the most part, workarounds for the limits of current models. If a model can't grasp broad context, you have to spell out the background at length; if it's sensitive to role-setting, you have to name a persona explicitly. As models improve and those limits shrink, the techniques built to route around them grow obsolete alongside them. A prompt pattern honed over three years can become noise that actually lowers output quality on the next version of the model.
A feel for model selection fades a little more slowly. Knowing that Claude is strong at analyzing long documents while GPT-4o is strong at holding conversational context is useful right now. But as models gradually become more general-purpose, that distinction blurs too. There is no guarantee that today's differences will hold in the next version.
What stays valid longest is editing ability. More precisely: the ability to judge what makes a good result, to find where it went wrong, and to decide which direction to fix it in. This stays inside the person, regardless of the tool's character. No matter how fast the tools change, the sense for telling good writing from bad, and the eye for telling a sound strategy from an unsound one, live outside the tool.
The Skeptical View
A good share of those 489 comments were skeptical. The objection was that the phrase "AI training AI" conjures a far stronger image than the actual current level warrants. RSI's contribution today happens under thorough supervision by human researchers, and is a long way from autonomous self-improvement.
There were more specific rebuttals as well. The metrics in the RSI report come mostly from experiments in controlled research settings, the argument went, and how quickly those feed into actual product development cycles is a separate question. Across the two years from GPT-3 to GPT-4, the bottleneck lay less in algorithms than in data quality and safety verification. Even if AI speeds up the algorithmic iteration rate, the bottlenecks of data collection and safety verification still demand human time. Even granting that RSI produces an acceleration effect, where that acceleration gets stuck is an open question.
This counterargument is sound. Manufacturing anxiety by overstating the pace of change is not the aim of this piece. Still, to conclude that "it isn't full RSI yet, so the current approach is fine to keep using" is to ignore that the actual pace of change over the past three years is already unlike the past. In 2021, most solo planners did not fold AI tools into their workflows. By 2024, not folding them in invites doubts about your competitiveness. That shift happened in three years. Measured against that pace, "there's still plenty of time" is not an easy thing to say.
What Solo Planners Should Audit Now
If the question is what to do in the face of this change, learning new tools faster is not the first answer. If tools change quickly, the ability to make tool-independent judgments stays useful longer than the ability to rapidly absorb how to use a given tool.
It is worth checking whether judgment is mixed into the work you have currently delegated to AI. Summarizing documents, proofreading text, listing ideas—these are fine to hand off. But if the structure has become one in which an AI's output effectively decides whether "this is the right direction," then when the tool changes, the basis for your judgment wobbles along with it. Whoever hands judgment to a tool has to receive their new standard from the tool whenever it changes. The person whose standard lives inside the tool and the person whose standard lives inside themselves end up in completely different positions when the tool gets swapped out.
It is also worth checking whether you can put the limits of your current tool into words. Which kinds of questions does it answer wrongly; in which contexts is its output hard to trust? If you can articulate this, you can quickly tell when the limits have moved in the next version. Someone who only knows how to use the tool won't notice when the limits have shifted, and keeps working the same way. Then they apply the prompt structures they built to route around the old model's limits straight onto the new model, and end up degrading its performance instead.
Periodically doing the same work without AI is useful, too. The longer you lean on a tool, the easier it is to forget where your own baseline sits. You have to know that baseline to gauge precisely how far the tool is assisting you. Without it, you can't tell whether a result the tool produced reflects your level or the tool's—and when the tool disappears or changes, you are caught off guard.
Someone who has run a café and spent years training a sense for judging coffee keeps that standard no matter how the espresso machine changes. Someone who learned only how to operate the machine starts over from scratch when it gets replaced. Planners who use AI tools face the same fork in the road. However fast the tool rewrites itself, the judgment that decides which direction to point it and at what moment to stop lives outside the tool.
