In 1993, Vernor Vinge wrote a paper arguing that once intelligence crosses a certain threshold, it starts designing better intelligence on its own — and each newly designed intelligence then designs the next generation, in an unbroken loop. He called this the technological singularity. At the time, the computer science world filed it away as an interesting but speculative thought experiment. Thirty-two years later, the AI lab Anthropic published a report tracking exactly where that loop now stands. Its finding: AI systems are already doing real work inside the process of training the next generation of AI.
When the report hit Hacker News, it drew 489 comments. A single technical article rarely pulls in that much traffic. And the reactions didn't converge. "This is a warning sign" and "this overstates where we actually are" used the exact same words to point in opposite directions. The absence of consensus is itself a signal — this shift hasn't settled into a stage anyone can confidently name yet.
Once a tool starts taking part in building the next tool, what happens to the person using it? That's the question this piece starts from.
What It Actually Means for AI to Train the Next AI
The term recursive self-improvementRSI has to be used with technical precision. What Anthropic's report calls the current stage does not mean AI is autonomously building its own successor with no human involved. Under the direction of human researchers, AI systems are contributing to the next generation of models — generating training data, helping set evaluation criteria, reviewing experimental designs.
Even so, once that contribution crosses a certain threshold, the character of the improvement rate changes. Unlike work done directly by humans, AI-assisted work runs in parallel, twenty-four hours a day. An experimental design process that would take 50 researchers three months can, with AI assistance, run far more experiments in that same window. More experiments running means faster identification of where the next model should improve. That's where the most concrete shift in the early stages of RSI actually shows up.
For users, this shift shows up directly in how fast their tools get replaced. GPT-3 launched in June 2020; GPT-4 arrived in March 2023 — roughly 32 months apart. After that, GPT-4o shipped in May 2024, and o1 followed just four months later, in September. Models with a genuinely different character than their predecessors started appearing on a four-to-five-month cycle. Claude, Gemini, and the Llama family have all shipped new versions at a comparable pace. It would be a stretch to credit all of this to RSI. But it's worth noting that this compression of release cycles coincides with the point where AI started assisting in AI development itself.
The Bar for "Skilled" Has Moved
Over the past three years, the people considered good at using AI tools tended to share a few traits: writing precise prompts, an instinct for which model suits which task, and speed at editing raw output. Of these, prompt precision is the skill losing value fastest.
The reason is simple. Most of today's precise prompting is really a workaround for the current model's limitations. If a model can't hold broad context, you write out a long backstory. If it's sensitive to framing, you spell out a persona. As models improve and those limitations shrink, the techniques built to route around them age out along with them. A prompting style refined over three years can become noise that actively lowers output quality on the next model version.
A feel for which model to use fades more slowly. Knowing that Claude tends to be stronger at parsing long documents, or that GPT-4o holds conversational context better, is useful right now. But as models keep converging toward general-purpose capability, that distinction blurs. There's no guarantee today's differences survive into the next release.
What holds up longest is editing ability. More precisely: the ability to judge what counts as a good result, spot what's wrong, and decide which direction to fix it in. This lives in the person, independent of whatever tool they're using. However fast the tools change, the instinct for distinguishing good writing from bad, and a sound strategy from an unsound one, sits outside the tool entirely.
The Skeptical Case
A large share of those 489 comments were skeptical. The critique: framing this as "AI training AI" conjures an image far more dramatic than what's actually happening. RSI's current contribution, the argument goes, happens under close human supervision and is nowhere near autonomous self-improvement.
There were more specific objections too. The metrics in the RSI report largely come from controlled research settings, and how quickly that translates into actual product development cycles is a separate question. During the two years between GPT-3 and GPT-4, the bottleneck wasn't algorithms — it was data quality and safety verification. Even if AI speeds up algorithmic iteration, data collection and safety review still require human time. Whatever acceleration RSI delivers, where that acceleration hits its next wall remains an open question.
This pushback is fair. Overstating the pace of change to manufacture anxiety isn't the point of this piece. But "it's not full RSI yet, so keep doing things the way you always have" doesn't hold up well against how fast things have actually moved over the past three years. In 2021, most solo operators didn't build AI tools into their workflow at all. By 2024, not doing so raised real doubts about their competitiveness. That shift happened in three years. Measured against that pace, "there's still plenty of runway" is a hard claim to make with confidence.
What Solo Builders Should Check Right Now
If the question is what to do about all this, learning new tools faster isn't the first answer. If tools are going to keep changing this quickly, tool-independent judgment holds its value far longer than the ability to rapidly absorb a new interface.
It's worth checking whether judgment has quietly crept into the work you've been delegating to AI. Summarizing documents, copyediting, listing out ideas — fine to hand off. But if "is this the right direction" has effectively become a decision the AI's output makes for you, then your judgment shifts every time the tool does. Someone who hands their judgment to a tool has to get their new judgment from that same tool once it changes. Someone whose standard lives inside the tool and someone whose standard lives inside themselves end up in very different positions the moment the tool gets swapped out.
It's also worth checking whether you can actually describe your current tool's limitations in words — which kinds of questions it gets wrong, in which contexts you shouldn't trust its output. If you can articulate that, you'll notice quickly when those limits shift in the next version. Someone who only knows how to operate the tool won't notice the limits have moved, and will keep working the same way regardless — sometimes even carrying over prompt structures built to route around an old model's weaknesses, only to find they now drag performance down on the new one.
Doing the same task without AI now and then is useful too. The longer you lean on a tool, the easier it is to lose track of where your own baseline sits. Knowing that baseline is what lets you accurately gauge how much the tool is actually contributing. Without it, you can't tell whether a result reflects your own level or the tool's — and you're caught flat-footed the moment the tool disappears or changes.
Someone who spends years running a café and training their palate to judge coffee keeps that standard no matter how the espresso machine changes. Someone who only learned to operate the machine has to start over from scratch the moment it's replaced. Builders working with AI tools face the same fork in the road. However fast a tool rewrites itself, the judgment of which direction to point it in — and when to stop — belongs to the person, not the tool.
