Original title: The Economics of Recursive Self-Improvement
Tom Cunningham, Lukas Althoff, Basil Halperin, Brian Jabarian +5 · September 2026
Could artificial intelligence rapidly upgrade itself in an unstoppable loop? This idea is called recursive self-improvementWhen an AI system helps design or improve the next version of itself, creating a loop where each generation gets smarter faster.. Researchers built economic models to test whether this could actually happen. They mapped AI development as networks of feedback loopsA cycle where an output of a process becomes an input that influences the process again, like a microphone held too close to a speaker creating a louder and louder screech., where AI tools help build even better AI tools. They found that whether progress explodes depends on how strongly each step boosts the next. They also separated narrow skills, like scoring high on coding tests, from broad skills that help the wider economy. After plugging current data into their framework, they ran a calculation. Their main finding is that today's feedback loops are not strong enough to trigger an explosive, self-sustaining loop on their own, although these loops do appear to be strengthening over time.
You are trying to get better at puzzles by solving puzzles. Normally, solving one puzzle makes you a little better at the next one, but you still need to buy new puzzles from a store. Now imagine that solving a puzzle itself generates instructions to make the next puzzle harder. The better you get, the faster you can produce more difficult puzzles. In this paper, the feedback loop is between AI improving its own design and then using that improved design to make even better AI. The strength of that loop determines whether progress snowballs or fizzles out.
If AI systems can improve themselves, we might get rapid progress or even an intelligence explosion. But this paper suggests the feedback loops needed for that aren't strong enough yet. Companies and policymakers can use the model to anticipate whether AI development is about to take off—or not.
The model relies on estimated parameters, which the authors acknowledge are uncertain. The calibration is a back-of-the-envelope calculation, so the results are approximate. The paper does not prove whether recursive self-improvement will or will not happen, only that current loops are not self-sustaining.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
The Economics of Recursive Self-Improvement — Tom Cunningham, Lukas Althoff, Basil Halperin et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.