Riedl argues recursive AI self-improvement will be gradual

The runaway 'intelligence explosion' scenario tends to swallow every AI-risk conversation, and a plainer accounting of what recursive self-improvement actually looks like, once the science fiction is stripped out, is the useful corrective. That is what Mark Riedl has attempted in a long essay on Medium, and the honest headline is that the scenario looks less dramatic than the classic I.J. Good framing from 1965.

Riedl defines RSI plainly: an AI system that creates a version of itself that is improved, which in turn creates a further improved version, and so on. He then observes that some of this is quietly happening already. His claim is that 80% or more of the Claude and GPT codebases were written by Claude and GPT respectively, but that the actual recursion is still confined to lower-level work: architectural details, hardware optimization, and pieces of the training and execution harnesses. It is real, but it is not the model rewriting its own mind.

The reason he thinks a hard take-off is hard to justify comes down to the boring constraints of data and compute. New data is harder to get, cheap data is tapping out, and by his framing, data can be created by people at the speed that people make data. Recursive tweaks to a training harness cannot conjure new tokens. Compute for radical experimentation is expensive, so each recursive step imposes a real cost. Stack these up and you get s-shaped curves rather than an exponential lift-off.

The caveat is that diminishing-returns arguments have a mixed track record in this field, and Riedl himself frames the piece as scenario analysis rather than prediction. What the essay does not give you is a specific tripwire, the data source or architecture shift or budget threshold that would flip a soft take-off into a hard one, nor a comparison with what labs claim to be seeing inside their own self-improvement experiments.

For leaders trying to price AI risk, the useful part is that it gives you concrete places to look, the training harness, the execution harness, hardware optimization, data pipelines, rather than a mystical black box. If any notion of the singularity is to occur, Riedl argues, a slow take-off is the more likely path, and that is the version worth planning around.