Andrej Karpathy · Independent · 2026-05-16
Read the Software 2.0 essay (the AutoResearch antecedent) on medium.com
Karpathy's Software 2.0 essay (2017) argued that we were moving from hand-written code to learned weights as the substrate of software. Nearly a decade later, his "AutoResearch loop" framing is the natural successor: not just weights replacing code, but a training loop that improves itself while running.
The technical content is narrower than the headlines suggest. AutoResearch describes continual or online RL: an agent that updates during deployment, against real tasks, with verifiable reward signal. There is no claim of arbitrary self-improvement, no claim of capability runaway. What it captures is the moment where the gradient between "AI doing research" and "AI doing AI research" stops being a slogan and becomes a deployable pattern. ICLR 2026's recursive-self-improvement workshop took the framing seriously enough to make it the organizing theme.
This matters for two reasons. First, the press cycle has already discovered "self-improving AI" as a story — Meta's Llama 5 launch claimed "Recursive Self-Improvement" as a feature, and that framing will be everywhere by 2027. Most of it will be marketing on top of fine-tuning that converges in a few iterations. Karpathy's loop is what the technically literate version of the claim actually looks like.
Second, the loop reframes what "scaling" means in 2026-2027. The pretraining-compute-doubling story still holds — Epoch tracks ~4-5x/yr. But the binding constraint has moved to inference and post-training. Continual RL is a way of spending inference compute as if it were training compute, in a loop, against the actual task distribution. That is structurally different from "train a bigger model."
The honest version of the claim: directionally meaningful, technically narrow, easy to over-extrapolate. Read more of Karpathy's writing before you accept any single take on what RSI does and does not mean.