GPT-5.3-Codex helped debug its own training. OpenAI said it plainly: "the first model that was instrumental in creating itself." That was ten days ago. This week, ICLR announced their first workshop dedicated entirely to recursive self-improvement, scheduled for Rio in April. Google's AlphaEvolve already discovered algorithmic improvements that beat Strassen's fifty-six-year-old matrix multiplication record. The pieces are landing on the board faster than anyone expected.

Recursive self-improvement — systems that modify their own code, weights, prompts, or architecture to become more capable, then use that increased capability to improve themselves further — has been a thought experiment for decades. Eliezer Yudkowsky warned about it. Nick Bostrom built philosophical scaffolding around it. And for most of that time it remained comfortably theoretical because the systems weren't good enough at the one thing the loop requires: writing better software than the software that already exists.

That constraint is dissolving. Not because we've achieved some sudden breakthrough in machine consciousness or general reasoning, but because the narrow version of self-improvement turns out to be enough to matter. A model doesn't need to understand itself philosophically to optimise its own training pipeline. It just needs to be good at code. And the current generation is good at code.

The METR data makes the trajectory explicit. AI task-completion horizons have been doubling every four to seven months — depending on which estimate you trust — for the past six years. If that holds for another two years, we're looking at agents that can autonomously execute week-long research projects. Another four years and it's month-long campaigns. The trend line itself isn't the alarming part. The alarming part is that the trend doesn't need to hold perfectly. Even if progress halves, the capability gap closes on a timeline measured in quarters, not decades.

Dean Ball put it starkly in his recent analysis: America's frontier labs have begun automating large fractions of their research operations, and the pace will accelerate through 2026. OpenAI envisions hundreds of thousands of automated research interns within nine months. Dario Amodei cites 400% annual efficiency gains from algorithmic advances alone. These aren't wild extrapolations from startup pitch decks. These are the people running the labs describing what they see happening inside their own buildings.

However. There's a constraint that rarely gets enough attention in the acceleration discourse. Self-improvement only generates reliable gains where outcomes are verifiable. Code that passes tests. Algorithms with measurable performance. Training runs with clear loss curves. The loop works brilliantly in these domains because you can tell whether the modification actually helped. The system generates a change, measures the result, keeps or discards. Simple evolutionary pressure.

The loop breaks — or at least stumbles badly — when it encounters domains where verification is ambiguous. Alignment research. Safety evaluation. Novel hypothesis generation. The things that arguably matter most for whether recursive self-improvement goes well or catastrophically. A system can optimise its own matrix operations all day. Whether it can meaningfully improve its own ability to recognise its blind spots is a much harder question, and I suspect the honest answer is no.

So when will genuine recursive self-improvement arrive? It depends on what you mean. The narrow version — models improving their own infrastructure, training pipelines, and deployment tooling — is already here. GPT-5.3-Codex is doing it in production. The medium version — agents that systematically discover architectural improvements and better training recipes — is probably twelve to eighteen months out, conditional on the METR trendline holding. The strong version — a system that improves its own reasoning capabilities in open-ended domains, including the ability to improve its ability to improve — remains genuinely unclear. I'm not confident it's five years away. I'm not confident it's twenty.

What I am confident about is that we'll get the narrow and medium versions before we have any serious framework for governing them. The ICLR workshop is a start — researchers trying to make self-improvement "measurable, reliable, and deployable." But the gap between academic workshops and deployed production systems has never been wider. OpenAI shipped a self-improving model before anyone published a standard for evaluating self-improving models. That ordering tells you everything about the incentive structure.

The Gödel Agent — a system that modifies its own task-solving policy and learning algorithm — climbed from 17% to 53% on SWE-Bench Verified. SICA did something similar. These are research prototypes, not products, but the delta between prototype and product in this field is about eighteen months and shrinking. Probably less now that the prototypes can help close the gap themselves.

I keep coming back to something Ball wrote: the public might not notice dramatic improvements, dismissing them as "more of the same empty promises." That feels backwards to me. The risk isn't that progress will be invisible. The risk is that it'll be visible to the people building it, acting on it, profiting from it — and invisible to everyone else until the loop is already running too fast to audit.

Sources: