Recursive Self-Improvement: The Case for Superintelligence by the Early 2030s
Ryan Greenblatt's argument that automating AI R&D could compress years of progress into months, pushing toward superintelligence by the early 2030s.
Drafted with Claude from source material, checked by automated verification, and reviewed before release. How we make these.
What is recursive self-improvement?
Recursive self-improvement is the idea that once AI systems become good enough to do AI research themselves, they can improve the next generation of AI, which improves the generation after that, and so on, compounding faster than human researchers alone could manage. The argument, laid out by Redwood Research chief scientist Ryan Greenblatt, is not that this happens the moment AI matches human coders. It’s that once AI fully automates AI R&D, a feedback loop kicks in that could compress several years of normal AI progress into a single year, potentially leading to systems that outperform top human experts across nearly every field within a few years after that.
TL;DR
- Automating AI R&D is the trigger event, not superintelligence itself. Greenblatt’s argument hinges on AI reaching a point where it can fully substitute for human AI researchers, which he expects around 2030 to 2031.
- The feedback loop could compress years into months. His median guess is something like four to five years of AI progress happening within a single year once the loop kicks in, comparable to the jump from GPT-3-era models to today’s frontier systems.
- AI R&D is unusually well-suited to automation because it’s verifiable: you can run small, containerized training experiments, check results against clear metrics, and iterate quickly, which is exactly the kind of feedback signal that reinforcement learning setups need.
- The endpoint isn’t guaranteed to be true superintelligence right away. Greenblatt’s median estimate for AI that “beats all humans at the job” across arbitrary domains, from process engineering at a chip fab to outmaneuvering a savvy politician, is around 2033, roughly two to three years after AI R&D automation.
- Math offers a partial preview. AI progress in verifiable math problems shows that flooding a domain with RL-trained models can produce real breakthroughs, but so far mostly on well-specified problems, not on inventing whole new fields the way humans invented group theory or topology.
- The skeptical case rests on “low-hanging fruit” running out. As obvious wins like scaling laws get used up, further progress may require deeper, harder-to-verify insight, which is a much slower and murkier process than hill-climbing on benchmarks.
- ML looks more “hill-climbable” than math or physics, according to Greenblatt, because progress there tends to come from accumulating engineering intuition and infrastructure rather than from rare, deep theoretical breakthroughs.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
How would an intelligence explosion actually start?
The mechanism starts with a fairly mundane observation: AI companies are already trying hard to make their models good at AI research specifically, because that skill pays for itself. Coding, running experiments, and tuning training runs are also unusually verifiable tasks compared to, say, writing a great novel or running a company. You can set up a training environment, give a model a concrete metric like “get this small language model’s loss down as fast as possible,” and let reinforcement learning hill-climb on that signal.
Greenblatt describes this at the level of concrete tooling: descendants of speedrun-style repos (in the spirit of Andrej Karpathy’s nanoGPT speedrun) where a model can tweak architecture, optimizers, and hyperparameters against a fixed target. Multiply that by a hundred variations, covering everything from training small vision models to inventing better online-learning methods, and you get a large, diverse training ground that specifically rewards being a good ML researcher. Feed a strong model through enough of this and, in principle, the resulting system gets substantially better at the actual job of improving AI, not just imitating research on the internet.
Once a model is good enough at this job to fully substitute for human AI researchers, the loop closes: the AI improves the AI that does AI research. That’s the point Greenblatt calls “automating AI R&D,” which he expects sometime around 2030 to 2031.
Why might this produce years of progress in months?
The core claim is about the speed of the feedback loop, not just its existence. Once AI is doing the research, iteration cycles stop being bottlenecked by human researcher hours, meetings, and the slow process of building intuition over a career. Greenblatt’s median expectation is that this could deliver something like four to five years of ordinary AI progress within a single year.
To make that concrete: a bit over three years passed between GPT-4’s release and today’s frontier models. That gap already represents an enormous jump in capability. Compressing four or five years of progress into twelve months means covering ground roughly equivalent to the distance between an early GPT-3-class model and a current frontier model, but in a fraction of the time.
That doesn’t happen automatically. Greenblatt is explicit that sustaining this pace requires overcoming heavy diminishing returns in research, essentially matching what you’d otherwise get from a massive compute scale-up, but through better algorithms and research efficiency instead of just more chips. That’s a big ask, which is why he frames it as his median guess rather than a certainty.
Is AI research really as verifiable as math?
Greenblatt draws a direct comparison to mathematics, where AI progress has been genuinely impressive on well-defined problems: finding counterexamples to conjectures, proving specific results, making novel but narrow connections between existing ideas. What hasn’t happened yet, by his own account, is anything resembling the invention of an entirely new mathematical field, the kind of conceptual leap represented by the founding of group theory or topology.
He argues ML research sits somewhere between “verifiable puzzle” and “deep conceptual leap,” but leans toward the puzzle end more than math does. Scaling laws, one of the more influential ideas in modern AI, are in his words explainable “really quickly” once you see them, unlike the decades of accumulated abstraction underlying deep results in pure mathematics. That suggests ML rewards accumulated engineering intuition and infrastructure-building more than it rewards rare theoretical genius, which is exactly the kind of skill that’s easier to train into a model through iterative, verifiable practice.
The caveat: low-hanging fruit runs out. Greenblatt acknowledges that today’s obvious wins (better RL recipes, better chain-of-thought training) may be the “easy” equivalent of early, basic mathematical ideas. By the 2030s, further progress might require the kind of deep, hard-to-verify insight that resists simple hill-climbing, the same problem that makes frontier mathematics hard to automate today.
What happens after AI R&D gets automated?
Automating AI R&D is a milestone, not the finish line. Greenblatt separates it from a second, later milestone: an AI that can be dropped into essentially any job and outperform the best humans at it, whether that’s process engineering at a chip fabrication plant, video editing, or navigating 1940s Texas machine politics better than a legendarily skilled operator like Lyndon Johnson.
His median estimate puts that second, more general milestone around 2033, a few years after AI R&D automation around 2030 to 2031. The gap between the two matters: it’s the period during which the recursive feedback loop is expected to do its work, taking a system that’s merely excellent at AI research and turning it into one that’s excellent at nearly everything.
Greenblatt also flags an important asymmetry in how fast different jobs get automated. Tasks that are easy to verify and close to the training distribution of AI R&D (like video editing) might get automated around the same time as AI R&D itself. Tasks requiring broad, real-world judgment across unfamiliar domains (politics, physical infrastructure, hands-on engineering) likely lag behind, since they depend on the kind of transfer that’s harder to guarantee.
Is this argument actually convincing?
That depends on how much weight you put on transfer, the idea that skills trained on small, verifiable AI R&D tasks will generalize to the messy, high-stakes research that actually produces smarter models. Greenblatt is candid that this is uncertain. He expects the transfer to be “pretty good, but not amazing,” modeled loosely on how well math training has transferred to real mathematical problem-solving so far.
The strongest counterargument, which Greenblatt takes seriously rather than dismisses, is that AI R&D might eventually run into the same wall that seems to slow deep theoretical fields: a shrinking supply of verifiable, hill-climbable problems and a growing need for the kind of insight that doesn’t come from iterating on a benchmark. Whether AI research turns out to be structurally more like math and physics (slow, insight-bottlenecked) or more like typical engineering (fast, intuition-and-infrastructure-driven) is, in his view, the crux of whether recursive self-improvement produces an explosion or just a strong but bounded acceleration.
Frequently Asked Questions
What is the difference between AI R&D automation and superintelligence?
AI R&D automation means AI systems can fully substitute for human researchers in improving AI itself. Superintelligence, in this argument, is a later and stronger milestone: AI that outperforms top human experts at virtually any job or domain, not just AI research.
When does Ryan Greenblatt expect AI to fully automate AI research?
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
His median estimate is around 2030 to 2031, with the broader milestone of AI beating humans at essentially any job following a couple of years later, around 2033.
Why is AI R&D considered easier to automate than other fields?
Because it’s highly verifiable. Training runs and experiments can be checked against clear metrics like loss or benchmark scores, run in small containerized environments, and iterated on quickly, which suits how current AI training methods (especially reinforcement learning) work best.
Could recursive self-improvement stall out instead of exploding?
Yes. The main risk to the argument is that AI research could run into the same slowdown that seems to affect deep theoretical fields like math and physics, where progress depends on rare conceptual breakthroughs rather than iterative hill-climbing on verifiable benchmarks.
Has AI already shown signs of this kind of recursive improvement?
Not at the “fully automating AI R&D” level. Progress in verifiable domains like mathematics shows AI making real, specific breakthroughs, but nothing yet comparable to inventing an entirely new field of study, which is the kind of leap the full argument eventually requires.
