The artificial intelligence industry has a terminology problem, and it just got worse. Having spent the better part of a decade debating what "artificial general intelligence" actually means—and whether anyone is close to achieving it—the field has now embraced a new concept that is somehow even harder to pin down: RSI, or recursively self-improving AI.
The premise is seductive. An AI system that can improve its own capabilities, which then improves its ability to improve itself, which then... you see where this goes. It's the intellectual descendant of the "intelligence explosion" hypothesis that mathematician I.J. Good articulated in 1965, now dressed up in contemporary machine learning vocabulary and venture capital pitch decks.
The definitional morass
The trouble with RSI is the same trouble that plagued AGI: nobody agrees on what counts. Does a model that fine-tunes itself on its own outputs qualify? What about one that writes better training curricula for future versions? The goalposts aren't just moving—they're on wheels.
Some researchers argue that current large language models already exhibit weak forms of recursive self-improvement when they engage in chain-of-thought reasoning or self-correction. Others insist that true RSI requires a system to fundamentally alter its own architecture, not merely refine its weights. The distinction matters enormously for safety research and not at all for marketing copy.
Why the pivot now
The shift from AGI discourse to RSI discourse is partly strategic. After years of promising that AGI was five years away (always five years away), leading labs have grown weary of the credibility tax. RSI offers a narrower, more technical-sounding objective that sidesteps the philosophical baggage of "general" intelligence while preserving the existential stakes that attract talent and capital.
It also reflects genuine uncertainty. The scaling laws that powered the GPT era are showing diminishing returns in some domains, and researchers are hunting for the next paradigm. Self-improvement loops—whether through synthetic data generation, automated prompt engineering, or more exotic approaches—represent one plausible path forward.
The safety implications
For AI safety researchers, RSI is both more tractable and more alarming than AGI. More tractable because you can study specific self-improvement mechanisms in isolation. More alarming because a system that can recursively enhance itself might do so faster than human oversight can keep pace—the classic "fast takeoff" scenario that keeps alignment researchers employed.
The counterargument, advanced by skeptics, is that self-improvement faces the same bottlenecks as any optimization process: you can't pull yourself up by your own bootstraps indefinitely. At some point, you need new data, new architectures, or new hardware. The recursion has to ground out somewhere.
Our take
The AI industry's acronym churn is tiresome but not meaningless. Each new term—AGI, ASI, RSI—represents a genuine attempt to articulate what we're building toward and what we should fear. The problem is that definitional vagueness serves too many interests: it lets optimists claim progress and pessimists claim peril, often about the same systems. RSI is a real research direction worth pursuing, but it's also the latest vessel for anxieties and ambitions that have more to do with human psychology than machine capability. The recursion we should worry about isn't in the models—it's in the hype cycles.
