The artificial intelligence industry has a branding problem, and it is solving that problem the way it solves most problems: by inventing new terminology. AGI — artificial general intelligence, the hypothetical system that can do anything a human can — has grown stale. It promised too much, arrived too slowly, and became a punchline among skeptics who noted that the definition shifted whenever a benchmark was met. Enter RSI: recursive self-improvement, the notion that an AI system could enhance its own capabilities in a self-perpetuating loop, eventually outpacing human intelligence without further human input.
The pivot is underway across the major labs. Internal strategy documents, investor presentations, and research roadmaps increasingly frame RSI as the true north star, with AGI demoted to a waypoint. The logic is seductive: if you cannot agree on what general intelligence means, perhaps you can agree that a system improving itself faster than humans can intervene would be transformative. The trouble is that RSI is, if anything, harder to pin down than AGI ever was.
The definitional shell game
AGI suffered from a moving-target problem. When GPT-4 passed the bar exam, critics noted that passing exams was never the real test. When models began writing competent code, the goalposts shifted to embodied reasoning, then to long-horizon planning, then to something else. RSI inherits this pathology and amplifies it. What counts as self-improvement? If a model fine-tunes itself on curated data, is that recursive? If it writes code that makes its inference faster, does that qualify? The term is capacious enough to describe everything from a modest AutoML loop to a hypothetical intelligence explosion — and therein lies its rhetorical utility.
For labs seeking continued funding, RSI offers a fresh narrative arc. AGI was always a binary: either you have it or you do not, and the industry conspicuously did not. RSI, by contrast, admits of degrees. A model can be a little self-improving, then somewhat more, then dramatically so. This gradient lets companies claim incremental progress toward a goal that remains conveniently undefined.
Why the shift matters now
The timing is not accidental. Scaling laws — the reliable relationship between compute, data, and model performance — are showing diminishing returns at the frontier. The next order-of-magnitude improvement will cost tens of billions of dollars and may not deliver proportional gains. Labs need a story that does not depend solely on throwing more GPUs at the problem. RSI supplies that story: if the model can improve itself, perhaps raw scale becomes less critical.
There is also a regulatory angle. Policymakers have begun drafting frameworks around AGI, attempting to define thresholds that would trigger oversight. RSI sidesteps those frameworks neatly. It is not AGI, after all — it is something else, something newer, something that existing proposals do not quite address. Whether this is strategic or coincidental, the effect is the same: the industry stays one vocabulary term ahead of the regulators.
Our take
RSI is not a scientific concept so much as a fundraising concept dressed in scientific clothing. That does not mean recursive self-improvement is impossible or unimportant; it means the term is doing more work in pitch decks than in research papers. The AI industry's habit of defining success in ways that are always just over the horizon is understandable — every ambitious field does this — but investors, policymakers, and the public deserve clearer metrics. Until RSI comes with falsifiable criteria and agreed-upon benchmarks, it is AGI with a fresh coat of paint: a promise that cannot be broken because it was never precisely made.
