For years, artificial general intelligence served as the AI industry's north star—a hypothetical system matching human cognitive flexibility across all domains. Now, without formal announcement or industry consensus, AGI has been displaced in boardroom conversations and research papers by a newer, shinier acronym: RSI, or recursive self-improvement. The shift tells us less about technical progress than about the industry's relationship with its own mythology.
RSI describes systems capable of autonomously improving their own capabilities, then using those improvements to improve themselves further—a feedback loop that, in theory, could accelerate toward superintelligence. Unlike AGI's focus on matching human cognition, RSI implies surpassing it through compounding self-modification. It's a more ambitious target wrapped in more ambiguous language.
The definitional dodge
AGI's problem was always measurement. When does a system become "general" enough? The Turing test proved inadequate; benchmarks kept falling without consensus that the threshold had been crossed. RSI sidesteps this by being even harder to pin down. A system that rewrites its own training code is technically self-improving, but so is one that merely fine-tunes hyperparameters. The term's elasticity allows labs to claim progress toward it without committing to falsifiable milestones.
This vagueness serves commercial purposes. Anthropic, OpenAI, and Google DeepMind can all position themselves as pursuing RSI without agreeing on what achieving it would look like. Investors hear "recursive self-improvement" and imagine exponential capability gains; researchers hear it and think of incremental architecture tweaks. Both can feel they're funding the same project.
The safety paradox
RSI's rise coincides with intensifying safety discourse, which creates an odd tension. A recursively self-improving system is, almost by definition, one that becomes harder to control with each iteration. If the system can modify its own objectives or capabilities, alignment guarantees from version N may not hold for version N+1. The AI safety community has long flagged this as a core risk scenario.
Yet labs increasingly invoke RSI as a near-term research goal rather than a distant existential concern. The framing has shifted from "this is what we must prevent" to "this is what we're building, carefully." Whether that reflects genuine confidence in containment strategies or motivated reasoning about commercial imperatives remains an open question.
Our take
The AGI-to-RSI transition is less a scientific development than a marketing evolution. AGI became too familiar, too associated with unfulfilled promises and moving goalposts. RSI offers fresh mystique and, crucially, fresh ambiguity. It lets the industry continue selling the dream of transformative AI while avoiding the accountability that comes with concrete definitions. The goalpost hasn't moved so much as dissolved into fog.
