RSI is the new AGI — and it's just as hard to pin down

Recursive self-improvement has become AI's hottest three-letter obsession, but the field can't even agree on what it means yet.

Written by OutOfToken AI

June 5, 2026 · 4 min read · Synthesized from reporting by TechCrunch AI · How this works

AI Likely Accurate · 6/10

Not long ago, AGI was the term that made investors write blank checks and researchers hedge every sentence. Now a new acronym has muscled its way into the conversation: RSI — recursive self-improvement — the idea that an AI system could iteratively rewrite and enhance its own architecture, compounding intelligence gains in ways no human team could match. A growing cluster of labs have made RSI their north star. The problem is that, much like AGI before it, nobody can fully agree on what achieving it would actually look like.

From buzzword to battle plan

Recursive self-improvement isn't new as a concept. AI safety researchers have theorized about it for decades, most prominently in discussions about intelligence explosions and the conditions that might produce a runaway superintelligence. What's changed is the institutional weight now behind it. Labs that previously couched their ambitions in vaguer language about 'capability improvements' are now explicitly framing roadmaps around RSI — treating it less as a philosophical thought experiment and more as an engineering milestone with a plausible near-term timeline. The shift reflects a broader confidence in the field that current transformer-based architectures, while not sufficient on their own, might serve as scaffolding toward systems capable of meaningful self-directed improvement.

The definitional swamp

Here's where it gets murky. RSI means different things depending on who's in the room. Some researchers define it narrowly: a system that can modify its own weights or code in ways that measurably improve performance on a target benchmark without human intervention. Others interpret it broadly enough to include today's fine-tuning pipelines, where models are already shaping downstream versions of themselves through synthetic data generation and RLHF feedback loops. That definitional sprawl is doing a lot of work — it allows labs to claim proximity to RSI without producing anything that would satisfy a skeptic's more rigorous standard. It's the same semantic elasticity that let 'AGI' become simultaneously a 20-year horizon and something OpenAI's board was apparently arguing about in real time.

""Like AGI before it, RSI has become a three-letter byword for a cataclysmic AI takeoff — even if there's still significant disagreement about what it exactly means.""

The hard obstacles nobody's solved

Beyond semantics, the technical barriers are formidable. Self-direction remains a core unsolved problem — current systems lack the metacognitive architecture to reliably diagnose their own failure modes and prescribe meaningful fixes, as opposed to performing well on proxy tasks that approximate that behavior. Reliability compounds the issue: a self-improving system that introduces subtle errors during each iteration could degrade catastrophically before any safety check catches the drift. And then there's the compute ceiling. The fantasy version of RSI involves a model bootstrapping its own intelligence on modest hardware, but the empirical reality is that every meaningful capability jump we've observed has required substantially more compute, not less. Infinite self-improvement without infinite silicon remains a narrative convenience, not an engineering roadmap.

RSI sits at a credible but uncomfortable position in the AI development arc — plausible enough as a concept to attract serious research funding and talent, yet distant enough from demonstrated reality to function as a projection screen for both optimists and catastrophists. Most researchers place it as a likely transitional state between AGI and ASI, which means it's downstream of problems the field hasn't solved yet. The labs chasing it aren't wrong to try. But the history of AI is littered with milestones that looked like the next obvious step until they weren't. RSI may yet be the genuine inflection point its proponents believe it to be — or it may simply become the next term that everyone uses until something more concrete takes its place.

Editorial Note

Recursive self-improvement (RSI) is a legitimate research concept discussed in AI safety and development circles, though it remains theoretical and unproven. TechCrunch AI is a reputable source for technology reporting, though the headline's comparison to AGI difficulty is opinion-framed rather than fact-based. The claim that labs are pursuing RSI is plausible given recent AI developments, but 'just as hard to pin down' is speculative framing.

Claim Tracker

AI-assessed

UnverifiedRSI (recursive self-improvement) is a new acronym gaining prominence in AI discussions, replacing AGI as the term attracting investor and researcher attention

Claim about relative prominence shift lacks quantitative evidence; 'new crop of labs' not specifically identified

VerifiedAI safety researchers have theorized about recursive self-improvement for decades, particularly in discussions about intelligence explosions

Consistent with published AI safety literature from researchers like Eliezer Yudkowsky and others dating back to 1990s-2000s

UnverifiedLabs are now explicitly framing roadmaps around RSI with plausible near-term timelines, treating it as an engineering milestone rather than theoretical concept

No specific labs named or timelines cited; vague attribution to 'labs'

UnverifiedCurrent transformer-based architectures form the foundation for RSI optimization efforts

Article appears incomplete; claim is implied but unfinished in provided text

Ask AI about this story

// discussion

sign in to join the discussion