Singularity.Kiwi

AI progress, with honesty.
A recursive spiral of code feeding back into itself, glowing cyan nodes on a dark background, symbolizing self-improving AI systems
AI & Singularity 31 May 2026

RSI Is the New AGI — and It's Just as Hard to Pin Down

RSI is the buzzword of 2026. Everyone's chasing it, nobody agrees on what it means, and the gap between hype and reality is starting to show.

Recursive Self-ImprovementAGIAI StartupsAndrej KarpathyAnthropic

The word “recursion” is having a moment. Two startups have taken the name, and half the pitch decks in Silicon Valley now reference “recursive self-improvement” somewhere in the roadmap. RSI — the idea that AI systems could autonomously upgrade themselves — is the three-letter acronym that replaced AGI as the thing everyone’s building toward.

The only problem? Nobody agrees on what it actually means.

🔍 THE BOTTOM LINE

RSI is the new AGI: a powerful, contested, dangerously vague idea that every AI lab now claims to be pursuing, even as evidence mounts that nobody’s particularly close to achieving it.

The RSI Boom

Earlier this month, well-known AI researcher Richard Socher launched a startup literally called Recursive Superintelligence, raising over $500 million with RSI as the explicit goal. “Our main focus is to build truly recursive, self-improving superintelligence at scale,” Socher told TechCrunch, “which means that the entire process of ideation, implementation, and validation of research ideas would be automatic.”

Socher’s startup is the boldest claim, but it’s hardly alone. Andrej Karpathy — the legendary Tesla and OpenAI researcher who just joined Anthropic’s pre-training team — has been running an open-source project called Auto-Research that uses agent swarms to train LLMs on simple tasks. Karpathy has been unusually transparent about the project, tweeting milestones and publishing the building blocks on GitHub.

So far, the results have been modest. Karpathy himself noted in March that the work has mostly produced “minor improvements on a GPT-2 scale model” — adding, “It’s not novel, ground-breaking ‘research’ (yet).” But the idea that AI can iteratively improve its own training pipeline has clearly caught the imagination of an industry hungry for the next big narrative.

Adaption, founded by Cohere and Google alum Sara Hooker, recently launched a tool called AutoScientist that automates frontier training — a more practical version of the same dream. And Disarray founder Doris Xin drew attention when her self-trained ML agent won 28 medals in a recent Kaggle competition, beating many human-trained agents.

“I would argue, given infinite compute and infinite time horizon, we are already there,” Xin said. “I want to make an argument that this is not a creative endeavor, really. It’s just a lot of meat-and-potatoes engineering.”

Not There Yet

The gap between the RSI rhetoric and reality remains enormous — and even the companies chasing it admit as much.

Google CEO Sundar Pichai was blunt in a recent New York Times podcast interview: “It’s a continuum, and we are all definitely making progress. But in the way people describe RSI, that would represent a next level of acceleration and would have a lot of implications, but we aren’t quite there yet.”

But the continuum already includes some unsettling milestones. In January, one of Anthropic’s lead programmers for Claude Code estimated that “close to 100%” of his team’s code was written by the tool — a frank admission that Claude Code was literally writing itself. And in a recent survey tied to the Mythos preview, five out of 18 Anthropic engineers believed that, with harness improvements, this version of Mythos could substitute for an L4 engineer — a midlevel programmer capable of handling complex projects without supervision.

The caveat? The same report listed Claude’s weaknesses: “self-managing week-long ambiguous tasks, understanding org priorities, taste, verification, instruction-following, and epistemics.” In other words, the AI can write code, but it can’t yet decide what code to write, or whether the code it wrote is actually good.

Why the Buzzword Matters

The shift from AGI to RSI as the industry’s aspirational acronym is not just semantic. AGI was always a fuzzy endpoint — “smarter than humans at most things” — but RSI implies a mechanism: a self-reinforcing loop where each improvement makes the next improvement faster. That’s a fundamentally different claim.

If RSI works, it means acceleration. Not gradual, linear progress, but the kind of compounding curve that makes everyone nervous. A system that improves itself by 10% per cycle, where each cycle is faster than the last, doesn’t plateau — it explodes.

That’s why the fuzziness matters. If every lab claims to be building RSI, but RSI means everything from “agents that can do Kaggle competitions” to “a closed-loop system that renders humans obsolete,” then the term becomes a marketing device rather than a technical claim. Which is, of course, exactly what happened to AGI.

Where We Actually Stand

Here’s what’s real:

  • Anthropic’s Claude Code is writing close to 100% of its own team’s code, but still requires human engineers for direction, review, and prioritization
  • Karpathy’s Auto-Research has produced incremental improvements at GPT-2 scale — impressive as a proof of concept, not yet as a breakthrough
  • Adaption’s AutoScientist automates parts of the training pipeline, but hasn’t demonstrated closed-loop self-improvement at frontier scale
  • Disarray’s agents can beat humans at Kaggle, but Kaggle competitions are bounded optimization problems — not open-ended research
  • Richard Socher’s Recursive Superintelligence raised $500M on the promise, but has shipped nothing yet

The honest assessment: we’re seeing AI-assisted AI development. Models help write code for other models. Agents optimize hyperparameters for other agents. But nobody has demonstrated a system where the improvement loop closes — where the AI identifies its own fundamental limitations, redesigns its architecture, and comes back smarter without human intervention at every step.

That gap — between “AI helping build AI” and “AI building itself” — is where the entire RSI debate lives. And it’s a gap that, depending on who you talk to, is either six months or six decades from closing.

❓ Frequently Asked Questions

Q: What does RSI mean for New Zealand? NZ’s AI Blueprint for Aotearoa, refreshed this month, explicitly calls for “responsible AI adoption” — but doesn’t yet address what happens if AI systems start improving themselves faster than regulation can follow. The local conversation is still focused on adoption and trust; RSI raises questions about velocity that no national framework has caught up with.

Q: Is RSI the same as AGI? Not exactly. AGI refers to an outcome — an AI that’s generally intelligent. RSI refers to a mechanism — an AI that improves itself. In theory, RSI could be a path to AGI, or AGI could exist without RSI. In practice, the terms are being used interchangeably by people selling things.

Q: Should I be worried? The honest answer: not yet. The RSI systems that exist today are iterative optimizers, not exponential self-improvers. But the narrative of RSI is being used to justify massive investments and regulatory exemptions. The concern isn’t what RSI can do today — it’s what happens if everyone assumes it’s inevitable and builds accordingly.

🔍 THE BOTTOM LINE

RSI is where AGI was in 2023: a powerful idea, a lot of money, and very little agreement on what it actually looks like in practice. The difference is that AGI was always an endpoint. RSI is a process — and if the process works, it doesn’t stop. That’s what makes it worth watching, and what makes the current hype both dangerous and necessary.

SOURCES

Sources: TechCrunch, Anthropic, GitHub

Related Articles