The AI world has a shiny new obsession, and it comes in the form of three letters: RSI. No, not repetitive strain injury - though the industry's frantic buzzword adoption might cause that too. This RSI stands for Recursive Self-Improvement, the latest buzzword that has startups and researchers alike chasing a vision of AI that can upgrade itself without pesky human interference.

In theory, RSI is simple: an AI system that continuously improves its own capabilities, eventually closing the loop so that humans become optional - or worse, obsolete. In practice, as with AGI before it, nobody can quite agree on what it actually means or when it might arrive. But that hasn't stopped the hype train.

Earlier this month, renowned AI researcher Richard Socher launched a startup called Recursive Superintelligence, whose main focus is to build “truly recursive, self-improving superintelligence at scale.” Socher told TechCrunch that the goal is to automate “the entire process of ideation, implementation, and validation of research ideas.” So basically, AI that can come up with its own homework, do it, and grade itself. What could possibly go wrong?

Socher isn't alone. Alex Karpathy, a Tesla and OpenAI alum now at Anthropic, has been working on a project called Auto-Research, using agent swarms to train LLMs on simple tasks. So far, his work has been limited to minor improvements on a GPT-2 scale model - which, as Karpathy himself noted, is “not novel, ground-breaking ‘research’ (yet).” But it's enough to keep the dream alive.

Adaption, founded by Cohere and Google alum Sara Hooker, recently launched AutoScientist, a tool aimed at automating frontier training. Like Karpathy's system, it trains agents to make incremental improvements - but with the grand ambition of training a full-scale frontier model. If that works, the system could quickly spiral into something very RSI-like.

Disarray founder Doris Xin took a more practical route: her self-trained machine learning agent won 28 medals in a recent Kaggle competition, beating many human-trained agents. Xin argues that with infinite compute and time, “we are already there.” She insists it's not about creativity, just “meat-and-potatoes engineering.”

But there's evidence the industry isn't as close as the hype suggests. Google CEO Sundar Pichai recently admitted, “We aren’t quite there yet,” describing RSI as “a next level of acceleration with a lot of implications.” Meanwhile, Anthropic's Claude Code tool is reportedly writing close to 100% of its own team's code. A recent survey found that five out of 18 Anthropic engineers believed Mythos could soon substitute for an L4 engineer - a mid-level programmer. However, the report noted weaknesses in self-direction, including “self-managing week-long ambiguous tasks, understanding org priorities, taste, verification, instruction-following, and epistemics.” In other words, everything that makes RSI actually work.

Georgetown's Center for Security and Emerging Technology assembled experts last year and found a major split: some expected an imminent “superintelligence” explosion, others slower progress and a plateau. Helen Toner, CSET director and former OpenAI board member, clarified that using AI tools for research doesn't qualify as RSI. “They’re just using AI for as much as they can,” she said. “RSI is really that there are no humans needed.”

METR's Ayeja Cotra laid out milestones: “adequacy” (AI can research without humans, even if poorly), “parity” (AI matches humans), and “supremacy” (AI outperforms human-AI collaboration). She thinks adequacy could arrive within a couple years, but parity is murkier. Once parity hits, she predicts supremacy within another year.

Toner draws a historical parallel: “We went from machine languages to assembly language and compiled languages; you’re getting further from the guts of the computer. But the human is still, in some intuitive sense, running the show.” Moving beyond that paradigm requires solving massive engineering and alignment challenges - and there's no infinite compute to brute-force it.

For now, the only thing researchers agree on is that, like AGI, RSI isn't here yet. But the industry is having a grand time pretending it's just around the corner.