Why AI Researchers Are Calling RSI the Next Big Thing (And Why It’s Still Elusive)

A new buzzword is sweeping the AI world: RSI, or “Recursive Self‑Improvement.” Many tech insiders say it could be the stepping stone to true artificial general intelligence (AGI), but just like the early days of AGI, it’s hard to pin down. The most vocal champion right now is Alex Karpathy, a former Tesla and OpenAI star who’s now experimenting with “agent swarms” – tiny AI helpers that work together to teach large language models (LLMs) simple tasks. He calls the effort “Auto‑Research” and has been unusually transparent, posting progress updates on Twitter and sharing code on a public GitHub repository. So far the experiments have nudged a GPT‑2‑size model a little better, which Karpathy modestly admits isn’t groundbreaking yet. Still, the modest gains have sparked a wave of interest, with dozens of labs chasing the same RSI dream. Now that Karpathy has moved to Anthropic, he’ll have access to far more computing power and data, giving the project a chance to scale up dramatically. In short, RSI is the newest hype in AI, promising a path toward smarter machines, but it remains a work in progress that researchers are still trying to fully understand.

Read more