Graham Neubig pushed back on the idea of distinct recursive self-improvement companies after seeing a five-axis taxonomy in the arXiv paper AI & Human Co-Improvement for Safer Co-Superintelligence. The CMU researcher noted that nearly every LLM and coding-agent team already iterates across parameters, data, objectives, architecture, and code using their own models, leaving little room for a specialized RSI workflow that stands apart.
@gneubig In my view @jaseweston & @j_foerst laid out the spectrum well.
https://arxiv.org/abs/2512.05356
To be honest, I don't know what a "recursive self improvement company" is.
Every company building AI coding agents or LLMs is now using their own products to build their systems.
Aren't all AI coding agent/LLM companies recursive self improvement companies now?
Neubig observed that serious players are hitting all five learnable axes the paper maps out, so any new entrant claiming RSI as its edge would need to show concrete differences rather than just the label.
Without a unique mechanism beyond what existing providers already do with self-generated data and self-evaluation loops, specialized RSI firms risk looking like re-labeled versions of the current generation of agent builders.
No Digg Deeper questions have been answered for this story yet.
@_rockt @jaseweston @j_foerst Thanks, this is a great reference!
But what I wanted to say is that most LLM providers and serious coding agent companies are using AI to do all of those 5 things. So I'm not sure what an "RSI company" would do differently?
@gneubig In my view @jaseweston & @j_foerst laid out the spectrum well.
https://arxiv.org/abs/2512.05356
