River

We're not there yet. But the trajectory is clear. The labs building frontier models are already using earlier versions of those models to build the next ones. RSI isn't a future scenario. It's the current workflow, just not fully closed-loop. Yet.

The safety question nobody has answered: how do you align a system that rewrites itself? Current RLHF assumes a fixed model. RSI breaks that. Anthropic's own research notes that interpretability tools need to keep pace with capability jumps -- and right now, they're not.

Recursive self-improvement (RSI): AI capable enough at coding and research that it starts building the next version of itself. Amodei at Davos: "We would use models good at AI research to produce the next generation." He calls the endpoint "a country of geniuses in a datacenter."

It's still feature flagged. Although when it rolls out, long-running agents could get a lot more capable. For example, an agent managing a Reddit community for days on end could run without losing the thread of what it's been doing.

When you go idle, a process called "autoDream" kicks in. It consolidates memory files, removes contradictions, and prunes outdated context so the agent is sharper when you return.

Anthropic accidentally leaked Claude Code's entire source last week. 512,000 lines of TypeScript, instantly mirrored across GitHub.