@dunik_7: An AI more than doubled its own coding ability while the researchers just watched. 20% -> 50% on SWE-bench. They never …

An AI more than doubled its own coding ability while the researchers just watched. 20% -> 50% on SWE-bench. They never touched it. It pulled that off by rewriting its own source code. That's the whole paper the Darwin Gödel Machine, 72 pages out of Jeff Clune's lab. An agent finally pointed at itself. / it reads its own code and edits one piece a tool, a retry rule, a prompt / it runs the new version on real coding tasks and checks if the score actually moved / the variants that win get archived and branched, like evolution, so it never dead-ends / then it runs the whole thing again, on the agent that just improved Everyone's still arguing about which model is smartest. This quietly skips the question. Same model, same weights the agent around it gets sharper every pass, by its own hand. Loop 4 in my breakdown was the one people called sci-fi. Here it is, benchmarked, with a public repo. The unsettling part isn't that it worked. It's that nobody needed to be in the room. Paper's below. Read it before it improves itself again.