source: techcrunch ai: what happens when ai starts building itself?
level: technical
recursive superintelligence, a san francisco startup, came out of stealth with $650 million in funding. it was founded by richard socher, known for you.com and imagenet work, along with peter norvig and tim shi. the company aims to build a recursively self-improving ai model that can spot its own weaknesses and fix them on its own, without people. this idea of ai improving itself is a long-sought goal in ai research.
the team uses a method called open-endedness to reach recursive self-improvement. this approach lets ai systems co-evolve, like animals adapting in nature. for example, one ai can attack another to find flaws, and the defender learns to become safer over millions of rounds. this rainbow teaming technique is already used in major labs. the startup believes this focus on open-endedness sets it apart from others working on similar goals.
socher says the company plans to ship products in quarters, not years, and wants to be more than a research lab. he notes that once recursive self-improvement works, compute becomes the key resource. the faster you run the system, the faster it improves. this could shift the focus to deciding how much compute to use for solving problems like diseases. the startup's backers include greycroft and gv.
why it matters: recursive self-improvement could speed up ai progress dramatically, changing how resources are used in ai development.
source: techcrunch ai: what happens when ai starts building itself?