Can A.I. get better at making itself better?

 Recursive self-improvement (R.S.I.) has gone from a nerdy, obscure concept to front-page news. That’s Bay Area slang for artificial intelligence using artificial intelligence to improve… artificial intelligence. Why does it have everyone so freaked out? And why is it linked so closely with fears of out-of-control superintelligent A.I.?

In part one of a two-part series about R.S.I., David Deming explains what R.S.I means, how A.I. models get trained, and what autonomous R&D looks like on the ground. He walks through the evidence from the frontier labs and academic researchers on how close we are to A.I. running its own research process.

He then also turns to the economics of R.S.I. A recent paper asks whether recursive self-improvement could cause A.I. capabilities to go vertical - what some people would call “takeoff.” David works through this argument and and tries to put together a speculative estimate of how close we might be. Spoiler alert, we're not there yet.

Part two will focus on the relationship between R.S.I. and alignment, which is the general challenge of getting the A.I. models to do what we want them to do.

For show notes and full episodes, subscribe to The Context Window with David Deming: ⁠https://thecontextwindowpodcast.substack.com