Recursive Self-Improvement
Computer, load up celery man.
Can AI build AI? Yes, and it already is. Sort of. I showcase the ability of AI agents like claude code to perform AI research, to build and optimize machine learning algorithms. I put various state-of-the-art LLMs like claude Mythos/Fable into an endless recursive research loop and have them build a neural network that learns the shape of the mandelbrot set. It is inspired by Andrej Karpathy's autoresearch. While we watch this loop, I express my thoughts on the concept of recursive self improvement, arguing that it is possible, hard, and dangerous.
Sorry for the bitrate issues.
Fractalsearch repo: https://github.com/MaxRobinsonTheGrea...
Neural Mandelbrot repo: https://github.com/MaxRobinsonTheGrea...
~SUPPORT ME~
Learn to code faster with Scrimba! (saves you 20% and support me): https://scrimba.com/?via=EmergentGarden
Patreon: / emergentgarden
Twitter: / max_romana
Bluesky: https://bsky.app/profile/emergentgard...
~SOURCES~
Autoresearch: https://github.com/karpathy/autoresearch
Mandelbrot Zoom: • Mandelbrot World Record Attempt - Part 1 (...
Celery Man: • Tim and Eric - Celery Man
Karpathy's Youtube: / @andrejkarpathy
Self-building Cranes: • How Tower Cranes Build Themselves
Darwin-Godel Machine: https://arxiv.org/abs/2505.22954
Hashgrid Paper: https://arxiv.org/abs/2201.05989
Anthropic's RSI Article: https://www.anthropic.com/institute/r...
Fable System Card: https://www-cdn.anthropic.com/d00db56...
My Music Guy: / @acolyte-compositions
"Equatorial Complex"
Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 3.0
http://creativecommons.org/licenses/b...
~TIMESTAMPS~
(0:00) Recursive Self Improvement
(3:14) fractalsearch
(9:56) RSI is Possible
(15:03) RSI is Hard
(21:52) RSI is Dangerous
(26:03) Results
(28:28) Cost
(29:29) Takeoff