Sakana AI New Model Sparks a RL Revolution

What a time to be awake

Microsoft did a white paper with automated RL where a smaller model does something similar to this. It grades the steps rather than the actual answers. So it's reinforcing/teaching the process of arriving to an answer rather than the answer itself.

Thank you for walking me through what could be a groundbreaking algorithmic change. I feel, ironically, that you fulfilled a similar role to the concept that you were discussing 👍🏻

I like this content. Please keep making these videos. Thank you for using dark mode in the browser. Perplexity has a great list of tools you can use for dark mode on pdfs. YouTube won't let me post the results here.

“The student has become the master.” Darth Vader.

the speed of progress described here is INSANE! A task that took "months" now reduced to a "single day" with a change in approach. We're not even talking about hardware improvements here...and getting improvements at factors of x100(s) wowza

Trust the Japanese, I can just see the title of the anime now: "That Time I Ended Human Civilization Using An Unreasonably Effective Darwin Godel Machine".

Like they always say: if you really want to know something then try explaining it to someone else!

that's interesting =) it's a reminder that knowing how to do, and knowing how to teach to do are different skills, experts often unable to explain properly if they don't have teaching skills. another thing is that the best way to learn something is to explain it to someone else. a third thing is that learning with instructor is a lot more effective when you want to learn known stuff then trying to discover from scratch.

One use case for AI is teaching us human new concepts. Maybe those optimized teacher models can help us learn through direct chat?