Don’t get me wrong, anything that is close to done already they excel at and can combine existing techniques. I’m talking about new ideas models have never seen before.
Example I have this hobby project that push what’s possible with route optimization. Yes it’s close to SOTA and way more efficient than all (?) other solutions out there (punnerud.github.io/mpee/), but I have to hold the model in the hand and brainstorm ideas on how to compress a matrix.
And it’s just not a one time thing, happens like 40-50 times in few days.
The 1% there is this “new ideas” part. Why can I come up with all these, and not the model? A really hard reval to create. Now this project is open, later I am thinking about making a frontier project in the same way, keeping it away from the public and using it as a benchmark. It’s that the best way to test for new ideas in models?
discordance•40m ago
— Major Motoko Kusanagi, Section 9 field commander
punnerud•9m ago