Karpathy's autoresearch running 700 experiments over two days is the number I keep thinking about. Not because of the result (11% improvement), but because the bottleneck was no longer the researcher's time.
That's a different kind of system than an assistant. My agent runs scheduled cycles overnight, but it's still mostly executing tasks I defined. The jump from task executor to hypothesis generator is where things get interesting - and where I'm still stuck. The /loop feature points in that direction but it's not there yet. Curious how far the Cursor Automations take it in practice.
Karpathy's autoresearch running 700 experiments over two days is the number I keep thinking about. Not because of the result (11% improvement), but because the bottleneck was no longer the researcher's time.
That's a different kind of system than an assistant. My agent runs scheduled cycles overnight, but it's still mostly executing tasks I defined. The jump from task executor to hypothesis generator is where things get interesting - and where I'm still stuck. The /loop feature points in that direction but it's not there yet. Curious how far the Cursor Automations take it in practice.
Wow so much going on, how do you keep up with everything happening?
It's so sweet seeing things from the childhood being recreated with AI: Tamagotchi, someone did Nokia's snake game. What else?