ARC-AGI-4 - We created the test for AGI, whether they actually control the world economy
“Humans currently control 98% of the world, AI <2% …. Most benchmarks tests whether models can learn and do tasks autonomously, we test whether they control tasks.”
e.g the things that become benchmarks are often the things that become progress (or what gets measured gets managed) and I very much do not think that autonomous agents that are able to learn faster than humans is a good thing for the empowerment status of humans.
2 years later...
ARC-AGI-4 - We created the test for AGI, whether they actually control the world economy
“Humans currently control 98% of the world, AI <2% …. Most benchmarks tests whether models can learn and do tasks autonomously, we test whether they control tasks.”
e.g the things that become benchmarks are often the things that become progress (or what gets measured gets managed) and I very much do not think that autonomous agents that are able to learn faster than humans is a good thing for the empowerment status of humans.
I agree it’s probably bad, but I still want to understand it, and I still want LW to understand it.