Hmm, yes. I agree that there’s something about self-guiding /self-correcting on complex lengthy open-ended tasks where current AIs seem at near-zero performance.
I do expect this to improve dramatically in the next 12 months. I think this current lack is more about limitations in the training regimes so far, rather than limitations in algorithms/architectures.
Contrast this with the challengingness of ARC-AGI, which seems like maybe an architecture weakness?
Hmm, yes. I agree that there’s something about self-guiding /self-correcting on complex lengthy open-ended tasks where current AIs seem at near-zero performance.
I do expect this to improve dramatically in the next 12 months. I think this current lack is more about limitations in the training regimes so far, rather than limitations in algorithms/architectures.
Contrast this with the challengingness of ARC-AGI, which seems like maybe an architecture weakness?