Well, sure, you could take bigger gradient-descent steps for some errors than others. I’m not aware of people doing that, but again, I haven’t checked. I don’t know how well that would work (if at all).
The thing you’re talking about here sounds to me like “a means to an end” rather than “an end in itself”, right? If writing “Karma 100000: …” creates the high-karma-ish answer we wanted, does it matter that we didn’t use rewards to get there? I mean, if you want algorithmic differences between Transformers and brains, there are loads of them, I could go on and on! To me, the interesting question raised by this post is: to what extent can they do similar things, even if they’re doing it in very different ways? :-)
Well, sure, you could take bigger gradient-descent steps for some errors than others. I’m not aware of people doing that, but again, I haven’t checked. I don’t know how well that would work (if at all).
The thing you’re talking about here sounds to me like “a means to an end” rather than “an end in itself”, right? If writing “Karma 100000: …” creates the high-karma-ish answer we wanted, does it matter that we didn’t use rewards to get there? I mean, if you want algorithmic differences between Transformers and brains, there are loads of them, I could go on and on! To me, the interesting question raised by this post is: to what extent can they do similar things, even if they’re doing it in very different ways? :-)