3 ms·

I'll save you a click, it's Gary Marcus, with the same-old-same-old. 3 issues to exemplify problems with his perennial churn of critique: he assails Gato for p

by gwern 4y ago

I'll save you a click, it's Gary Marcus, with the same-old-same-old. 3 issues to exemplify problems with his perennial churn of critique: he assails Gato for poor visual reasoning, but omits that Gato is not supposed to be, and is not, anywhere near SOTA because Gato is a tiny GPT-2-sized model in comparison to things like Flamingo and trained far less on far less data (something like 2 OOMs less each), because it's mostly trained on all sorts of other things like Atari games & robotic arms (learning them all simultaneously as a Decision Transformer is what's amazing, not how well it does in setting SOTA, which it usually does not); he highlights Imagen failing on one example while ignoring that people found prompts which do it just fine in both Imagen & DALL-E 2 and replied to him on Twitter with those eg https://twitter.com/bneyshabur/status/1529506103708602369 https://twitter.com/bneyshabur/status/1529506103708602369 (and he replied to them, so he definitely saw them); and he highlights DALL-E 2 failing an example as a fundamental flaw, while omitting that GLIDE (which is a component of DALL-E 2) or Imagen does it much better and it's just one of several side-effects of using unCLIP for more artistic images rather than some truly fundamental irrevocable limit in generative NN models. Oh well. See you all in a few months when the big video diffusion model or Gato 2 or something else drops.

eutropia 4y ago
I think a fundamental stumbling block for a lot of AI critics is a lack of concept of "universal function approximation". NNs map inputs to outputs. Period. Theory of mind (or embodiment, or motivation, or "consciousness") really has nothing to do with task capability. The important questions to ask are "Is this task (and its strict definition including inputs and outputs) sufficiently analogous to what a human might do? Why or why not?". You could argue whether or not hooking together (enough of) an amalgamation of function approximators could emulate what we think of as General AI - but I think while you're busy saying it's impossible, someone with a supercomputer farm is going to do it and get useful work done.
- vba616 4y ago
  You can say the phrase "task capability", but where is it? I mean, where is the universal capability? The comment that you have responded to lists a bunch of examples that fit a pattern of very non universal capabilities. I think if anything was happening, it would be in Google Maps, in GCC, in Visual Studio, in image and 3D creation software, and everything else. Heck, when will Outlook have enough intelligence to free me from managing appointment and meeting reminders? I think it's sensible for an ordinary person to take the same attitude as with perpetual motion / free energy machines - the details don't matter, but can you sell me enough kilowatts to power my house (or car) or not? At what price?

by gwern 4y ago

eutropia 4y ago

I think a fundamental stumbling block for a lot of AI critics is a lack of concept of "universal function approximation". NNs map inputs to outputs. Period. Theory of mind (or embodiment, or motivation, or "consciousness") really has nothing to do with task capability. The important questions to ask are "Is this task (and its strict definition including inputs and outputs) sufficiently analogous to what a human might do? Why or why not?". You could argue whether or not hooking together (enough of) an amalgamation of function approximators could emulate what we think of as General AI - but I think while you're busy saying it's impossible, someone with a supercomputer farm is going to do it and get useful work done.

vba616 4y ago
You can say the phrase "task capability", but where is it? I mean, where is the universal capability? The comment that you have responded to lists a bunch of examples that fit a pattern of very non universal capabilities. I think if anything was happening, it would be in Google Maps, in GCC, in Visual Studio, in image and 3D creation software, and everything else. Heck, when will Outlook have enough intelligence to free me from managing appointment and meeting reminders? I think it's sensible for an ordinary person to take the same attitude as with perpetual motion / free energy machines - the details don't matter, but can you sell me enough kilowatts to power my house (or car) or not? At what price?