4 ms·

Pretty cool. Although it's interesting that it can't seem to render an image from a precise description that should have something like an objectively correct

Pretty cool. Although it's interesting that it can't seem to render an image from a precise description that should have something like an objectively correct answer. I tried prompts like "Middle C engraved on a staff with a treble clef" and "An SN74LS173 integrated circuit chip on a breadboard" both of which came back with images that were nowhere close to something I'd call accurate. I don't mean to detract from the impressiveness of this work. But I wanted a sense of how much of a "threat" this tech is to jobs or to skills that we normally think of as being human. Based on what I'm seeing, I'd say it's still got a ways to go before it's going to destroy any jobs. In its current form, it mostly seems like a fun way to generate logos or images where the exact details of the content don't matter.

nerdponx 4y ago
I am generally of the "it's not threatening yet and won't be for a while" camp, but in this particular case it's probably just for lack of trying. These algorithms are essentially enormous pattern-matching engines, so given enough data and some task-specific engineering effort, I wouldn't be surprised if you could build an "AI" circuit designer, like Copilot but for electronics instead of code. Next-level autorouting would be cool, but it's still not going to put the electrical engineering field out of business.
- davesque 4y ago
  Sure, but this model is a general image synthesizer that is trained on a massive amount of data that is openly available on the internet. Given that, I would assume that it would have seen many images of 74xx series chips and also musical notation. So I would think that the most "likely" image to generate would involve either a chip with 8 pins on either side (for the 74ls173) or a note on a single leger line just below a staff with a treble clef. I imagine there must be hundreds of 74xx chip images that would establish that fact and also thousands of images of musical notation that would establish the other. I guess the takeaway really is that the model does not function in such a way that it can recall its training data. Which is fine. I don't think I should expect it to. On the other hand, GPT-3 can be made to produce specific facts that are established by its training data. Although, admittedly it often gets things wrong. Maybe the problem of image modeling is just naturally harder than language modeling. After all, language already "directly" represents meaning in some sense much more than arbitrary images do. I'm sure someone could design a targeted model that would solve the issues I'm talking about. But I feel as though they shouldn't have to if we really had something that sees the world the way humans do. In any case, this work definitely seems cutting edge and represents a huge leap in that direction.

nerdponx 4y ago

I am generally of the "it's not threatening yet and won't be for a while" camp, but in this particular case it's probably just for lack of trying. These algorithms are essentially enormous pattern-matching engines, so given enough data and some task-specific engineering effort, I wouldn't be surprised if you could build an "AI" circuit designer, like Copilot but for electronics instead of code. Next-level autorouting would be cool, but it's still not going to put the electrical engineering field out of business.

davesque 4y ago
Sure, but this model is a general image synthesizer that is trained on a massive amount of data that is openly available on the internet. Given that, I would assume that it would have seen many images of 74xx series chips and also musical notation. So I would think that the most "likely" image to generate would involve either a chip with 8 pins on either side (for the 74ls173) or a note on a single leger line just below a staff with a treble clef. I imagine there must be hundreds of 74xx chip images that would establish that fact and also thousands of images of musical notation that would establish the other. I guess the takeaway really is that the model does not function in such a way that it can recall its training data. Which is fine. I don't think I should expect it to. On the other hand, GPT-3 can be made to produce specific facts that are established by its training data. Although, admittedly it often gets things wrong. Maybe the problem of image modeling is just naturally harder than language modeling. After all, language already "directly" represents meaning in some sense much more than arbitrary images do. I'm sure someone could design a targeted model that would solve the issues I'm talking about. But I feel as though they shouldn't have to if we really had something that sees the world the way humans do. In any case, this work definitely seems cutting edge and represents a huge leap in that direction.