Creative AI progresses slowly because its results are hard to verify

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 28:53

   that human creativity is basically connecting dots in the unconscious that you've had seen or experienced all of these things in your life or in your set of inputs sensory inputs and from those inputs you connect dots and then you extrapolate a forward view of things that hasn't perhaps been seen in the world before do you agree with that view on creativity and is it distinct from actually what ai does and again i want to go back to the uniqueness of the human in an ai era well i i think that um one thing that happens with humans is they're always seeking novelty like our reward function is different than like an ai's also our loss function like if i was to open this bottle of water and splash it all over you and you got bad at me like you bet i'm not going to do that again but in ai like you know with rl you gotta like really tell it a bunch of times uh uh that

2. 29:54

   no you don't do that um robots stop kicking me yeah well and maybe that'll change with architectures over time but like overall the brain is different uh than the current systems that we've made with ai and i i think that uh it's unknown we don't know much about neuroscience uh overall i'd love to get to the point where like we all know more about neuroscience but something is fundamentally different there when it comes to creativity and i also think design is different than creativity it's not just creativity it's also problem solving like i mentioned earlier and the combination of the two is i i just find it fascinating because you have a user you're trying to solve their needs you're trying to do so in a way that stands out as creative and it connects with them and also like you have to connect to wider systems the marketing

3. 30:47

   brand the culture the business problems uh and so on and and like holding all that in your head at different levels of abstraction while trying to like deliver a screen or a flow uh that like from a search-based standpoint can be a commentorial explosion is just pretty cool it's pretty fascinating i'm actually personally really excited about the idea of reinventing existing art form and media where ai can enable more dynamic generation where the creator steps up a level sets a bunch of parameters but doesn't actually necessarily render the final product but the final product sort of like in a video game is rendered on the fly with the user in front of it but the creator the creative team behind a video game for example is building all of the objects They're building the storyline, they're building the narrative flow, and the realization of

4. 31:41

   that is much more dynamic, but I feel like it's going to find its way, it's going to leach its way into many more different media. We have products around that too, like in Figment Design, you can have generative tools and shaders where you can do very much that, or we have a product called Weave, and Weave is basically a node-based. You can connect model outputs into other inputs and such. And the whole idea is like, it's not bad to generate with AI, but you need a way to craft that clay. You need a way to mold it. And with Weave, people do like the most amazing video and video work. And overall, just like the outcomes that are achievable with it are incredible.

5. 32:26

   uh so yeah i think that if you've got the intention if you actually care to do the work um you can always get to a better result and people you know new model drops there's like this ai psychosis moment that people are in for a few days and it's like i think also aided by algo feeds totally the best results rise to the top everyone's like this model is everything yeah And then a few days later, it's, it's okay. Okay. I know the kind of like main tropes of the model. Um, and you adapt fast, like the novelty is gone. It no longer is a thing that's, that's kind of like rewarding you or firing off your dopamine system or whatever. And instead you're like, okay, well, yeah, it always does that thing. How do I move past it?

6. 33:13

   And that seems to be the continued loop we're in, which is sort of fascinating because models keep getting so good when it comes to these formal domains, math, cybersecurity, IQ points just stack and stack and stack. And then for creativity and for design, it's very incremental progress overall. Maybe that will change at some point, but I think the fact that these are non-verifiable domains has something to do with it.

7. 33:42

   And it's probably the fact that maybe that is the human role in the AI toolkit, which is conceive, create, design, and let the AI do more of the work, but it gives you the ability to do much more work.