It’s hard when it was trained on everything pretty much. That’s the same problem as with GPT3. In my mind it’s still brute forcing a solution but instead of endless computation it’s endless examples
Is this "bruteforcing" really different from what our brains do? We see thousands, millions of little things (examples) every day. Then we combine what we've seen into something new. Probably the only difference is that DALLE's training was done once while our brains are trained every day for 80+ years.
I would say yes it is. For humans there is plenty in the world that is completely novel, and we can (and have to) reason from abstractions or first principles. Being exposed to millions of every day things doesn’t give us intrinsic knowledge of writing fiction, algebra, artistic expression etc. instead, we have to apply abstract knowledge and reason. This may well be possible for an AI system, but I haven’t seen GPT3 or DALLE do this.
It’s hard to test this on DALLE or GPT3 as the internet is a summation of all known knowledge in effect. True unbounded Original thought is hard, and given they have seen everything before, it’s impossible to know if it’s original or just seen it previously. It would be interesting in a decades time to see how DALLE or GPT3 deals with novel ideas that it was never exposed to.
