> We never had technology which can write code like this. I prompted ChatGPT to write a very basic java tool which renders an image from an url and makes it bigger on a click. It just did it.
It's worth noting, that it can do things like that because of the large amount of "how to do simple things in java" tutorials there are on the internet.
Ask an AI to _make_ java, and it won't (and will continue to not) be able to.
That's the level that AI will fail at, when things aren't easily indexed from the internet and thus much harder / impossible to put into a training set.
I think the technology itself (transformers and other such statistical models) have exhausted most of their low hanging fruit by now.
Sora, for example, isn't a grand innovation in the way latent space models, word2vec, or transformers are, it's just a MUCH larger model than DALLE-3.
which is great! but still has the limits inherit to statistical models. They need the training data.
> It's worth noting, that it can do things like that because of the large amount of "how to do simple things in java" tutorials there are on the internet.
Much like the same points made elsewhere with regard to AI art: It cannot invent. It can remix, recombine, etc. but no AI model we have now is anywhere close to where it could create something entirely new that's not been seen before.
It's worth noting, that it can do things like that because of the large amount of "how to do simple things in java" tutorials there are on the internet.
Ask an AI to _make_ java, and it won't (and will continue to not) be able to.
That's the level that AI will fail at, when things aren't easily indexed from the internet and thus much harder / impossible to put into a training set.
I think the technology itself (transformers and other such statistical models) have exhausted most of their low hanging fruit by now.
Sora, for example, isn't a grand innovation in the way latent space models, word2vec, or transformers are, it's just a MUCH larger model than DALLE-3. which is great! but still has the limits inherit to statistical models. They need the training data.