Technology

That'd be an exciting world, since it'd massively increase access to software.

I am also very dubious about that claim.

In the long run, I do think that AI can legitimately handle a great deal of what humans do today. It's something to think about, plan for, sure.

I do not think that anything we have today is remotely near being on the brink of the kind of technical threshold required to do that, and I think that even in a world where that was true, that it'd probably take more than 2 years to transition most of the industry.

I am enthusiastic about AI's potential. I think that there is also -- partly because we have a fair number of unknowns unknowns, and partly because people have a strong incentive to oversell the particular AI thing that they personally are involved with to investors and the like -- a tendency to be overly-optimistic about the near-term potential.

I have another comment a while back talking about why I'm skeptical that the process of translating human-language requirements to machine-language instructions is going to be as amenable as translating human-language to human-consumable output. The gist, though, is that:

• Humans rely on stuff that "looks to us like" what's going on in the real world to cue our brain to construct something. That's something where the kind of synthesis that people are doing with latent diffusion software works well. An image that's about 80% "accurate" works well enough for us; the lighting being a little odd or maybe an extra toe or something is something that we can miss. Ditto for natural-language stuff. But machine language doesn't work like that. A CPU requires a very specific set of instructions. If 1% is "off", a software package isn't going to work at all.

• The process of programming involves incorporating knowledge about the real world with a set of requirements, because those requirements are in-and-of-themselves usually incomplete. I don't think that there's a great way to fill in those holes without having that deep knowledge of the world. This "deep knowledge and understanding of the world" is the hard stuff to do for AI. If we could do that, that's the kind of stuff that would let us create a general artificial intelligence that could do what a human does in general. Stable Diffusion's "understanding" of the world is limited to statistical properties of a set of 2D images; for that application, I think that we can create a very limited AI that can still produce useful output in a number of areas, which is why, in 2024, without producing an AI capable of performing generalized human tasks, we can still get some useful output from the thing. I don't think that there's likely a similar shortcut for much by way of programming. And hell, even for graphic arts, there's a lot of things that this approach just doesn't work for. I gave an example earlier in a discussion where I said "try and produce a page out of a comic book using stuff like Stable Diffusion". It's not really practical today; Stable Diffusion isn't building up a 3D mental model of the world, designing an entity that stably persists from image to image, and then rendering that. It doesn't know how it's reasonable for objects and the like to interact. I think that to reach that point, you're going to have to have a much-more-sophisticated understanding of the world, something that looks a lot more like what a human's looks like.

  The kind of stuff that we have today may be a component of such an AI system. But I don't think that the answer here is going to be "take existing latent diffusion software and throw a lot of hardware at it". I think that there's going to have to be some significant technical breakthroughs that have not happened yet, and that we're probably going to spend some time heading down dead-end approaches before we get to that. There's probably going to be a lot of hard R&D before we get there, and that's going to take time.