Superposition code is a well known concept in information theory - I think there is certainly more to the story then described in the current works, but it does feel like they are going in the right direction
there is no source it is pure conjecture, but I would say that there are many many fine tunes available of various image generation models so it is clearly possible to make many styles. thus it must be a conscious choice on the part of a API provider to render by default to a distinct style. there are many plausible reasons they would want to do this. Surely, but I don't have actual evidence from the internal management processes of these organizations that they were doing this for one reason or another.
Most of these bizarre properties can be traced to the mixed valent character of plutonium - that is the electrons in the partially filled 5f shell of the plutonium atom hybridize in the solid phase with the valence electrons, and thus for some properties they are valence like and others they are core like. This is quite similar to some lanthanide (4f) elements like cerium, which also show strange variety of allotropes. The mixed valent character causes plutonium to have a variety of exotic physics effects, completely unrelated to it's most notorious use in nuclear weapons. Examples of these include heavy fermion metals, novel superconductivity, etc..
how is thermodynamic linear algebra different from monte carlo algorithms? I'm willing to believe that there are efficiency gains to be made by exploring the time/energy trade-off, but i bet they have a long long way to go before they are better than general purpose compute (CPU/GPU) running well designed monte-carlo algos...
well said, I would note that both sides recognize that "AGI" will require new uncertain R&D breakthroughs beyond merely scaling up another order of magnitude in compute. given this, i think it's crazy to blow the resources of azure on trying more scale. rapid commercialization at least buys more time for the needed R&D breakthrough to happen.
I just want to say how much I love your product. So many of the best things never make it financially, but I really really hope you all make it. I'm 100% willing to support obsidian financially, just need to know where and how to do that. I don't want an account or a sync thing, I just wanna give you all money for the great product.
I've played around with these mesh inference methods, and while I certainly value and respect the progress that has been made by people working in this field - Something feels fundamentally off about it - as if they aren't being quite ambitious enough.
Consider shape-net - while being able to generate samples from the shape-net distribution would be a very impressive achievement, and these would certainly be useful as "assets" (decorations in a game scene), they don't have any of the real structure of these objects that would allow real interaction. They are effectively just inflatable decoys compared to real objects.
In short, "AI" isn't going to replace the artists, it's going to massively amplify their productivity and we don't know of any limit to consumer demand for realism and rich interactivity in games. If anything the number of game artists is going to increase, only the bar will be raised on the quality of their product.