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.
It's industry standard procedure to tune your model to output a consistent distinct style, to prevent malicious actors from abusing it and presenting fabricated (but very convincing) images as real.
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..
Is it specifically the emissivity that's so unusual and makes it the preferred material? I don't know that "low emissivity in infrared + high emissivity in visible" is a rare property—I recall reading about space satellite thermal engineering and seeing long lists of common materials sorted by visible/infrared emissivity ratios [0]. And there are a lot of them, in every category. I suspect the key thing is ThO2 is a super-refractory with a melting point of (approximately—these are hard to measure!) 3,350° C [1]. (About the same as the tungsten filaments in the old-school type of lightbulbs—something with pretty similar considerations). I suspect "things that maintian structural integrity in hot gas flame" is really the key discriminator here, the rare property that prunes out most candidates.
I'm not any sort of expert on this, to make very clear! Just a curious geek.
[0] (That's basically a proxy for the radiative equilibrium temperature in space: visible emissivity measuring absorption of sunlight, infrared emissivity measuring emission of waste heat. (To those unfamiliar, absorption and emission are exactly the same, at a specified wavelength: the physics is reversible). ThO2 for example, you'd expect would get extremely hot in space).
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...
I think the benefit here is that noise is not something you need to calculate with eg a seed, but rather the device can focus on the actual computations required.
Only because there is under-supply of housing at the moment. If we built more houses then landlords wouldn't be able to charge higher rent just because people had more money.
To be clear, I think that we need to address both of these issues: general wealth distribution and lack of affordable housing.
A majority of the population's landlord is the bank that holds their mortgage, and they don't just raise your mortgage payment because your income increases.
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.
do we really know that scaling compute an order of magnitude won't at least get us close? what other "simple" techniques might actually work with that kind of compute ? at least i was a bit surprised by these first sparks, that seemingly was a matter of enough compute.
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.
