1,775 karma · joined January 18, 2019
After you link it up with the apparatus, it is pulled into the system as a new part of it, and so on. The more stuff you pull into the system, the less number of different configurations you sum over, eventually you end up with a single configuration with an observer that has a clear measurement result in her mind. This process is what is called "wavefunction collapse" in old physics texts (and a lot of modern pop-sci accounts).
What sets this apart from other purely probabilistic theories is that it's non-local, the entire linked system from the proton to the measurement apparatus to the observer has to be taken into account in the calculations if you are to be thorough. In local theories you can separate the parts of the system and handle them separately, like you could say "there is a 1% probability that a charm quark pops out in the upper left part of the proton every second" etc. You can't really do this in QM/QFT, and this is what in essence leads to all of the counter-intuitive results and confusion..
What they mean by the "haze of probabilities" is that you need to iterate over all possible configurations of the gluons and quarks in the proton to produce experimental predictions.
It is for sure computationally ludicrous, but conceptually it's really not better or worse than to say that particles are billiard balls moving around by Newton's Equations. You're just more used to the latter, but spend some time working with for example Lattice QCD and you get completely used to the former being the "actual" underlying physics rules.
Add to that a company environment that seems to be built on money-crazed stock option piling engineers and a CEO that seems to have gotten power-crazed.. I mean they grew far too fast I guess..
Text-prompting is just a very coarse tool to quickly get some base to stand on, ControlNet is where the human creativity again enters.
This allows the rest of the network to be smaller while still generating a usable output resolution, so it's a performance "hack".
It's a really good idea to explore it and hack into it like in the article, to "remaster" the image so to speak!
If you quantize those 4 floats per 8x8 block, is that encoding better than say the old venerable JPG 8x8 DCT + quant?
Of course if the entire industry is corrupt it doesn't work, but to a certain degree I guess it is robust. This time it led to a potential disaster but it will become safer as a result..
But the main problem is probably not reading thrown out drives - it's that stuff is just too transient nowadays. People put something up on the net and decides to withdraw it next year. archive.org can't store the walled gardens, and even if a social network for example wants to archive stuff they might not be allowed to for legal issues (not that it has prevented them before, but anyway...)
I agree fully with the OS issues, but for the purposes of gaming, it turned out to not be an issue really I guess..
This was ages ago, but as far as I can recall, the main obvious limitation of the AGA chipset compared to the PC cards was 1 byte-per-pixel modes (I think they call it "chunky" modes now) so the simple texture-mapping 3D loops you could hand-code on the x86 were a pain on the Amiga. I'm sure now 30 years later people have figured out cool hacks to do this better but back then it didn't work, and adding to this was the fact that the PC cards were improved like every 6 months. Commodore went bankrupt 1 year after I bailed, and I didn't look back.. (though I did miss AmigaDOS real multitasking for some PC/Mac generations after :).
The A1000 I had was a great machine in 1987-1990 but somewhere after that the whole Amiga concept started to feel dated (IMO). The PC architecture won because it was essentially hardware open-source, and hundreds of vendors could add features to the platform together. Commodore would have had to do every single decision spot on correct for years in order to stay competitive as a single company. Look at what happened when IBM tried to tame the beast and released the PS/2 that was "closed source"..
Remember TSENG LABS? :)
I also think I saw Ultima Underworld or some of the variants on PC before this, rendering full free-view 3D in a sub-window at least. Was really blown away by that, this was before Doom came out.
But then again, it helps the entire industry that it apparently has reached a different audience this time who gets hyped again.
I just wish there could be new use-cases than just the same old recycled use-cases like virtual monitors, watching movies, looking at how it is to stand on the moon, or virtual conferencing. Maybe there really are no killer VR apps, except beat saber :/
Wasn't AVP supposed to be an AR headset? Or they just haven't really figured out what to do there? Now it's just a VR headset with passthrough, it isn't really "using" the external world.
The next thing we did was make a version of our CPU with an MMU, designed to work optimally with Linux (the first version was on the uClinux concept, with a kernel without MMU support and user-space programs that couldn't rely on fork() or mmap() fully). After a year or 2 with MMU-less Linux, it was like heaven to be able to run on an MMU :)
Beyond Surface Statistics: Scene Representations in a Latent Diffusion Model
"... In this work, we investigate a basic interpretability question: does an LDM create and use an internal representation of simple scene geometry? Using linear probes, we find evidence that the internal activations of the LDM encode linear representations of both 3D depth data and a salient-object / background distinction. These representations appear surprisingly early in the denoising process−well before a human can easily make sense of the noisy images. ..."
The models will generalize (because that's the most efficient way of storing concepts) and you can make an argument that that means they understand a concept. Claiming "it's not learning concepts, only statistical probabilities" trivialises what a modern neural network with billions of parameters and dozens of layers is capable of doing. If a model learns how to put a concept like line width 5, 10 and 15 pixels into a continuous internal latent property, you can probably go outside this at inference at least partially.
I would argue that improving this is at this point more about engineering and less about some underlying unreconcilable differences. At the very least we learn a lot about what exactly generalization and learning means.
I've seen "infographics" spread by anti-AI people (or just attention-seekers) on Twitter that tries to "explain" that AI image generators blend together existing images, which is simply not true..
It is however the case that different AI models (and the brain) generalize a bit differently. That is probably the case between different humans too. Not the least with for example like you say those with photographic memory, autists etc.
What you call creativity in humans is just noise in combination with a boatload of exposure to multi-modal training data. Both aspects are already in the modern diffusion models. I would however ascribe a big edge in humans to what you normally call "the creative process" which can be much richer, like a process where you figure out what you lack to produce a work, go out and learn something new and specific, talk with your peers, listen to more noise.. stuff like that seems (currently) more difficult for AIs, though I guess plugins that do more iterative stuff like chatgpt's new plugins will appear in media generators as well eventually..
So essentially you can ask chatgpt to fetch and summarize a paywalled article, because openai has subscribed their crawler?
I think that logic ties in with the theme of the article as well. Supply and demand regulates markets in a subtle way, it's not a binary input like "people want to live in SF or not". Maybe it's enough that 1/10 of the people who wants to move there changes their minds when they see the housing market.