hwers

I agree, it can be incredibly frustrating at times. My rule is that if it “compiles” in my brain as an understood idea then i accept it. I also push back a lot (sometimes it points out good errors in my thinking, sometimes it admits it hallucinated). Real humans hallucinate a lot as well or confidently state subtly wrong ideas, it’s a good habit anyway. It’s basically the same approach when presented with a “formula” for something in school. If i dont know how to derive/prove it then i dont accept it as part of my memorized or accepted toolkit/things i use (and try to forget it). If it fits with the rest of my network of understood ideas i do. It’s annoying but still more time efficient than trawling through lecture slides with domain specific language etc

Its unfortunate that there’s mode collapse around what the consensus “best way” to use these things are. It’s too bad we didn’t have a period where these things were great teachers but didn’t attempt to write code because in my opinion the ideal way to use them is not by agents mass producing sloppy buggy disorganized code, but to teach you things way faster than the old alternatives, rubber duck, and occasionally write snippets of functions when your brain is too tired or it’s throwaway cli code or some api you’re not familiar with.

My conspiracy theory hat is that somehow investors with a stake in openai as well is sabotaging, like they did when kicking emad out of stabilityai

The default of making a public discord for your project/company always seemed like a bad idea anyway. It’ll always devolve into some drama or distracting overhead to moderate it

I definitely think they’ve nailed the personality better than others too. Gemini and grok are always paragraphs and paragraphs of text to sift through for something that with openai is usually digested to much less

I’ve moved to more closed source projects for this reason (just for the fun of coding rather than sharing). Though I suspect they still use private github repos in their deals to microsoft

I was in management i probably also wouldn’t like my designers to use AI. I pay them good money to draw original pieces and everyone can tell and it looks generic when AI is used. I’d want my moneys worth

You’re perfectly free to use it for private use, model output have been deemed public domain

68% felt like they were being watched but didn’t feel safe to admit so because they didn’t trust the report were truly anonymous

If price per compute keeps going down we effectively keep having moore’s law for parallel compute like GPUs. (Just get better at making ML models that don’t have as big of a communication bottleneck.)

Release it, fine. There’s been lots of fine tuned and continued trained SD models. Just don’t call it “1.5” which is the specific label for the model stability is training internally. Again the license ‘permits’ them to do it but seems like a very bad business decision since runway given what their service does would likely benefit hugely from early access to eg stability’s future text2video models (etc), which they now likely won’t get until everyone else (leaving someone else to possibly take market share in their field, and if this is the trade - because they got ‘impatient’ - that seems awfully not smart).

Well if that’s the case that’s still a pretty shitty thing to do on runways part. Just be curteous to what stability’s needs are, keep good business relations. Weird behaviour and I wouldn’t be surprised if in the future runway are silently excluded from before-public releases (which seems to be many in the years ahead).

Sounds to me like runway released it without consulting stability, called it “1.5” - which according to the license they’re allowed to do but pretty scammy since emad had hyped a model with that label. And now stability is deciding to call this the official release to be nice to runway and avoid a general PR thing and community infight.

Seems really useful in a wasm context

This is google, they for sure aren’t releasing the weights

I’m seeing a lot of R&D solely focused on giving them the chance to extract rent through future silly things like buying artificially scarce houses in fbs metaverse. I’m barely seeing any “giving back” type research like AT&T created with their research in the internet.

This is extremely well written

That this would have come out was totally predictable and expected to most in the know people in the ML world (we basically had proto versions of it summer 2021). Not really the unpredictable trajectory I associate with a singularity

Only downside with this is that each mesh takes like 5 hours to generate (on a v100 too). Obviously it’ll speed up but we’re far from the panacea

These things take months and months to train (hardly fast progress). Any new model that’s coming out is generally known in the atmosphere (not unpredictable) and these applications were pretty expected the day stable diffusion came out.

This sounds really interesting but I’m not sure I follow. Having a hard time expressing how I’m confused though (maybe its unfamiliar nerf terminology) but if you have the time I’d be very interested if you could reformulate this alternative method somehow (I’ve been stuck on this very issue for two days now trying to implement this myself).

Hey as a european where I doubt anything like this would ever happen maybe I should take the perspective on this as a good thing for me in the sense of a lot less future competition

Well they even have a “downlod model” so yep you definitely can. I wouldn’t think of this as an amazing panacea though, since once everyone has access to it that suddenly means whatever reason making assets like this was valuable before, will now be dirt cheap for all and thus actually net negative for people in that industry. Just saying and warning, not to be a bummer

I genuinely think stable diffusion is better than dalle. There’s a really obvious ugly artifact on almost all the dalle image’s I’ve seen that SD doesnt suffer from.

But anyway, SD is far superior even if you consider dalle better per image since you can create 1000 SD outputs and just pick the one you like best (which for sure will have one that’s better than the dalle output you got)

“They have literally floated the idea of whether murdering people to prevent AI-takeover is acceptable,”

Where? I probably believe you but it almost makes me worried about the well being of the stability ai founders (on a long term horizon)

Oh that discussion thats been going on in the last few day is so silly. The vae is hardly any kind of bottleneck. You barely gain anything by working with the latents directly

What’s interesting about datasets like this is that you can likely use it to distil an even more compressed SD generator from it.

People recently figured out how to export stable diffusion to onnx so it’ll be exciting to see some actual web UIs for it soon (via quantized models and tfjs/onnxruntime for web)

4.5*400 is a lot lower than 700 so it’s actually a power reduction (vs using multiple A100s).

Oh sorry I guess i need to mention that you need to put the text encoder on the cpu (or precompute the text embedding somehow). (Im using a custom codebase to make that possible idk how trivial that is to achieve with StableDiffusionPipeline.) Only the unet and vae should be on the gpu.

For your case with 8 gb you shouldn’t need to do either of those things (run it all on gpu), just make sure you have batch size 1 and are using the fp16 version.