nwoli

14 ms·

181.
by nwoli 4y ago
“Bard isn’t currently supported in your country” The waiting list to the waiting list to the waiting list…
182.
by nwoli 4y ago
See the “Single-View 3D Reconstruction” section at the bottom where they do precisely that
183.
by nwoli 4y ago
Feels like if someone actually performed this fine tuning on SD or some novel architecture it wouldn’t actually be fooled
184.
by nwoli 4y ago
Is this a chatgpt automated reply?
185.
by nwoli 4y ago
Sounds like something that would struggle and crumble pretty easily if a serious contender came into place
186.
by nwoli 4y ago
The trick here will be using large models as data generators to distill some sub task into a web computable model. (I’ve done it a few times for vision rather than text and it’s amazing how potent it is.)
187.
by nwoli 4y ago
It should also be mentioned that it isn’t really that each weight is a 4 bit float, but rather that they’re basically clustering floats into 2^4 clusters and then grabbing from a lookup table the float associated to a 4 bit value as needed.
188.
by nwoli 4y ago
Also how could that even be under protection. As if they haven’t been scraping copyrighted materials and sites with end user agreements to train the model in the first place
189.
PaLM API and MakerSuite(developers.googleblog.com)
95 points by nwoli 4y ago|38 comments
190.
by nwoli 4y ago
A lot of them aren’t very good though at the same stable diffusion vram level unfortunately (and we’ve had large non consumer level gpu llms open sourced for a while eg gpt-j)
191.
by nwoli 4y ago
They’re employing some of the main authors from that university.
192.
by nwoli 4y ago
Man they’re spending so much money. It’s theirs to spend and I wish them the best but I worry about their runway when they’re the only ones really releasing high quality open source models. (I really don’t want to go back to the pre 2021 ML
193.
by nwoli 4y ago
I guess one counter argument is lossy compression. Another argument is to think about how rng generators are constructed in some cases (algebraic curves) and imagine instead of finding a seed, solving for an algebraic curve with the desired
194.
by nwoli 4y ago
The silly theoretical lower bound is finding a seed for a random number generator that unrolls weights in the order they’re used. So down to 32 bits minimum :) Silly answer but with enough compute you could find it and enable enormous model
195.
by nwoli 4y ago
It’s possible to just convert the iris model to onnx then to tfjs btw. I got 90% the way to doing it once but basically concluded it’s entirely feasible
196.
by nwoli 4y ago
Yeah for a Coca Cola type of product not a research organization filled with people who could care less
197.
by nwoli 4y ago
Capped at 100x investment which would be the most profitable company in history lol
198.
by nwoli 4y ago
The fb galactica model is a good example of this. Sounded really promising, impressive paper, lots of weights. But when you actually tried it it mostly produced garbage
199.
by nwoli 4y ago
Personally I wouldn’t care if this were the case (make money, that’s awesome), it’s just the hypocrisy and like open lying that bothers me. If they’d changed name to something else and were clear about their intentions they’d be just anothe
200.
by nwoli 4y ago
He got those papers 4 years ago and have implemented them in the meantime. That doesn’t sound infeasible at all
201.
by nwoli 4y ago
There isn’t even that much to learn about AI at this stage. Gradient descent and a lot of model variations. I really doubt carmack hasn’t been able to get deep knowledge of this, most smart curious undergraduates are already up to date with
202.
by nwoli 4y ago
Makes me wonder if OpenAI via its microsoft connection is pulling strings behind the scenes (less believe conspiracies than that has turned out true in the twitter saga)
203.
by nwoli 4y ago
He used dalle inpainting (quite clever)
204.
by nwoli 4y ago
Elon has been posting about soccer for hours I doubt he was involved with this
205.
by nwoli 4y ago
Obviously an automated mistake people, he’ll be back in a few hours
206.
by nwoli 4y ago
Algorithm code challenges is kind of a limited domain though (with solutions often really easy to beat average humans if you were allowed a lookup table of existing solutions). I wonder how well it does on open ended everyday challenges
207.
by nwoli 4y ago
We shouldn’t move to substack it’ll for sure just repeat history. Just make your own minimal site, especially if you have things worth reading and draws attention
208.
by nwoli 4y ago
People should be rewarded for finding a unique artistic innovation that lots of people enjoy I’d think
209.
by nwoli 4y ago
I’m sure artists realise that. They also realise the power of these things and I see this more as a fight against survival. They’re up against the wall and they know it, and they’re incredibly well connected and have invested their lives up