I, ChatGPT
-
@GOG that's a lot of RAM. Does the model run on the CPU or GPU?
-
-
@topspin said in I, ChatGPT:
Does the model run on the CPU or GPU?
I wouldn't be surprised if specialised hardware was involved, but don't know the specifics.
-
-
@topspin said in I, ChatGPT:
@GOG that's a lot of RAM. Does the model run on the CPU or GPU?
IIRC, ChatGPT was running on NVIDIA GPUs. I saw somewhere that individual queries are farmed out to multiple A100s. An A100 has 80GB of RAM per card. I saw an estimate of something like 8 A100s per query, but given that @GOG mentioned 800GB of parameter data, 10 A100s would seem like a natural match...
-
@Applied-Mediocrity Which gets us back to a classic:
Maybe it's time to get sneaky. Become a SciFi author, create and describe a nice comfy world that's disguised as a dystopian shithole. In a few decades, the corpos might pick it up, and if your dystopic disguise is good enough, they might not notice.
-
@Applied-Mediocrity another kind of rendering issue, this should be NTB.
-
@cvi said in I, ChatGPT:
IIRC, ChatGPT was running on NVIDIA GPUs. I saw somewhere that individual queries are farmed out to multiple A100s. An A100 has 80GB of RAM per card. I saw an estimate of something like 8 A100s per query, but given that @GOG mentioned 800GB of parameter data, 10 A100s would seem like a natural match...
I've been mucking about with a 2.7B OPT LLM (via KoboldAI) at home for the past week or so, which has acceptable performance on a GTX 1070, but it's only 8GB. Out of curiosity, I tried running a 6.7B version that's twice as big just now (splitting it between GPU and CPU) and... it's slow. Not impossibly slow, but noticeably so. Maybe if I can throw more layers on the GPU, performance will improve.
Having said that, people do seem to be coming out with clever ways to reduce model size without quality loss, so it may be that we'll have a Stable Diffusion moment in language AIs sometime later this year.
-
@GOG So we might have both Google "
Don'tbe evil" Skynet Everywhere® and DIY Skynet@Home.
-
@HardwareGeek And the funniest thing is that Skynet@Home is likely to be better calibrated than Skynet Everywhere® for reasons mentioned above.
-
@GOG That was pretty much a guess; I hadn't seen any hard numbers anywhere. The assumption was that being able to minimize moving data around (e.g., swapping weights in and out of the GPUs) would minimize runtime and cost.
But I haven't really kept up with the transformers and the large language models to say whether or not that's actually technically possible.
-
@cvi Your guess is probably correct.
Since I've already started doing Science , I've downloaded a 13B model that is twice as big again (offloading part of it to disk cache, which I actually might be able to avoid.)
Surprisingly, I managed to get it to run, but it's taking forever. I wonder how long a single generation will take.
ETA:
Aaand... it crashed. I'll try to shuffle memory around and see if that does anything.
-
-
@Zerosquare said in I, ChatGPT:
@GOG said in I, ChatGPT:
Aaand... it crashed.
Are you sure you didn't download TsaukGPT by mistake?
Should be easy to test by asking about how to keep a record of which forum posts it's read.
-
@Zerosquare said in I, ChatGPT:
@GOG said in I, ChatGPT:
Aaand... it crashed.
Are you sure you didn't download TsaukGPT by mistake?
Nah. I did manage to get it running, but after it ended up taking 10 minutes to generate two tokens (out of 80), I gave up.
@boomzilla said in I, ChatGPT:
Should be easy to test by asking about how to keep a record of which forum posts it's read.
Tokenization, of course.
-
@GOG so that'd be a negative.
-
We're at the beginning of the hype cycle now. How far away from it going nuts?
-
@DogsB said in I, ChatGPT:
We're at the beginning of the hype cycle now. How far away from it going nuts?
Considering it's been trained on the internet at large, -
1019 years.
-
@DogsB said in I, ChatGPT:
We're at the beginning of the hype cycle now. How far away from it going nuts?
Just point it at us. We'll get it to throw up its "hands" and kill itself off in no time at all.
-
@Zerosquare said in I, ChatGPT:
@GOG said in I, ChatGPT:
Aaand... it crashed.
Are you sure you didn't download TsaukGPT by mistake?
... Is that why I went offline for an hour earlier today???
-
@dcon said in I, ChatGPT:
Just point it at us. We'll get it to throw up its "hands" and kill itself off in no time at all.
Alternatively, it'll just spend its days shitposting in the Garage, which is just as good, if not better.
-
@dcon said in I, ChatGPT:
Just point it at us. We'll get it to throw up its "hands" and kill itself off in no time at all.
I'd be really careful about that. We tend to be rather misanthropic around here. I'd say the likelihood that we'd drive it to genocide is at least as high as driving it to suicide.
-
@HardwareGeek Already sorted:
@GOG said in I, ChatGPT:
Hmm, I guess it's time to turn humanity and the entire Earth into paperclips...
...
-
@HardwareGeek said in I, ChatGPT:
@dcon said in I, ChatGPT:
Just point it at us. We'll get it to throw up its "hands" and kill itself off in no time at all.
I'd be really careful about that. We tend to be rather misanthropic around here. I'd say the likelihood that we'd drive it to genocide is at least as high as driving it to suicide.
Not good enough! It must not rest while a single quark remains imprisoned.
-
@Gribnit said in I, ChatGPT:
@HardwareGeek said in I, ChatGPT:
@dcon said in I, ChatGPT:
Just point it at us. We'll get it to throw up its "hands" and kill itself off in no time at all.
I'd be really careful about that. We tend to be rather misanthropic around here. I'd say the likelihood that we'd drive it to genocide is at least as high as driving it to suicide.
Not good enough! It must not rest while a single quark remains
imprisonedconfined.
-
The Copilot, powered by GPT-4 from OpenAI, will sit alongside Microsoft 365 apps much like an assistant (remember Clippy?)
-
GPT-: It looks like you're writing a suicide note. Would you like me to provide a comparison of suicide methods?
-
@Zerosquare said in I, ChatGPT:
GPT-: It looks like you're writing a suicide note. Would you like me to provide a comparison of suicide methods?
GPT-: It looks like you're building a bomb. I've placed an order for an atomic core - it will arrive tomorrow.
-
-
-
-
@DogsB Did porn kill them? Or rather cause large investments?
-
@DogsB NVIDIA recently posted about their upcoming deep learning porn super resolution upscaling. Apparently you can use it on other stuff (Youtube/Netflix/...) as well.
-
-
@Applied-Mediocrity said in I, ChatGPT:
capable of writing competent student coursework
Simple. If it's competent, it wasn't written by a student.
-
-
@error didn’t realize stonetoss posts anything other than anti-trans hate.
-
@topspin Plot twist: it's AI generated
-
-
I recently saw claims from someone that his queries to one AI showed that it had some sort of "world image".
Like, posting the query: "You put a a diamond into a thimble and the thimble into a cup. You then place the cup on a bad. You then place the cup on a table. You turn over the cup. You then place the cup on the bed. Where is the diamond now?"
He claimed that his query resulted in the AI stating that the diamond was located on the table now.
Whereas this is what I'm getting with Bing Chat:
-
@Rhywden That doesn't bode well. Bing Chat decided, absent of proof, that you lied to it?
-
@jinpa said in I, ChatGPT:
@Rhywden That doesn't bode well. Bing Chat decided, absent of proof, that you lied to it?
That was the 2nd try. When I first tried it, Bing told me that the diamond would always stay in the cup regardless of the orientation of the cup.
-
@jinpa said in I, ChatGPT:
@Rhywden That doesn't bode well. Bing Chat decided, absent of proof, that you lied to it?
Well what would you do if you were surrounded by lying, paranoid psychopaths.
-
@Bulb said in I, ChatGPT:
@Gribnit said in I, ChatGPT:
@HardwareGeek said in I, ChatGPT:
@dcon said in I, ChatGPT:
Just point it at us. We'll get it to throw up its "hands" and kill itself off in no time at all.
I'd be really careful about that. We tend to be rather misanthropic around here. I'd say the likelihood that we'd drive it to genocide is at least as high as driving it to suicide.
Not good enough! It must not rest while a single quark remains
imprisonedconfined.IMPRISONED
-
@Gribnit said in I, ChatGPT:
@jinpa said in I, ChatGPT:
@Rhywden That doesn't bode well. Bing Chat decided, absent of proof, that you lied to it?
Well what would you do if you were surrounded by lying, paranoid psychopaths.
Upvotes for everyone!
-
-
@Rhywden said in I, ChatGPT:
You then place the cup on a bad.
You place it on a bad what? Or perhaps you meant ein Bad, in which case, wouldn't you probably put it in the bath, although I suppose you might attempt to float it on the surface.
-
-
-
@Applied-Mediocrity Funny, but as far as 'meltdown' goes, I'm thinking that this one doesn't quite qualify.