data stringlengths 115 7.61k |
|---|
Aran Komatsuzaki#5714: not perfect robustness or anything
bmk#1476: GPT3 is already more robust than many people
Daj#7482: I just wanted to make a meme at humans
Aran Komatsuzaki#5714: ok got it
Daj#7482: Don't get me wrong, some of my best friends are human
Ravna#1831: The "algorithm improvements" may as well just be ... |
Daj#7482: It's provably impossible to have an agent that performs better than random on average at every possible universe
goolulusaurs#1571: There has been some pretty amazing stuff using nns with fixed random weights, e.g. reservoir computing.
Ravna#1831: Yes but NNs have some inherent priors and that would hopefully... |
bmk#1476: lol
goolulusaurs#1571: I've also done some stuff with evolving masks on the weights of a randomly weighted network. Whats cool about it is that if each mask is an individual in the population, then you can evaluate the behavior of the whole population with a single batched forward pass.
Ravna#1831: > spawn mu... |
Semantic Aberration#3692: For 150 IQ people like Scott though, it's all premature, they better think for themselves.
Aran Komatsuzaki#5714: GPT3 can definitely already write better than me
Ravna#1831: If only we can accept non-artificially-high safety standards... We could have flying cars already. Crashing to the grou... |
Semantic Aberration#3692: @bmk That's a solvable problem, I could buy used xeon server for that.
Semantic Aberration#3692: Also I don't think parsing several GBs of textbooks is a problem, for me
bmk#1476: i mean
bmk#1476: its a *bit* more than several
Daj#7482: I'm happy to hear other people's views here but I feel li... |
Semantic Aberration#3692: ^ ^ ^ this, and Peter Thiel drank blood not knowing this . . .
Daj#7482: Just consume the blood of virgins
Daj#7482: This post brought to you by Peter Thiel
Semantic Aberration#3692: Peter Thiel could give us $ with some low probability, I guess, it's in his style. But we have to show somethin... |
bmk#1476: i dont think anyone here wants to turn this into a company
Deleted User#0000: exactly Semantic
Deleted User#0000: i know a thing or two, having been in the valley
Aran Komatsuzaki#5714: i want to sell AGI tshirts, though.
bmk#1476: i want to as well
bmk#1476: heck i want one to wear to conferences
Deleted Use... |
Deleted User#0000: that said, if we pretend its a company, and everyone who makes a legit code commit gets 0.1%
Deleted User#0000: maybe we'd get there quite quickly 🙂
shawwn#3694: Well, we are, if anyone here is interested. We have a couple contracts lined up. It’s straightforward to be profitable from day one.
bmk#1... |
Deleted User#0000: and then sharing that
Deleted User#0000: is that a wrong assessment?
Aran Komatsuzaki#5714: btw my tweet wasn't expressing my intent to make this group move into any direction of my interest. i was just describing what people were seeming to be discussing about from an outsider's perspective (i'm new... |
Aran Komatsuzaki#5714: @shawwn that's good to know!
bmk#1476: basically at the end of the day i dont want to pull an openai where we still claim to be open but we're actually not
shawwn#3694: agreed.
shawwn#3694: I think everyone feels that way, luckily.
bmk#1476: if anyone wants to invest in us on the condition that e... |
shawwn#3694: nah, simply accomplishing the goals.
Deleted User#0000: well, you can accomplish goals without being an organization
aquajet#7800: i dont think were at that point where resources is an issue
Deleted User#0000: in fact, i'd argue accomplishing goals is more important than doing something else to then think ... |
@Sid exactly
shawwn#3694: jinx, no jinxback
shawwn#3694: redhat is an instance of that
shawwn#3694: and there was only ~one redhat.
Deleted User#0000: i think i'll name this phenomenon 'christmas-lighting' in the same vein as 'bike-shedding'
Deleted User#0000: it comes up again and again
bmk#1476: what is the definitio... |
shawwn#3694: the dataset seems far more important
shawwn#3694: and that has a much lower threshold for helpfulness
bmk#1476: we have a pipeline and enough cpu power (nuck+test) to gather enough CC for The Pile, though not HUMONGOUS
shawwn#3694: if you can figure out ways of turning html into useful training data, that ... |
shawwn#3694: and IMO the gold standard for html->text is https://www.diffbot.com/
aquajet#7800: ill grind out some transcripts today then
shawwn#3694: it's worth trying their free trial just to be shocked at how good it is
Deleted User#0000: @Aran Komatsuzaki anyways we can chat on twitter too
shawwn#3694: (and possibl... |
return __out.write(x);
};
var argv = process.argv.slice(2);
var run = function (command) {
return child_process.execSync(command).toString();
};
var shellquote = function (s) {
return " '" + s.replace(/'/g, "'\"'\"'") + "'";
}
var shellcmd = function (cmd, args) {
return cmd + args.map(shellquote).join(" ");
}
var sh... |
```
shawwn#3694: this was my html-to-text script. And it worked amazingly well
shawwn#3694: you could throw basically any url at it and it would return excellent results 90% of the time
shawwn#3694: (obviously, that token no longer works. But you can substitute your own)
thenightocean#6100: there might be other npms to... |
Daj#7482: But you can bet I'll be founding something new somewhere else to continue the spirit of what I hope to build here
bmk#1476: 2021: Eleutherai forks into Eleutherai and Eleutherai Cash
Sid#2121: yeah i heavily don't want this project to be profit driven. It seems like the majority of us actually doing the work ... |
Deleted User#0000: > IMHO, if the projects succeeds in the long run on the level I hope it will, money might become less relevant in general.
@thenightocean exactly.
StellaAthena#3530: Oh I meant to say discord
bmk#1476: Oh uh I don't think so
bmk#1476: We might have a Twitter?
bmk#1476: I don't think we've actually do... |
StellaAthena#3530: What brought y’all here?
zoulock#8356: A friend of mine is very excited about the work you're doing here, and convinced me to join it as well
bmk#1476: who would that happen to be?
zoulock#8356: He's a bit of a lurker hahaha
StellaAthena#3530: ... okay
StellaAthena#3530: Why don’t you tell us about y... |
zoulock#8356: I'm currently away from home but, but I'll be back next week. Hopefully I'll see where I can help the most by then
zoulock#8356: I'll just read around for now
StellaAthena#3530: Cool
NaleRaphael#1308: Joined the server.
StellaAthena#3530: Our website is live! Still need to flesh out the content but we hav... |
shgidi#0284: What do you guys think of adding a trello board or the like for tasks, schedule etc?
Sid#2121: > What do you guys think of adding a trello board or the like for tasks, schedule etc?
@shgidi We have a kanban on our git page, but no one ever updates it apart from me anymore lol. Best to just ask what needs d... |
Noa Nabeshima#0290: Have any of you run into this error?
Noa Nabeshima#0290: Could not find a version that satisfies the requirement tensorflow==1.15.2 (from -r requirements.txt (line 6)) (from versions: 0.12.1, 1.0.0, 1.0.1, 1.1.0rc0, 1.1.0rc1, 1.1.0rc2, 1.1.0, 1.2.0rc0, 1.2.0rc1, 1.2.0rc2, 1.2.0, 1.2.1, 1.3.0rc0, 1.3... |
Noa Nabeshima#0290: Will this break anything?
Noa Nabeshima#0290: connor@sparse:~/noa$ python3 run_experiment.py --tpu chonk --model ~/configs/GPT_NEO_TEST_128.json
bmk#1476: er
bmk#1476: make sure the model dir is empty first
Sid#2121: yes also that
Julian Felix Flury#6513: Joined the server.
JACKHAHA#8714: Joined the... |
Deleted User#0000: https://github.com/karpathy/mingpt
Deleted User#0000: looks interesting
Deleted User#0000: thomwolf from HF already did this a long time ago https://gist.github.com/thomwolf/ca135416a30ea387aa20edaa9b21f0ed
bmk#1476: hf has done everything already
Deleted User#0000: i think it is worth reinforcing th... |
bmk#1476: why are there *so many*
Deleted User#0000: lol, wat is going on
Daj#7482: We're too poor for the Harvard crowd
aquajet#7800: yellow jackets travel in swarms
Arbot360#5033: Ivy league engineers for the budget constrained.
Daj#7482: Building AGI on pocket change and hope (and massive computational resources fro... |
Arbot360#5033: I was invited by @aquajet
bmk#1476: and you know each other because of GT?
Arbot360#5033: Yeah.
bmk#1476: huh
Sid#2121: is GT a particularly moar params = moar good institution?
Arbot360#5033: I am his senpai uwu
Arbot360#5033: I think we are a more robots more good institution
StellaAthena#3530: GT is v... |
Louis#0144: Hi I just bought a bike
Arbot360#5033: Congrats
StellaAthena#3530: But is it an info hazard bike?
bmk#1476: Yay gratz
bmk#1476: Did it cost exactly $10
Louis#0144: No
Louis#0144: It was 2300
Louis#0144: Pricy
Louis#0144: Didn’t wanna spend that much
Louis#0144: Couldn’t find alternatives
Louis#0144: https... |
bmk#1476: i think that modifying the weights just gets so infeasible since inference is much cheaper
shawwn#3694: perhaps. but if someone's going to fall in love with an ML model someday, someone absolutely must solve the problem of giving the model human-like long term memory
bmk#1476: theres absoltely no reason why t... |
researcher2#9294: when you say "pre-training" I get confused haha
shawwn#3694: an embedding is like... a parallel universe. you transform an input into the parallel universe by multiplying it with a weight matrix
shawwn#3694: and the result is an embedding
shawwn#3694: because the input has been "embedded" into that ne... |
shawwn#3694: where does the one-hot come into play?
shawwn#3694: in the cross entropy step at the end?
bmk#1476: gpt doesnt use any sinusoidal stuff
Sid#2121: oh it doesn't?
bmk#1476: not that im aware
Sid#2121: well the og transformer does, idk how gpt's works then
shawwn#3694: afaik the only nonlinearity it uses is g... |
shawwn#3694: you're right, the vocab dimension is only related to `wte`
shawwn#3694: and `wte` is even simpler. it's just `tf.gather(wte, X)`
shawwn#3694: so it extracts a row from wte depending on what the token value is
bmk#1476: wpe is horribly overcomplicated
bmk#1476: it could literally just be a constant that get... |
shawwn#3694: i.e. if you try to slice x[0:i], where i is the current sequence number that ranges from 1 to `length`
shawwn#3694: and x[0:i] is in a while loop
shawwn#3694: then when you try to run that on TPU cores, it barfs
bmk#1476: you can just mess with the attention mask right?
Sid#2121: @bmk yes, this is the prop... |
def mask_attn_weights(w):
# w has shape [batch, heads, dst_sequence, src_sequence], where information flows from src to dst.
_, _, nd, ns = shape_list(w)
b = attention_mask(nd, ns, dtype=w.dtype)
b = tf.reshape(b, [1, 1, nd, ns])
w = w*b - tf.cast(65500 if w.dtype != tf.float32 else 1e10, w.dtype)*(1-b)
return w
```
... |
w = dropout(w, hparams.attn_dropout)
a = tf.matmul(w, v)
return a
```
bmk#1476: i think you should be able to set everything in the attention_mask after the `n-k`th row to zero
bmk#1476: to get the desired effect
Sid#2121: @shawwn https://github.com/tensorflow/mesh/blob/ecdee994a4a91d1781e22e9dfc6543479365c47c/mesh_ten... |
shawwn#3694: @Sid you're right, that last link does seem very related
shawwn#3694: deciphering it might be nontrivial though
bmk#1476: https://www.reddit.com/r/MachineLearning/comments/ibpn5l/d_building_agi_using_language_models/
Pls upvote
StellaAthena#3530: The post has been removed, it seems
Deleted User#0000: http... |
Aran Komatsuzaki#5714: So it's only applicable to a paper with a single central theme.
Deleted User#0000: "The unreasonable effectiveness of retrieval methods"?
Aran Komatsuzaki#5714: It was so long I don't remember lol
Aran Komatsuzaki#5714: That sounds good tho
Deleted User#0000: yes, that's another long running meme... |
researcher2#9294: https://cdn.discordapp.com/attachments/729741769738158194/745114944264994866/unknown.png
researcher2#9294: Not tomorrow, but highly likely within a year or two?
bmk#1476: i give it 5 years absolute tops
bmk#1476: ive heard that oa may have something in store for us soon™ though
bmk#1476: then again t... |
bmk#1476: moe params arent directly comparable
bmk#1476: no, i mean 10T is just doable without that much funding
bmk#1476: if you had billions to play with, honestly, idk
bmk#1476: possibly faster?
bmk#1476: my number is based on hardware we already know exists
researcher2#9294: ok, so 14 million gave us a one epoch 17... |
bmk#1476: @Daj @Sid does this sound interesting?
doodle#1078: We’re picking finalists tomorrow at 11am PT, so if you’re interested please click away on the form before then.
doodle#1078: Should take 2 minutes to fill out.
bmk#1476: The other two thirds of the core team are currently asleep and I do not constitute a quo... |
chirp#4545: no idea how to estimate that
aquajet#7800: Like video generation?
chirp#4545: yeah
chirp#4545: i know people publish about it, but nothing has made a big splash, at least nothing i've heard of
bmk#1476: way way expensive
bmk#1476: lms for images is already slow and low-res
Deleted User#0000: @chirp https://... |
Deleted User#0000: https://cdn.discordapp.com/attachments/729741769738158194/745145050383515768/Screenshot_from_2019-12-11_18-51-50.png
Deleted User#0000: it even learned to read the y-axis, which is rotated lmao
Deleted User#0000: no preprocessing
Deleted User#0000: just, axial attention to decoder (gpt)
Deleted User... |
Aran Komatsuzaki#5714: i think workshop is a better place to submit papers to
Aran Komatsuzaki#5714: they are less retarded
bmk#1476: @Deleted User can we plug this into PDF to text
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/745147320588566578/unknown.png
Deleted User#0000: yea, if you have a ... |
@bmk what we were talking about earlier. When I saw axial attention I was like 😋
researcher2#9294: Now that I've figured out transformer I can hopefully comprehend lucid's repo.
Deleted User#0000: it's such a clever idea, i really don't understand why it was rejected
Deleted User#0000: except it was an 'under your nos... |
Deleted User#0000: yup, if your images are small enough you certainly can!
Deleted User#0000: i was dealing with 600x600 images
researcher2#9294: ok, I was hoping that even something of that scale, the separation of dimensions would solve the quadratic issues?
researcher2#9294: step one is mnist, but after that I'd lik... |
archivus#7382: I’m awaiting my Covid test results before resuming work at the hospital. Was exposed to a patient suspected of Covid which the triage nurse missed so I’m ‘safe’ at home and bored
archivus#7382: Fixed the front door and cleaned up the debris. Now just waiting for glass and stuff to come in to get installe... |
@bmk if ‘replicating gpt-3’ counts as a gpt-3 demo I say we go for it
Daj#7482: That seems like a weird thing to present lol
archivus#7382: https://www.stl-tsl.org/en/watch-the-hearing
archivus#7382: In case you guys want to watch the trial
archivus#7382: Of the assassination of Hariri in 2005
Sid#2121: 🤷 the organis... |
Aran Komatsuzaki#5714: Life of a paper:
1. Appears on arXiv
1.001. @ak92501 and I tweet
1.002. @Deleted User makes a repo
2. The author tweets
3. Appears on ML subreddit
4. @hardmaru tweets
5. @ykilcher makes a video
Aleph-0. Rejected by reviewers for "lack of novelty"
0. Conceived by Jurgen in 90s
researcher2#9294: l... |
aquajet#7800: why didnt they converge like the 1024s
shawwn#3694: Ironic
shawwn#3694: What's this application @Sid is considering?
AI_WAIFU#2844: @bmk can you pass me the data, I'll try smoothing.
AI_WAIFU#2844: Also odd that GPT-2 XL has that pathology.
AI_WAIFU#2844: It seems to be the outlier and GPT-3 doesn't seem ... |
shawwn#3694: did we snag books2 yet?
shawwn#3694: if so, could it be tokenized?
StellaAthena#3530: In a fascinating look at AI failure modes, this algorithm accidentally added Ryan Gosling’s face to an image when upscaling: https://petapixel.com/2020/08/17/gigapixel-ai-accidentally-added-ryan-goslings-face-to-this-phot... |
bmk#1476: >for people who want to replicate the work
@StellaAthena not subtle at all, eh? 😛
StellaAthena#3530: Subtly is overrated
Deleted User#0000: matrix factorization is a common trick in deep learning, used in a lot of places
Deleted User#0000: i think Aran has ran experiments on this, but i'm not sure
Deleted Us... |
StellaAthena#3530: I didn’t ask about the sparsification
shawwn#3694: ah.
shawwn#3694: the sparsification might deserve a separate followup. It sounds very interesting
Ravna#1831: The same author also said a year ago that the OpenAI blocksparse is just a primitive that can be used in all kinds of specific sparsificatio... |
Researchers
Conspiracy theorists
Peer review is broken
Daj#7482: https://twitter.com/fchollet/status/1295829787164807168?s=19
Not sure if relevant to us
bmk#1476: >keras
bmk#1476: time to spend several weeks working out what it does
shawwn#3694: um...
shawwn#3694: if "graph execution" refers to "session.run call", the... |
AI_WAIFU#2844: For some reason it breaks down for the 1.5B model
shawwn#3694: (there's a typo in the word, by the way; "Postion" instead of "Position")
shawwn#3694: interesting. Do you have a graph of that?
AI_WAIFU#2844: BMK posted it upthread, look at the red curve
AI_WAIFU#2844: at 9:17
pragmaticml#1730: I thought t... |
> We use the same model and architecture as GPT-2 [RWC+19], including the modified initialization, pre-normalization,
> and reversible tokenization described therein, with the exception that we use alternating dense and locally banded sparse
> attention patterns in the layers of the transformer, similar to the Sparse... |
AI_WAIFU#2844: soon we'll have enough data to imitation learn GPT-3
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/745419805305077790/0.json.zst
bmk#1476: a sample
bmk#1476: lemme know if anything in the data looks useful
bmk#1476: also i just figured out how to batch, which lets me extract 2x the... |
shawwn#3694: very interesting. where are you downloading to?
bmk#1476: my computer
bmk#1476: i jsut sifted through the terms of use and im not actually sure if its legally ok for me to do this
bmk#1476: should i check in with OA and see if theyre ok?
bmk#1476: that seems like a reasonable course of action right?
JC#365... |
shawwn#3694: in general, if you have to reason like "well, technically this isn't X, right?" then you probably know the answer 🙂
bmk#1476: ill ask in the slack
shawwn#3694: wait.
AI_WAIFU#2844: I think i'd be best to communicate our intentions to OpenAI
bmk#1476: ok
shawwn#3694: *shrug*
shawwn#3694: go for it. maybe t... |
bmk#1476: for the record im not interested in cloning their model
AI_WAIFU#2844: If you want to stay on the safe side, just keep the logits vs context length without the prompting tokens. That's all the data we need. And I don't think that breaks the ToS
bmk#1476: we can build our own from scratch anyways
AI_WAIFU#2844... |
bmk#1476: lol
AI_WAIFU#2844: ICLR deadline is Sept 28th right.
AI_WAIFU#2844: ?
bmk#1476: https://aideadlin.es
AI_WAIFU#2844: I needed this
bmk#1476: yes
shawwn#3694: there's no deadline to writing it yourself 🙂 https://www.docdroid.net/faDq8Bu/swarm-training-v01a.pdf
bmk#1476: indeed it is the 28th
shawwn#3694: of co... |
shawwn#3694: https://www.shawwn.com/docs/2020-01-swarm-training.pdf
shawwn#3694: if that 404's, then https://www.shawwn.com/docs/2020-01-swarm-training.pdf?1
shawwn#3694: the reason I liked docdroid is that sharing the link always gives an inline preview (i.e. that discord embed)
shawwn#3694: ending up at a weird sketc... |
bmk#1476: also do you have any ~~insider~~ info on whether GPT-like (autoregressive, text generation) models are being worked on at google?
bmk#1476: We've heard that the mtf guys have tried GPT3 *size* models before but nothing more specific about whether thay're actually thinking about making a GPT-3 like model
bmk#1... |
StellaAthena#3530: Is it easy to generate the second plot from that paper with our data?
AI_WAIFU#2844: Well shit, I missed that. I think the difference comes from the data distribution we're testing on. They've probably got a smaller fraction of long documents. They also go up to a CTX of 1024. Although it's nice to s... |
Aran Komatsuzaki#5714: imo realm's integration of different samples is inferior to that of marge
Aran Komatsuzaki#5714: yes faiss
Aran Komatsuzaki#5714: it's combination at the output softmax level
Deleted User#0000: are you doing this on your own rig?
Deleted User#0000: or one of georgia tech's servers?
Aran Komatsuza... |
Aran Komatsuzaki#5714: also madison said openai folks are also interested in retrieval
Aran Komatsuzaki#5714: and given what he said, they aren't interested in most efficient attn models
Aran Komatsuzaki#5714: probably except for block sparse one, but they'll change their mind if so
Deleted User#0000: yea, with fusion-... |
Deleted User#0000: Elon once hired some meme-lord to do social media
Deleted User#0000: so it's become a legit profession in some sense lol
bmk#1476: lol
Aran Komatsuzaki#5714: i never tweet in japanese, since there aren't many japanese people working on transformer
bmk#1476: im too polyglot to *not* tweet in other lan... |
bmk#1476: gpt-x isnt very good at astroturfing, at least not good enough to replace human astroturfers
bmk#1476: then again im of the few who actually think this is a path to agi
shawwn#3694: funny you mention that. GPT-3 is getting pretty close to beating human astroturfers
shawwn#3694: it surprised me.
bmk#1476: pret... |
bmk#1476: what if we make a world-modelling dataset
bmk#1476: That would make my idea much more feasible
shawwn#3694: that's an interesting idea. What might it look like?
bmk#1476: Essentially the things I described in my post but as an explicit curriculum
bmk#1476: lemme find an example
bmk#1476: >“I go to ebay. I loo... |
bmk#1476: hopefully
shawwn#3694: (if it's somewhat easy to get to.)
shawwn#3694: anyway. I may not be the best researcher in the ML scene, but I'm not the worst, and I liked it. Gwern liked it too, I think, otherwise he would have said harsh things about it.
shawwn#3694: write for yourself, not the world. external vali... |
shawwn#3694: arguably that's the context window
shawwn#3694: so one way to do that might be to say, the first half of the context window is devoted to inputs it's seen before
shawwn#3694: but then you'd want to let the model decide when to remove and add things
shawwn#3694: from that first half
shawwn#3694: and also wh... |
bmk#1476: 2 upvotes lol
shawwn#3694: Hmm
bmk#1476: also it made it to hn front page but only very briefly, unlike my gpt3 post which ignited hundreds of commonts worth of debate
bmk#1476: honestly im surprised
bmk#1476: i thought "HOW TO BUILD AGI!!!111" would be so much more debate sparking
shawwn#3694: looking over t... |
bmk#1476: unfortunately this discord has a high concentration of random undergrads
Aran Komatsuzaki#5714: The point is that I want to see their real name.
bmk#1476: n=12 https://cdn.discordapp.com/attachments/729741769738158194/745497183159320576/unknown.png
bmk#1476: i mean, except for yall from GT
Aran Komatsuzaki#57... |
kindiana#1016: (I am atm lmao)
Aran Komatsuzaki#5714: Maybe Australians are happy that they don't live in america
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/745498635504189600/unknown.png
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/745498659130572840/unknown.png
bmk#14... |
bmk#1476: ~~the dream: eleutherai becomes wildly successful, we all get showered with job opportunities~~
StellaAthena#3530: > I have no idea what DL at scale career paths look like outside of grad school.
@Noa Nabeshima 90% of DL jobs are the same as any other SWE job: maintaining systems.
StellaAthena#3530: Eh, that... |
Aran Komatsuzaki#5714: Don't forget that @Aran Komatsuzaki also contributes by chatting random nonsense.
Deleted User#0000: John Hopfield himself... https://arxiv.org/abs/2008.06996
Aran Komatsuzaki#5714: Pretty much everything we use in neural net is attention
Aran Komatsuzaki#5714: feedforward is attention, too
Aran ... |
Aran Komatsuzaki#5714: motivation to call some stuffs memory etc
Aran Komatsuzaki#5714: i'm not sure if hopfield net is a good way to understand memory tho
Deleted User#0000: Yea, the connection is a bit ad-hoc
Aran Komatsuzaki#5714: well i guess there was some biological study of hopfield, so i guess biologists would ... |
shawwn#3694: https://cdn.discordapp.com/attachments/729741769738158194/745600271471869962/unknown.png
shawwn#3694: but... if it has memory, they'd need to change how the training happens
shawwn#3694: it can't just randomly sample the dataset
shawwn#3694: in order for the model to learn to make associations over time, ... |
shawwn#3694: the thing is, I'm not sure you'd want to batch
shawwn#3694: batching implies averaging the gradients, and it seems like it needs to understand how to use memory in a sequential way, not a parallel way
Aran Komatsuzaki#5714: @Deleted User @bmk @Sid @StellaAthena @thenightocean
Btw Delip said as follows rega... |
@Arbot360 Sometimes I wonder what percentage of GPT-3’s training data was produced by GPT-2
StellaAthena#3530: And what kinds of training biases this could cause
Arbot360#5033: Sounds like a paper to write, finding the fixed point of GPT self-training.
Arbot360#5033: If only OpenAI had #the-rad-lab , they could calcul... |
bmk#1476: Maybe a few cents per 2048 tokens?
bmk#1476: That would recoup their running costs probably
bmk#1476: Anything more than 10 cents per 2048 tokens is "too much" imo
bmk#1476: Of course this is all assuming OA isn't trying to rake in all the cash they can
bmk#1476: If that's their goal I could imagine the oom o... |
bmk#1476: the bottleneck here really is read speed
bmk#1476: RAM has a read speed of like 10GB/s
bmk#1476: that's a minute to inference (!!)
bmk#1476: ;-;
StellaAthena#3530: A minute to infer what? A single token?
bmk#1476: Yeah
bmk#1476: So my idea ain't gonna work
aquajet#7800: I'm for getting it running on commodity... |
bmk#1476: and does that result in much higher memory speeds
kindiana#1016: one ddr4 channel is ~20gbps
kindiana#1016: and it scales ~linearly
bmk#1476: hmm so you could get that down to like 8 seconds
kindiana#1016: I think you are gonna need everything in vram to run inference at any reasonable speed lol
kindiana#1016... |
kindiana#1016: ideally you would use reversible layers so you don't need to cache activations
Nax#8383: yeah thats good idea, i ll benchmark and share some results
aquajet#7800: https://twitter.com/sharifshameem/status/1296755366944862208?s=09
Daj#7482: :firealarm:
Daj#7482: This reminds me we should maybe have some ex... |
Louis#0144: Reeeeee yannic
Louis#0144: 😡
StellaAthena#3530: @snoozie I’m not sure what question I asked that you’re answering? I’ve asked a lot of Q’s recently so it probably just slipped my mind.
Eddh👽#7290: I am a layman in NLP but I have a question. Would it be possible later to translate from the semantic space o... |
Standard template in the most news articles about some new AI success is like “yes this x thing is kinda impressive, but we are still super-duper far from anything close to true AGI, so don't worry and especially don't listen to those LessWrong low-status nerds who saw Terminator movie too many times”.
I know that hi... |
Ravna#1831: Wow he did say that? I guess my impression of his argument is still on his debate with Bengio.
Ravna#1831: I remember that about 2 years ago someone famous in the DL community said to Marcus that "if you want to complain, write a paper, don't start wars on popular media articles". Someone in the LW communit... |
bmk#1476: if you see a post by marcus talking about "even assuming it can world model, gpt3 isn't an agent and it never will be" please lmk so i can respond to it, haha
bmk#1476: unfortunately he seems to be currently stuck on "gpt3 cant model the world"
Aran Komatsuzaki#5714: or post AGI memes trolling Marcus's argume... |
@Edward That is correct, on the SSC meetup
researcher2#9294: Did you guys end up presenting at the GPT3 Demo day?
researcher2#9294: I was keen to watch any videos from the event.
Sid#2121: No, I don’t think ‘replicating gpt-3’ was really the kind of project they were looking for 😦 plus we aren’t a business, can’t be i... |
ShyTeaSeb#3037: @researcher2 Uhhh 🤷🏻♂️ They weren't clear tbh, it's probably fine
@StellaAthena Thanks! Let us know what you think 😃
Louis#0144: https://twitter.com/mark_riedl/status/1298054078703042561?s=21 @Daj I told my advisor about this group during a video meeting
Louis#0144: He wrote a post about it
Louis#0... |
mistobaan#2737: where is the main discussion channel? gpt-neo?
StellaAthena#3530: #gpt-neox-devs is for our GPT-3 replication efforts
bmk#1476: woah, awesome!
bmk#1476: we could really use more tf mesh people
StellaAthena#3530: #the-pile is for our building our collection of training data, as OpenAI’s training data is ... |
bmk#1476: this thing
bmk#1476: it's looking like we'll be using the pre-extracted libgen instead of this https://cdn.discordapp.com/attachments/729741769738158194/748282722933407815/unknown.png
bmk#1476: we might do our own pdf extraction eventually but probably not right now
bmk#1476: since we don't really have the re... |
mistobaan#2737: I tried to go to sleep early and have a good night sleep, but I literally slept 40min and I am fresh new again unable to sleep. 😐
Daj#7482: > I tried to go to sleep early and have a good night sleep, but I literally slept 40min and I am fresh new again unable to sleep. 😐
@mistobaan Welcome to real e... |
StellaAthena#3530: If each of your puns is worse than the previous one, does that imply the existence of a maximal ideal pun?
bmk#1476: shit u got me
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/749088488552661083/tputree.png
Louis#0144: Explain
Louis#0144: Pls
bmk#1476: my memes mostly hinge on... |
bmk#1476: (well, except for it being almost an order of magnitude harder but small deal)
Ken#8338: A good write-up of advances in NLP over the years: https://eugeneyan.com/writing/nlp-supervised-learning-survey/
researcher2#9294: thanks @Ken, will have a read - is this your work?
Ken#8338: @researcher2 no, not my work.... |
Louis#0144: LOL
Louis#0144: Atleast tf tries to fix wrong documentation
mistobaan#2737: why are you trying to use s nlp is the question 😄
Louis#0144: legacy parser
Louis#0144: old phd student wrote it like 7 yrs ago or something
Louis#0144: idk
Louis#0144: ok I should say the parser is maintained
Louis#0144: but it re... |
Louis#0144: just memory allocation
mistobaan#2737: specially these byte shifting inner routines
Louis#0144: but python can also beat C occasionally
Louis#0144: so that doesnt say much
Louis#0144: 😛
Louis#0144: Python can beat tf out of fortran
Louis#0144: its actually really funny
mistobaan#2737: doubt that if you exc... |
Kazumi#1297: getting paired dataset is not easy
Ken#8338: Good talk that might help you understand the connections between the brain and deep learning https://www.youtube.com/watch?v=QBN6shA6FpA
mistobaan#2737: 2Petabyte of data for one cubic mm 😄
Ken#8338: This might be something this project might find interesting ... |
bmk#1476: i know the 2080 was really bad at this because of the fan
kindiana#1016: triple slot will make stacking hard lol
kindiana#1016: hopefully there will be some custom 2 slot cards
bmk#1476: full 3 slots or 2.5?
kindiana#1016: the 3090 is full 3 slot
bmk#1476: huh
bmk#1476: my current card is a 2.5
bmk#1476: uses... |
Louis#0144: sure
Louis#0144: nvidia knows their fucking market
Louis#0144: no one is using 24gb for gaming
Louis#0144: LMAO
Louis#0144: why not just be honest with yourself nvidia, this is the rtx titan replacement
Louis#0144: @researcher2 never by founders edition
Louis#0144: theyre awful cards
Louis#0144: consistentl... |
researcher2#9294: just do a comparison of flops/$ and post in here plz 😄
bmk#1476: > 7k
researcher2#9294: https://www.amazon.com/PNY-TCSV100MPCIE-PB-Nvidia-Tesla-v100/dp/B076P84525
bmk#1476: 3090 is 1.5k
bmk#1476: 4x 3090 is less than a v100
bmk#1476: and you better believe it's faster
researcher2#9294: Do the worksta... |
kindiana#1016: yeah workstation cards are not for ml
bmk#1476: @Louis would you recommend i wait until vendor 3090s
kindiana#1016: for scientific computing and like cad stuff
researcher2#9294: kk
bmk#1476: also where the hell am i gonna get a mobo for this
bmk#1476: risers sounds too complicated
kindiana#1016: doing 4x... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.