data stringlengths 115 7.61k |
|---|
StellaAthena#3530: I’ll check it out
bmk#1476: https://arxiv.org/pdf/1902.09469.pdf here's an arxiv-paper version of it, which i presume is more polished than the AF version
bmk#1476: on my last read-through, i *really* got stuck around the counterfactuals and Löb's theorem bits
bmk#1476: https://cdn.discordapp.com/at... |
StellaAthena#3530: The example with the $5 and the $10 does not actually illustrate their point
bmk#1476: hmm
bmk#1476: i mean isnt their point that naively thinking about counterfactuals can cause weird behavior?
bmk#1476: and if the weird math stuff is correct it should justify that
StellaAthena#3530: Right, but noth... |
3dprint_the_world#6486: seems like that's what that's implying. Unless I misunderstand.
bmk#1476: so "I want to want to enjoy math" + "I am a sufficiently powerful agent to change my own desires" -> "I begin wanting to enjoy math"
bmk#1476: right?
StellaAthena#3530: Right
StellaAthena#3530: Löb’s theorem is a similar p... |
StellaAthena#3530: Read section 2.4
bmk#1476: hm, that makes sense
bmk#1476: so even if you only do it once, you still want to go through with the plan that gives you the higher expected return
bmk#1476: even if you already know which "branch" you're on
bmk#1476: this feels a lot like game theory
bmk#1476: i dont see h... |
bmk#1476: not really
StellaAthena#3530: An AGI puts two boxes before you. In one, you see $100. The other is opaque. The AI tells you “you may take any box(es) you wish. I have predicted how you will act. Had I predicted you will take both boxes, the opaque box will be left empty. Has I predicted you will take one box,... |
$1,000,000 there, $1,000,000 not there
StellaAthena#3530: > Eliezer and those who agree with him can prove that a rational agent will take two boxes.
My response has always been that I hope that that proof makes them happier than my $1,000,000 makes me.
bmk#1476: ok so
bmk#1476: the AI precommits right?
StellaAthena#35... |
bmk#1476: right?
bmk#1476: even in the "ai predicts take one" universe i'll still want to pick only opaque
StellaAthena#3530: I disagree with the line argued in this paper, so I’m not a good person to try to convince you
bmk#1476: i mean, convince me of your argument
bmk#1476: more arguments better
StellaAthena#3530: M... |
asparagui#6391: literally you can conceive an infinite universe, but not a denizen who will disagree with you for the sake of doing so
bmk#1476: I dislike this concept of super unrigorousness
bmk#1476: I can conceive of a computer program that is completely useless and wrong, and yet I wouldn't want to make one despite... |
Daj#7482: Haha I think people quietly dropped functional
Daj#7482: But I'm not an expert
Daj#7482: I still haven't read the functional DT paper
asparagui#6391: you all like math
asparagui#6391: so let me throw in mr gödel
asparagui#6391: my decision is the decision not in the set of decisions in decision theory
bmk#147... |
bmk#1476: i dont think i can handle this level of math this late
bmk#1476: Idk the halting problem seems kinda intuitive to me
Daj#7482: Scott Aaronson had a good paper on that on his website
bmk#1476: Meanwhile I couldn't prove gödel on my own
Daj#7482: Yea My point is that seeing incompleteness as a halting problem i... |
3dprint_the_world#6486: the problem is all these Haskell people who discovered they can make their language look like it's really cool if they used Monads to do printf()
3dprint_the_world#6486: but just because they like wearing hair shirts, doesn't mean everyone else does too.
3dprint_the_world#6486: (j/k, I like Hask... |
Daj#7482: Thank you :)
StellaAthena#3530: > this is news to me. Elkezers whole point of TDT is that you should take one box
@Daj I could have worded this better. The whole point of TDT is that you should take one box. However the *reason to develop it* is that E felt that under other frameworks you can prove that a rat... |
Daj#7482: Yea, Seen it, pretty cool stuff
Daj#7482: Not super relevant to us unfortunately since we don't have tons of GPUs
Youssef Abdelmohsen#8707: If only I was a gazillionaire! :/ I'd donate
Youssef Abdelmohsen#8707: I saw you guys were looking for OCR tools for math, you're also prob already familiar with this but... |
Youssef Abdelmohsen#8707: Are you guys planning to include Libgen sci papers as well?
Youssef Abdelmohsen#8707: (Or only books?)
Youssef Abdelmohsen#8707: I checked the Libgen sci paper archive size and that was around 76 TB last time I checked.
Youssef Abdelmohsen#8707: But that's only in PDF size / not text of course... |
bmk#1476: So I'm not a Professional Economy Person™, but my prior is that it probably wouldn't work
bmk#1476: There probably are ways to make it work, but they would be highly nontrivial
3dprint_the_world#6486: I am not a Professional Stock Maker either, but check out https://towardsdatascience.com/stock-predictions-wi... |
bmk#1476: or whatever
3dprint_the_world#6486: ok
bmk#1476: ~~we should totally spin off an algotrading business after we replicate gpt3~~
3dprint_the_world#6486: lol
bmk#1476: im not kidding, i actually think gpt3 could be very useful for algotrading, and moreover i feel confident in stating that most trading firms pro... |
59.70 GiB ± 968.85 MiB
59.62 GiB ± 967.86 MiB
59.76 GiB ± 946.74 MiB
59.70 GiB ± 833.13 MiB
```
bmk#1476: im way too proud of myself haha
bmk#1476: tfw you actually find a chance to use something from a stats textbook in the real world and not on some contrived problemset
3dprint_the_world#6486: how does it work
bmk#14... |
researcher2#9294: My first thought was to try to get curators on board for their field of specialty.
researcher2#9294: But that wouldn't enhance what you have, just add an entirely new feature.
researcher2#9294: What I really want is to be able to stick a cable in the back of my head to download it all, can you do this... |
Davidweiss2#3174: How can I make my computer gpu available for use?
Daj#7482: We could definitely use large amounts of CPUs (and network bandwidth) for data processing. We're currently getting our hardware for training from Google's TFRC program, so we don't need GPUs unless you have hundreds of them available
Daj#7482... |
@Daj I am no coder expert, buy my team is team of young hackers that will love to get their hands dirty.
Daj#7482: > wait, is the original GPT-3 model unreleased? this project makes a ton more sense now
@stellie Correct
Daj#7482: > @Daj I am no coder expert, buy my team is team of young hackers that will love to get th... |
stellie#3553: ah, i thought it was all common crawl
StellaAthena#3530: https://cdn.discordapp.com/attachments/729741769738158194/755808261210374144/unknown.png
StellaAthena#3530: CC is public, but they don't document how they process and filter it. WebText2 is public, Books1 and Books2 are not public and they're vague... |
Anyone here have intuitions about the probability of powerful insights from mathematics that change AI timelines in a surprising way?
StellaAthena#3530: @Noa Nabeshima most of that comment was about AI impacting mathematics, and then you switched to mathematics impacting AI. Was that deliberate?
StellaAthena#3530: We’v... |
bmk#1476: Define basic
bmk#1476: Ive been bitten by people who have.. very different definitions of basic than me
StellaAthena#3530: If you’ve been to the first two meetings of an intro course on group theory you’re fine
StellaAthena#3530: “What is a group” is necessary
StellaAthena#3530: Almost nothing about the theor... |
StellaAthena#3530: The convolution is a slight modification of this concept, where instead of shifting by g you shift *backwards* by g. The reason for this is abstract and unimportant, but it’s related to why you typically want to conjugate the transpose of a complex matrix.
StellaAthena#3530: I would actually represen... |
bmk#1476: > International Conference on Fun with Algorithms
bmk#1476: This sounds like an amazing conference
StellaAthena#3530: @bmk it is! You can see the list of accepted papers here: https://sites.google.com/view/fun2020/acceptedpapers
StellaAthena#3530: @Louis the preprint came out two years ago. Fun with Algorithm... |
Daj#7482: I'm technically an undergrad too lol
Abecid#5364: Yeah that's insane
Daj#7482: I had a few years before undergrad of dicking around haha
StellaAthena#3530: I’m currently “under” a “grad” but I don’t think that’s what you mean lol 😛
bmk#1476: https://twitter.com/numpy_team/status/1306268442450972674?s=19
bmk#... |
bmk#1476: I'd recommend you to maybe take a look at the established literature? A "robust model of intelligence" is very, very hard (one could argue that it's an unsolved problem), so the probability of any of us coming up with one is quite slim
Dal#7192: I can account for why GPT2 was underestimated and where there ar... |
bmk#1476: and?
Dal#7192: And trust it? Or don't
bmk#1476: theres no alignment or safety anywhere here
Dal#7192: I can map data to human compatible communication, beyond that is beyond me for now
bmk#1476: ok so you're thinking less about safety and more about getting to AGI in the first place then?
Dal#7192: Intelligen... |
Dal#7192: The nature of communication requires modeling what your partner knows and what information is salient to them
Dal#7192: Achieving the goal of communication means finding the most salient information, and as such being familiar with the partner's own world model to whatever extent
Dal#7192: In that way it's ut... |
Consequently, the control problem will always be inexact, defined by whatever parameters we're able to communicate. I'd posit that our best bet is set up a Reverse Genie, an oracle that takes our imperfect wish and reinterprets it as we would desire it to be understood.
```
Dal#7192: ```Specific poorly informed hypothe... |
Stephen5311#6349: Because he kinda became imo the best Minecraft PvPer in the game of all time
Stephen5311#6349: 2nd would be Tylarzz
Stephen5311#6349: Who is now a pro Fortnite player
Stephen5311#6349: But yeah, he went from no FPS experience
Stephen5311#6349: To pro Overwatch in a year
researcher2#9294: damn
Stephen5... |
Stephen5311#6349: I didn't know this till now
Stephen5311#6349: But I guess this was his first streaming playing it LOL
researcher2#9294: ugh, the precision
researcher2#9294: never quite had that level of refinement
Stephen5311#6349: Same lol
researcher2#9294: Any character that didn't require consistent headshotting I... |
Ken#8338: I am guessing the doubling time will slow downs, but who knows.
gwern#1782: yeah, but I don't think it's going to keep going for the next 6 years
gwern#1782: I think it may be not too far from breaking down. you see anyone training GPT-3 successors beside eleutherai?
gwern#1782: even MS's deepspeed stuff only... |
gwern#1782: you think Google put up $1m in prizes for lee sedol because it was vitally important to datacenter optimization research?
gwern#1782: so let's see... Hernandez is 'cost halves every 16 months / 1.3 years'. if you want to bring the current $10t ($5m * 2.2e6) down to just $1b (1e9), you have to wait...
bmk#1... |
bmk#1476: but 10 years from now.. that's a *long* time
bmk#1476: DL didnt exist as a field 10 years ago
Ken#8338: I think I remember you mentioning that you can imagine 100Trillion parameter models by 2030, ot was it 1 quadrillion?
bmk#1476: sounds like something i would say
bmk#1476: yeah wouldnt be surprised if that ... |
gwern#1782: yeah, once you start hitting 10 to 100t, I think you either need to start training on social media (how many words ya need? twitter's got'em!), or finally stop dragging your heels and go multimodal
bmk#1476: is twitter *that* big?
bmk#1476: X - doubt
Ken#8338: agree with the multi-modal - would think it wou... |
bmk#1476: well
bmk#1476: i hope haha
bmk#1476: we have nowhere near the hardware necessary
bmk#1476: no, 100T will be done by OA or google or M$ or fb or someone big
Ken#8338: All comes down to affording the compute
bmk#1476: and probably about 10 years from now
bmk#1476: maybe a bit earlier
Ken#8338: At least your are... |
bmk#1476: i mean, MoE parameters dont map cleanly so thats why i dont include them in any of my estimates
bmk#1476: i think a 100T MoE will be achieved much sooner than 10 years
gwern#1782: we should sacrifice a hectatomb of 1080tis at the shrine of saint schmidhuber that he may guide the way of MoEs into the light so ... |
genai (Immortal Discoveries)#0601: I'm not sure what you mean, so it can learn, predict, learn, predict, etc ?
bmk#1476: learning doesnt conflict with predicting
genai (Immortal Discoveries)#0601: cuz i'm trying to figure out why GPT hasn't been successfully ran on the Hutter Prize, even if it doesn't exactly follow th... |
genai (Immortal Discoveries)#0601: > gpt3 is 350GB
@bmk But above I asked: And 2) if we train GPT as it sees data, then we would know it isn't storing the definite answers, and would get the actual compression, even though the net would still be huge - we know it cannot be that much novel information in there at least.... |
bmk#1476: uh, im not sure i understand
genai (Immortal Discoveries)#0601: batches?
kindiana#1016: at t=0. the compressor trains the model on a part of the data, and sends that part of the data across without compression
kindiana#1016: the decompressor receives it and trains an identical model
kindiana#1016: then that m... |
kindiana#1016: theres a optimal network size for given data constraints, and gpt3 is wayyyy too big
genai (Immortal Discoveries)#0601: but its not gpt3, nor that big, only sees the enwik (100MB) bit by bit as it tests on it...
bmk#1476: off the top of my head, you're probably looking for about 10M parameters for 1GB of... |
Then it would become really easy to replicate your work... It could encourage others to build on your shoulders and multiply the impact your project has
Daj#7482: I'm in favor of good documentation for sure, and also plan on writing some extended blog posts on both the meta and technical aspects of the project. Though... |
zphang#7252: yea that quote mentions nothing that is "exclusive"
Daj#7482: wtf is wrong with OpenAI's PR
Daj#7482: Seriously
Daj#7482: Why can't they communicate _anything_ clearly?
bmk#1476: you gotta phone up jack and rant his ears off
Daj#7482: I'm considering it lol
Daj#7482: But I honestly can't think even of anyt... |
zphang#7252: so MSFT is the only reseller
bmk#1476: if that's the case, they picked possibly the worst way to word that
Daj#7482: What if I build an app that offers GPT3 functionality?
Daj#7482: Seems weird
Daj#7482: > if that's the case, they picked possibly the worst way to word that
@bmk this
bmk#1476: oa has been v... |
bmk#1476: doesnt change the prior much
bmk#1476: so yeah uh M$ bad
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/758017506219655219/unknown.png
bmk#1476: GPT-Clippy
Daj#7482: > GPT-Clippy
@bmk _literally a paperclip_
Louis#0144: NO
Louis#0144: NOT LIKE THIS
Daj#7482: If this was a black mirror ep... |
Daj#7482: Does feel that way sometimes ngl
Daj#7482: Especially lately
bmk#1476: so uh
bmk#1476: we need to really focus on alignment
bmk#1476: i've been studying measure theory in preparation for reading the infrabayesian sequence
bmk#1476: if i understand correctly, that's pretty close to where the cutting edge is an... |
Daj#7482: The canonical intro is Rationality: A-Z haha
Daj#7482: And then the three AF sequences I guess
bmk#1476: is it even about alignment
Daj#7482: And then a disheveled mess of hundreds of disconnected blog posts
Daj#7482: > is it even about alignment
@bmk yes
bmk#1476: rationality a-z is more just about rationali... |
Daj#7482: You wanted to know the canon
Daj#7482: You can get in through other ways
Daj#7482: IDA is pretty independent
bmk#1476: ok so yeah, let's do a reading club
bmk#1476: there is absolutely no way i'll get through rationality a-z on my own before the singularity without external pressure
Daj#7482: That might be fu... |
bmk#1476: I've always wanted to read thinking fast and slow
Daj#7482: Good book!
Daj#7482: If I could choose I'd recommend history books
Daj#7482: I've profited immensely from history books
Daj#7482: Currently reading Rise and Fall of the Third Reich
Daj#7482: What a tome
bmk#1476: Eh I personally am more interested in... |
Daj#7482: Page is not equal page imo
Daj#7482: Some books I need two weeks for 100 pages, others I can do in an afternoon
Daj#7482: (and infra I read at 1 sentence/week)
bmk#1476: Fair
bmk#1476: So if you think one book per week is doable then let's do it
Daj#7482: I think it's absolutely doable but you need to read ev... |
Daj#7482: This'll be fun! I've been meaning to reared the sequences anyways
bmk#1476: Awesome
Daj#7482: Think we can finish the first book this week? I think I could
bmk#1476: I've already read it once so it'll be fairly
easy
Daj#7482: Then how does maybe Sunday sound for book club meetup?
Daj#7482: We could also make ... |
Louis#0144: ya
Louis#0144: scroll up
Louis#0144: 😛
Quill#9732: ClosedAI
bmk#1476: what an original joke
Quill#9732: thank you
bmk#1476: in fact, it literally is pretty original, it's only ever been used one other time in this server https://cdn.discordapp.com/attachments/729741769738158194/758038904434589716/unknown.p... |
bmk#1476: and then the editorializing
bmk#1476: **AAAAAA**
bmk#1476: is this what gell mann amnesia feels like
Daj#7482: I have to hate that paper because it's from the LMU, my university's sworn enemy
Daj#7482: Haha
Daj#7482: The common joke is if the worst half of TUM's (my uni) students went to the LMU, both would i... |
bmk#1476: no, tfrc
bmk#1476: so yeah the major catch is that there's absolutely no guarantee how many tpus we can reserve
bmk#1476: TPU podcast had a major glut recently where they couldnt create a single tpu
bmk#1476: so having more cores to run experiments on is always a good thing
trsohmers#5863: I may be able to re... |
bmk#1476: at least
bmk#1476: probably more if you want it done in a reasonable timeframe
trsohmers#5863: 8x RTX 8000s would be 384GB of VRAM; OpenAI's paper lists 353GB of VRAM necessary for the full 175B parameter model
bmk#1476: you need to fit the activations too
Sid#2121: @trsohmers just reiterating we'd be very in... |
en3r0#3241: Not sure where to throw this, but the discord link in the EleutherAI Core Document is broken. https://docs.google.com/document/d/1yOnxEMlU57M8YFlQC3XNOvyMVX2EpU5LeIWhEBcwQNk/edit#heading=h.vjo85qqx8r3b
Daj#7482: Thanks! Pinging whoever's in charge ( @StellaAthena ? @bmk ?)
en3r0#3241: That was where I found... |
en3r0#3241: Haha very true!
en3r0#3241: If you are serious about funding though, I think you may find it. Especially with Microsoft announcement yesterday.
Daj#7482: It's not something we've been particularly keen on pursuing
Daj#7482: Most of us have day jobs, and it would taint the "purity" of this place if we were e... |
Sid#2121: definitely interested but unless they have a lot the training times are likely too high
Daj#7482: I'd be interested in GPUs for side projects other than Neo
Daj#7482: Neo is too big for GPUs unless we have like a few hundred at least
Daj#7482: But could be also useful for testing, or even dataprocessing since... |
Daj#7482: assuming no loss of efficiency etc
kindiana#1016: has there been any work with controlling generation by just adding vectors to the residual conenctions going through the network? (i.e. for sentiment from movie reviews, and take the mean of the hidden representation of all the positive samples at each layer, ... |
Daj#7482: I think this could be a Eleuther flagship project
Daj#7482: "The internet's communal AGI, train it to do anything!"
Daj#7482: I don't think it'll work ofc, but could also result in a cool dataset of paired commands and outputs if we allow the user to enter what they actually wanted
Daj#7482: Should focus on N... |
StellaAthena#3530: NeurIPS papers published from 2008 to 2014 average ~40 citations
StellaAthena#3530: (Ignoring recent years because of expected future citations)
StellaAthena#3530: https://www.microsoft.com/en-us/research/project/academic/articles/neurips-conference-analytics/
Daj#7482: Yea you're probably right
Daj#... |
StellaAthena#3530: Nice!
StellaAthena#3530: What degree are you working towards?
Daj#7482: Just a CS B.Sc.
Daj#7482: nothing exciting
StellaAthena#3530: http://www.artificialintelligencebits.com/self-driving-cars-will-hit-the-indianapolis-motor-speedway-in-a-landmark-a-i-race/
StellaAthena#3530: ^^ Don't let the result... |
StellaAthena#3530: (For those wondering, we do literally zero systematic outreach)
Daj#7482: You know I'm bullish on GPT3 and related tech hah
StellaAthena#3530: > i thought everyone would have forgotten about gpt3 by now
@bmk have you met ML people? 😛
bmk#1476: if the month on the arxiv id is more than one month in t... |
Daj#7482: Hahaha
bmk#1476: something something a difference in degree so big that it's a difference in kind
Daj#7482: GPT3 is software 3.0
StellaAthena#3530: > Prompt design as programming
@Daj this wasn’t invented by GPT-3. People were talking about it for years before GPT-3. It was made practical by GPT-3.
Daj#7482: ... |
PlumpTheBrave#4056: Plus open research means better feedback loops and convergence to impact (i hope lol)
PlumpTheBrave#4056: Both! Dr. Joyner also has been involved (to a very small degree lol)
PlumpTheBrave#4056: Let me add you, I can dm you about it
StellaAthena#3530: Also, if you (or anyone else) is interested in A... |
"You have to figure out how to convince the AI to give you the answer in a human-like manner and the AI is an intelligence that you can not understand" would have been deemed unrealistic SF 5 years ago
Daj#7482: @PlumpTheBrave That'd be awesome! Best to just be shameless and ask around rather than relying on our someti... |
bmk#1476: welcome to the dungeon of ai
bmk#1476: we only do weird stuff here
Adam Fisher#5266: I'm so down. Chain me to the floor! 🤪
kindiana#1016: https://ai.googleblog.com/2020/09/advancing-nlp-with-efficient-projection.html
genai (Immortal Discoveries)#0601: new paper https://arxiv.org/pdf/2009.07118.pdf
genai (Imm... |
StellaAthena#3530: > for instance, I know a research team at the Swedish Royal Library are at least in the phase of collecting a dataset for Swedish on the same order of magnitude :]
@bfredl do you know where they get their data from? Maybe I’m biased by speaking only English, but finding such data online is quite hard... |
bfredl#1945: GPT-3 already goes ways towards that, but the strong English bias is noticable..
bmk#1476: Our open gpt3 will be strongly english biased too but we plan on going even bigger and building a much more balanced model
bmk#1476: Also, you speak German?
bfredl#1945: I suspect the somewhat swedish overrepresentat... |
bmk#1476: then you're in good company here
StellaAthena#3530: > Even if I do not work with it directly, I am quite interested in the works of preserving endangered dialects / minority lang variants
@bfredl this is dope
bmk#1476: one major advantage of *multilingual* gpt3 for minority languages is that it could transfer... |
Alm#9130: yeah but if the wrong combinations are there
bfredl#1945: (I e GPT-3 swedish text becomes shorter, as the chars are more expensive to encode)
bmk#1476: We plan on redoing BPE for multilingual
bmk#1476: But sticking with OA BPE for monolingual
Alm#9130: well.i i get better score with larger vocab in swedish on... |
Louis#0144: The issue is masked photo of someone goes in, their address and credit card and name information comes out
Louis#0144: This isn’t the model right now
Louis#0144: But by implementing their model, they’re collecting data for this
Louis#0144: Think of the implications in HK or at BLM protests
Louis#0144: It’s... |
Louis#0144: We just don’t know
bmk#1476: I mean if the adversary is the govt
bmk#1476: And they want to do it
bmk#1476: Nobody's able to stop them anyways
Louis#0144: Mhm....
bmk#1476: If anything, at least if Uber does this everyone knows they're being watched
bmk#1476: (this has been an exercise in devil's advocacy)
... |
StellaAthena#3530: See, this is why I don't get libertarianism.
Louis#0144: man the amount of techbros defending uber
StellaAthena#3530: My response is "... duh?"
Louis#0144: is actually absurd
bmk#1476: > See, this is why I don't get libertarianism.
is this in response to "lean libertarian"?
bmk#1476: if so, i dont ac... |
Louis#0144: canada has it bad too...
Louis#0144: I doubt bmk likes trudeau at all
Louis#0144: his management of the virus has been pretty bad
Louis#0144: like even compared to the US
Louis#0144: its pretty bad...
Louis#0144: canadian populations arent dense tho
Louis#0144: so it isnt as obvious
StellaAthena#3530: https... |
bmk#1476: when i say lean, i'm somewhere between the first two points
bmk#1476: and i still really dont know
bmk#1476: Capital L Libertarianism is somewhere around the 3rd point
bmk#1476: anarchy is 4/5
bmk#1476: > I doubt bmk likes trudeau at all
@Louis the general opinion of trudeau that i get is that he's very good ... |
Louis#0144: The goose flies to Ottawa
Louis#0144: Giant tournament
Louis#0144: The last goose standing wins
Louis#0144: That person is then elected
bmk#1476: lmao
bmk#1476: so disclaimer that im a bit rusty on this
bmk#1476: (i know more about how the us system works than canada just because of the sheer volume of us c... |
bmk#1476: https://github.com/kuangliu/pytorch-cifar
StellaAthena#3530: hmmmmm
bmk#1476: which seems consistent with paperswithcode https://paperswithcode.com/sota/image-classification-on-cifar-10
bmk#1476: unrelated but: X - doubt https://cdn.discordapp.com/attachments/729741769738158194/758890491344322590/unknown.png
... |
kindiana#1016: i mean, if you don't do any data augmentation or regularization, its not surprising
bmk#1476: It still shouldn't be that bad, no?
StellaAthena#3530: I believe we are doing random crop and flipping
bmk#1476: What lr? That's responsible for about 50% of problems
kindiana#1016: the train loss is like 2e-5, ... |
bmk#1476: that's *way* too high
StellaAthena#3530: lol
bmk#1476: all of ML is voodoo like this, you just know things are off from hours of bashing your head into things
StellaAthena#3530: What's a good lr
StellaAthena#3530: 0.001?
bmk#1476: eh 1e-4 is a good starting point
bmk#1476: also my instincts are mostly calibr... |
researcher2#9294: > try cifar100
@ntakouris That could be a good next step, thanks.
ntakouris#5483: I would also suggest to try some colour augmentations and preprocess the dataset with per-channel mean before doing model input.
ntakouris#5483: Cifar 10 and 100 are so tiny datasets that you can even get top1 accuracy o... |
kindiana#1016: it gives better model accuracy at iso size
kindiana#1016: i.e. the most significant bits of weights are more important, so with more quantized weights, you get more MSBs
researcher2#9294: Ok, so lots of wasted space, basically you want this for phones.
researcher2#9294: Or to cram more in memory for larg... |
EfficientNet defines capacity with certain criteria, in order to find out proper hyperparameter values for size.
If you are building on tensorflow and need a huge dataset like imagenet, try using GCP trial to get a VM that packages everything to sharded TFRecords and then use TPUs on colab for super-fast and free tra... |
kindiana#1016: yeah amp is cool, although I'm not sure if I'd call fp16 quantization lol, its rare that you lose any model accuracy moving to it
ntakouris#5483: it is quantization. you're dropping half of the bit resolution. bfloat16 are more efficient than fp16 in deep learning because they changed the mantissa and ex... |
bmk#1476: Or something similarly hilarious
bmk#1476: The paper needs to be as bad as physically possible
StellaAthena#3530: If we replace "GDP" with "yearly mean distance to pluto" I bet nothing about the paper would meaningfully change
bmk#1476: What if we trained a CNN to classify images of proofs as valid or not val... |
Daj#7482: :berk:
StellaAthena#3530: They also seem to have mysteriously forgotten to mention that they applied LOESS before doing the analysis?
Daj#7482: Do you know https://www.tylervigen.com/spurious-correlations ?
StellaAthena#3530: Yup!
Daj#7482: We should replicate like a dozen bad papers
StellaAthena#3530: I've b... |
@StellaAthena count me in
Daj#7482: No more 4am hacking binges feels bad man
bmk#1476: this is fucking glorious
Daj#7482: Agreed
Daj#7482: I've always wanted Eleuther to be the merry pranksters of ML lol
bmk#1476: Ah shoot, I don't know R
Daj#7482: > R
Ah I see where this paper went wrong
bmk#1476: https://cdn.discord... |
Daj#7482: But Elon is nothing if not high variance
bmk#1476: oh god please do not let elon know about the existence of eleuther
bmk#1476: that would have a 90% chance of fucking all our shit up
Louis#0144: downloading a 600GB dataset rn
Louis#0144: a small chunk of common crawl
Louis#0144: Ive been training LMs this la... |
Louis#0144: 😉
Louis#0144: lol
Louis#0144: wait 8 months
Louis#0144: Ill publish it by then
bmk#1476: wait till pile is done
bmk#1476: then extract from that
bmk#1476: pile is much higher signal than CC which is mostly garbage
Louis#0144: proof of concept rn
Louis#0144: Ill use the pile in a few months
3dprint_the_worl... |
Sid#2121: best to ask this all over in TPU pod but: - month of tpu time, you tend to get allocated a certain amount of non-preemptible smaller pods, and a quota of preemptible larger ones
Louis#0144: so if I need to finetune a massive LM I can do that np?
Sid#2121: no hour limits although in practice the larger ones ge... |
Louis#0144: wait but I can use massive models with TPUs on colab
Sid#2121: the TPU CPU's ram has something like 300gb of ram
Louis#0144: I was using reformer-large
Louis#0144: sorry wait
Louis#0144: T5-large
Louis#0144: and like
Sid#2121: pretty sure T5 uses mesh under the hood
Louis#0144: that doesnt fit in my local 4... |
Sid#2121: VRAM is just gpu RAM though, right
Sid#2121: i assumed Louis meant the TPU's RAM
Louis#0144: so how can TF and TPUs fit such a massive model into only 12GB
bmk#1476: Yes but it's disjoint from normal ram
Louis#0144: T5-Large is HUGE
Sid#2121: @Louis mesh
Louis#0144: but colab only uses 1 TPU
Louis#0144: no?
S... |
Louis#0144: otherwise theres no way it would fit
bmk#1476: Also each tpuv2 is only 4 cores apparently
Sid#2121: i don't think that's right
Sid#2121: where did you get that from
bmk#1476: Nvm
bmk#1476: Ok so each tpuv2 core has 8gb
bmk#1476: Vram
Louis#0144: can I run on TFRC using a colab notebook
Sid#2121: no
Sid#2121... |
archivus#7382: https://cdn.discordapp.com/attachments/729741769738158194/760380730110640148/image0.png
archivus#7382: I’m tempted to tweet this but a bunch of doctors follow me 😂
ntakouris#5483: @Louis tfx transform component map reduce is fast
Daj#7482: I don't think anyone would be in the mood for rewriting our wor... |
Noa Nabeshima#0290: I currently don't believe it's a big deal, so I'm looking for other perspectives/other people's knowledge.
bmk#1476: I don't think it is either
bmk#1476: But the communication was absolute garbage
bmk#1476: I don't think they could have phrased it worse
StellaAthena#3530: I think it moved a lot of p... |
Human: The
GPT-3: A
Actual: Attention
Prompt:(NEWSER) – Attention,
Human: all
GPT-3: all
Actual: NFL
Prompt:(NEWSER) – Attention, NFL
Human: lovers
GPT-3: fans:
Actual: announcers.
Prompt:(NEWSER) – Attention, NFL announcers.
Human: You
GPT-3: You |
Actual: It's
Prompt:(NEWSER) – Attention, NFL announcers. It's
Human: going
GPT-3: time
Actual: Patrick
Prompt:(NEWSER) – Attention, NFL announcers. It's Patrick
Human: Mckenzie
GPT-3: Peterson
Actual: Mahomes
Prompt:(NEWSER) – Attention, NFL announcers. It's Patrick Mahomes,
Human: the
GPT-3: not
Actual: not
|
Prompt:(NEWSER) – Attention, NFL announcers. It's Patrick Mahomes, not
Human: the
GPT-3: Patrick
Actual: Pat
Prompt:(NEWSER) – Attention, NFL announcers. It's Patrick Mahomes, not Pat
Human: Mahomes
GPT-3: Mahomes.
Actual: Mahomes.
Prompt:(NEWSER) – Attention, NFL announcers. It's Patrick Mahomes, not Pat Mahomes.... |
Human: today
GPT-3: ask
Actual: ask
Prompt:(NEWSER) – Attention, NFL announcers. It's Patrick Mahomes, not Pat Mahomes. Just ask
Human: him
GPT-3: the
Actual: his
Prompt:(NEWSER) – Attention, NFL announcers. It's Patrick Mahomes, not Pat Mahomes. Just ask his
Human: followers
GPT-3: mom.
Actual: mother.
Prompt:(N... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.