data stringlengths 115 7.61k |
|---|
bmk#1476: actually, i feel like this has already been done before, no?
Noa Nabeshima#0290: sure seems like it should have, but I don't know about it.
bmk#1476: pretty sure shannon ran one
bmk#1476: pretty sure many others have done so too
Noa Nabeshima#0290: Oh, you mean the entropy of language?
bmk#1476: yeah
Noa Nabe... |
bmk#1476: Don't quote me on that
kindiana#1016: take a dataset you are interested in and run it through a gpt tokenizer, it might vary a little depending the specific dataset
Noa Nabeshima#0290: If 4.3, and it makes sense to multiply bits/token by tokens/char, we get .4 bits/char which is less than Shannon's empirical ... |
>they're doing some dank research!
>huh, elon left... prolly just a technicality or something.
>oh, they have an alignment guy!
>what's their approach?
>IDA... wait, that doesn't solve the first step where a human trains values...
>gpt-3 is cool!
>alright, they're not releasing it, probably it's too dangerous...
>okay,... |
Noa Nabeshima#0290: Say you actually had OpenAI's stated mission of making the world better by making AGI before other people and then using it in a responsible and principled way. If the main bottleneck to better systems is money (for researchers and compute), you need to get money *somehow*. If each 100x parameter sc... |
StellaAthena#3530: GPT-3 can easily generate hundreds of millions in revenue over the next 10 years
Noa Nabeshima#0290: At the beginning of 2018 OAI had 10.5M in assets, at the end OAI had 26M in assets
bmk#1476: if i was at OA i would have called it "allowing M$ to use GPT3 for their products" or something
bmk#1476: a... |
bmk#1476: OA themselves still offer GPT3
Noa Nabeshima#0290: making it more important
bmk#1476: or at least it's not precluded
bmk#1476: afaict the only thing that exclusive means is that M$ are the only people outside OA who also have direct access to the weights
bmk#1476: https://cdn.discordapp.com/attachments/72974... |
bmk#1476: this marketing only makes sense from the perspective of someone who doesn't give a shit about alignment
Daj#7482: Which seems in character for MS
Daj#7482: And out of character for the star cast at OA
bmk#1476: ok i guess it makes sense
bmk#1476: but it should be an "oh shit" moment for alignment that OA is n... |
Daj#7482: You guys should read the Biological Anchors report
Daj#7482: When you have a free month
Daj#7482: The "digital professionals" task is a good framing
Daj#7482: But I think Stella is on the right track too
FractalCycle#0001: where is that report?
Daj#7482: And this is just the US government
Daj#7482: Uhh let me... |
bmk#1476: govts suffer from nirvana fallacy more than we do
Daj#7482: And interpretability and accountability for AI is coming and easier than for human brains
StellaAthena#3530: You’d happily pay 100 mill to save 33% on 400 mill.
Daj#7482: I basically do exactly this at my day job
Daj#7482: Government are slow as fuck... |
Noa Nabeshima#0290: Palantir?
FractalCycle#0001: Palantir?
StellaAthena#3530: Lol
StellaAthena#3530: Booz Allen Hamilton
FractalCycle#0001: ah
Daj#7482: Hey that's similar to my company except we're just 5 people in a hole
StellaAthena#3530: > No, but the government thinks we have
To clarify this a bit, we do not have ... |
@Noa Nabeshima Anyways, that’s my rant on why the answer to this question *should be* yes but isn’t.
Daj#7482: I get to do Eleuther stuff during work hours hell yeah
StellaAthena#3530: I mean, it’s 1 pm for me right now
Daj#7482: We won't tattle
StellaAthena#3530: My direct boss doesn’t care how I spend my time as long... |
Noa Nabeshima#0290: Yeah, that's mostly true
bmk#1476: We have enough things we need to do as is
Noa Nabeshima#0290: You don't need to train it, just do inference for the entire dataset in 24 hours
bmk#1476: ?
kindiana#1016: if you ship the model in the compressor/decompressor, you need to double count the bits in the ... |
bmk#1476: and i'm having trouble seeing what we'd do with in-between amounts of money
bmk#1476: then again im not the best at this whole budgeting thing
gwern#1782: it is a dark day in GPT-3 land. the Slack is filled with the moans of the fallen, who have blown through their tier's tokens already (even people on paid t... |
Daj#7482: That's like, the one thing I thought missing for super human AGI
Daj#7482: But that bottleneck is now broken too
bmk#1476: Black mirror is behind the curve on AI
bmk#1476: Still worrying about pesky things like morality
Daj#7482: I can't stress this enough, but I think it's _actually happening_ this time
spir... |
bmk#1476: Yes
Daj#7482: https://openreview.net/forum?id=PdauS7wZBfC
spirit-from-germany#1488: Ok
spirit-from-germany#1488: Thx
Daj#7482: This means that the biggest bottleneck in backprop can be pretty trivially broken
Daj#7482: GPT4 next year confirmed
spirit-from-germany#1488: I am currently in the theme park with my... |
researcher2#9294: I knew it
bmk#1476: Theme parks open already?
spirit-from-germany#1488: Yes, here in Germany had they been reopened since June Or so
researcher2#9294: This seems a bit challenging. How the hell do you learn/remember anything with time based learning when the signal latency is unpredictable? https://cd... |
cfoster0#4356: That would be assuming you're trying to do this over a super unreliable network
cfoster0#4356: If you've got a cluster, just responding in an event based fashion could work well enough, possibly
researcher2#9294: Yeah this would ~~definitely ~~ probably work "in the same room", or even "in the same build... |
researcher2#9294: https://tenor.com/view/elmo-hell-fire-gif-5073559
Daj#7482: I'm already having enough of an existential crisis, thx
Sid#2121: “Wow, do you guys remember *checks notes* five nanoseconds ago when we thought this would be totally impossible?”
Daj#7482: https://media.tenor.com/images/d46785e4ab591e604e61c... |
I'm actually not convinced of this. You still need to store activations and wait for gradients to propagate.
Daj#7482: > I'm actually not convinced of this. You still need to store latents and wait for gradients to propagate.
@AI_WAIFU But only one layer
AI_WAIFU#2844: ?
Daj#7482: You can calculate the gradients of lay... |
cfoster0#4356: The Krotov & Hopfield paper has the recipe
Daj#7482: It seriously is happening, isn't it?
cfoster0#4356: Possibly
Daj#7482: Even if this doesn't turn out to be the paper that cracks it, it seriously feels like the research is converging on world models, locality, attention etc
Daj#7482: Anyone read the D... |
AI_WAIFU#2844: :cake: :cake: :cake: :cake:
StellaAthena#3530: @AI_WAIFU cool. My function should be expressed as containing noise.... it takes on the value `a` with probably `p` and value `b` with probability `1-p` where `a`, `b`, and `p` are polynominal functions of the input
AI_WAIFU#2844: So in that case if p is a f... |
bmk#1476: isnt static ram generally just worse than dram?
bmk#1476: slower, lower density
bmk#1476: the only advantage i know is it's easier to work with because you dont need a clock to drive it
kindiana#1016: sram is much faster but requires more chip area and power
bmk#1476: oh
AI_WAIFU#2844: looks like I've been ou... |
kindiana#1016: (but the problem with flash based ssds is you can only rewrite each cell like 1000 times, so you can't really train off an ssd)
AI_WAIFU#2844: You could always just buy 100x more ssd.
kindiana#1016: ssds are ~20c/gb, ram is ~1000c/gb, so its not really worth it
AI_WAIFU#2844: Yeah, which gets back to my ... |
cc_#1010: unfortunately my finances arent in a position where i can really help much yet but hopefully Soon
bmk#1476: > The model is finished
to elaborate on this, the *training code* is finished
bmk#1476: well, mostly finished
cc_#1010: i am now going to to close this folder for a bit and return to trying to make dril... |
bmk#1476: speaking of hard drives there was recently a big sale on amazon (at least, here in canada there was)
bmk#1476: 10TB for 230 C$
cc_#1010: that is a lot of money and a lot of terabytes
bmk#1476: that's 170 USD
gwern#1782: _shakes his head. even with kryder's law dying, storage is crazy cheap compared to when he... |
Louis#0144: I’ll get back to you
bmk#1476: which evaluation metric is bad?
bmk#1476: the mturk one?
Louis#0144: The issue is that their methods for testing story coherency go against current literature
Louis#0144: Yes
bmk#1476: what's wrong with it?
Louis#0144: Humans are awful at measuring coherency
Louis#0144: Typica... |
bmk#1476: what about a generic "do you like this story better"
Louis#0144: That’s not testing coherency
Louis#0144: That’s testing bias
bmk#1476: im asking if that's ok
Louis#0144: That’s ok yeah
bmk#1476: ok phew
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/762494944065093692/unknown.png
Louis#... |
bmk#1476: gpt3 is too expensive
bmk#1476: we need to ramp up our efforts
chirp#4545: out of curiosity what are you doing with gpt3?
Deleted User#0000: Are there any distributed computing projects like Folding At Home, but focused on ML training. Every now and then I have some gpu capacity to spare, and I'd like to dona... |
StellaAthena#3530: Not that the data has been deduped
researcher2#9294: Good, didn't want anyone getting too excited haha
StellaAthena#3530: @bmk I would wait until eval datasets are done
bmk#1476: Good idea
StellaAthena#3530: https://www.zdnet.com/article/digital-pioneer-geoff-huston-apologises-for-bringing-the-intern... |
gwern#1782: frigging podcasts. no transcript either
Ken#8338: but a few interesting tidbits in it.
bmk#1476: Tfw BPE encoding isn't a homomorphism
gwern#1782: _takes credit for everyone hating on bpes now_
gwern#1782: (me on my deathbed: "either the bpes go or I go, this world ain't big enough for the both")
bmk#1476: ... |
bmk#1476: a while ago, i ran character frequency stats on a small subset of commoncrawl to make this: https://gist.github.com/leogao2/6f0cb98e63126cc40759e58df7c511a8
genai (Immortal Discoveries)#0601: BPE is a way to make your network smaller yet still keep most the accuracy....
genai (Immortal Discoveries)#0601: "ano... |
bmk#1476: gpt2 was 1.542B
kindiana#1016: its just whatever fits well on their cluster lol
bmk#1476: gpt1 was 0.110B
bmk#1476: so gpt1 -> gpt2 was only 10x
bmk#1476: i fear the next one may well just be 1T, oh well
FractalCycle#0001: 1Q when
bmk#1476: but if someone makes it to market first with 1T they might aim higher... |
kindiana#1016: would be funny if AGI is just solved with MLE, no rl required
bmk#1476: where di you find this quote, btw?
bmk#1476: > would be funny if AGI is just solved with MLE, no rl required
@kindiana i've been saying this for a while now
bmk#1476: i don't believe RL is necessary for AGI
aquajet#7800: They have ha... |
kindiana#1016: theoretically tpus are great, you don't need those cuda cores, just mxus, but practically programming them is a big pita
bmk#1476: not to mention only google has em
bmk#1476: and OA is already friends with M$
kindiana#1016: are there tpus for training (not necessarily by google) you can buy without signi... |
bmk#1476: and i *strongly* suspect they optimized it just a tad too much for CNNs
bmk#1476: probably would fall flat on its face for transformers, just considering how it works, and even setting aside the *absolutely puny* amount of memory
FractalCycle#0001: just read the above article on OpenAI. Thinking about the par... |
What do you look for when hiring at OAI?
bmk#1476: Slight modification of that first question: any advice for people who want to get to to speed in alignment?
bmk#1476: I think there's a lot of resources out there for learning DL (and there are way too many DL people around) and not many Alignment people
FractalCycle#0... |
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/763955397584224286/8R9xGpL.png
aquajet#7800: Has someone finetuned a LM of lesswrong?
FractalCycle#0001: What safeguards does OAI have in place to prevent harmful accidents?
researcher2#9294: Are you looking into separate storage and retrieval systems... |
Eddh👽#7290: https://www.wired.com/story/opinion-ai-is-an-ideology-not-a-technology/ good article. I disagree in some parts but still like it a lot. That's how I know a good critique.
bmk#1476: i think i read this (unless i'm thinking of something else) and i *strongly* disagree on many levels
Noa Nabeshima#0290: > Doe... |
Daj#7482: > And it seems true to me that there is an ideology behind the drive to develop AI in tech, which has political implications.
@Eddh👽 sure, but that's a vacuous statement imo
Daj#7482: There's "an ideology" behind toothbrushes too
thenightocean#6100: also if I punch the writer in the face, problem is not the ... |
Daj#7482: (more fun metaphors: AI alignment is neo-catholicism. Be chaste and work hard in this life ("pre singularity") and you will be rewarded with heaven (aligned singularity) later, or if you don't you will go to hell (misalignment))
Daj#7482: In many ways, science is religion except, you know, it works
Chlorokin#... |
Daj#7482: Oh god it _actually_ has a name
Daj#7482: Finally I have a word to bully people that aim for weak alignment
Chlorokin#6581: Yep. Think it was from the Interpersonal Entanglement post in the sequences.
Daj#7482: It's from the Fun Theory Sequence
Daj#7482: Was that in the original sequences? I don't remmeber
bm... |
AI_WAIFU#2844: I get that a lot
Chlorokin#6581: Hotz said something interesting in his Lex interview, basically the most efficient interface between an AI and a human is a catgirl/catboy.
AI_WAIFU#2844: But yes catgirls are a nessesary part of the post singularity utopia.
Daj#7482: > Hotz said something interesting in ... |
Daj#7482: Eh, if humans (as anything other than uploads or post human demigods) still exist we basically failed
bmk#1476: if your definition of human is untampered monkey meat suits then i agree
bmk#1476: i think human + neuralink as a neo-neo-cortex is good enough
Daj#7482: I can't see how that would prevent e.g. wire... |
Daj#7482: Neocortex isn't a cpu
Daj#7482: It's a GPU
bmk#1476: sure, gpu
Daj#7482: Basal Ganglia and midbrain are the CPU
AI_WAIFU#2844: > @AI_WAIFU except we can directly rewrite what people want
To quote EY: "AAAAIIIIIIEEEEEEEEE"
https://www.lesswrong.com/posts/Py3uGnncqXuEfPtQp/interpersonal-entanglement
bmk#1476: t... |
@bmk yea, so with only a disconnected GPU there's not much "intelligence" there
Chlorokin#6581: What you have to realize is moral realism is true and it is catgirls.
Sid#2121: Pinned a message.
bmk#1476: my point is we can turn off parts of the primitive mind without just yeeting it entirely out of existence
Daj#7482: ... |
bmk#1476: we can write a better bootloader for our neocortex
Sid#2121: ok, you first
AI_WAIFU#2844: > I'm saying can and will do so if we don't solve the problem
Ok, although I think we need to be really careful with that because "preference rewriting" is some proper evil supervillan shit.
Daj#7482: That's like saying ... |
@AI_WAIFU
bmk#1476: everything is preference modification
Daj#7482: Then why not modify yourself to desire nothing?
Chlorokin#6581: How can you be "weakly aligned" when edge instantiation is a thing.
bmk#1476: i feel like this is a much more minimal preference modification
Daj#7482: > How can you be "weakly aligned" wh... |
Chlorokin#6581: > We don't just do surgery without washing our hands, either
@Daj Yes, but presumably the weak alignment people have a response to that line.
AI_WAIFU#2844: > Then why not modify yourself to desire nothing?
Was this addressed to me?
Daj#7482: > Was this addressed to me?
@AI_WAIFU To bmk
Daj#7482: > @Daj... |
Daj#7482: Yep, still not giving upbecause of that
Chlorokin#6581: One of the upsides of being a midwit is I don't feel morally responsible for contributing to the imminent destruction of our light cone.
Daj#7482: That's an excuse, being someone that can make a large contribution doesn't _feel_ like being big brain
Daj#... |
thenightocean#6100: so maybe we already have our best shot for now
Daj#7482: > But Paul Christiano is working for OpenAi, right? Is he is considered the best expert in Ai alignement atm ?
@thenightocean He's kinda the poster boy for prosaic AI alignment
Daj#7482: I don't like deferring to authority like this
Daj#7482: ... |
Daj#7482: I can try out arguments here before releasing them to the public
Daj#7482: But yes, experiments are good if we can do them
Sid#2121: I also feel like there's a big problem in the alignment field with everyone feeling like they have to do everything on their own
Sid#2121: seems like there's a lack of teamwork ... |
bmk#1476: what's an example of a thing that philosophy has solved?
Daj#7482: Deconfuse the philosophy, solve the science
bmk#1476: or not solved but progressed
Daj#7482: Lets not go into "solved" because you can always find someone who disagrees
bmk#1476: it feels like people just keep departing further into their own ... |
AI_WAIFU#2844: Do we even need that much philosophical progress though? IMO the first reasonable AI safety proposal came from bostrom and was basically "we kick all the complex philosophy down the road after dealing with the immediately pressing issues"
Daj#7482: Denett solved consciousness for all I'm concerned
Daj#74... |
Daj#7482: > You're a negative utilitarian?
@Chlorokin Ehhhh kinda
Daj#7482: I'm a utilitarian that thinks that the average suffering is much worse than the average good thing
bmk#1476: i'm not sure i would call myself a utilitarian
Daj#7482: Relevant MIRI post btw: https://intelligence.org/2013/11/04/from-philosophy-to... |
bmk#1476: i'm asking a really technical question
Daj#7482: I know, that wasn't aimed at you
bmk#1476: what does it mean for one state to be twice as desireable as another?
Daj#7482: _That's what we have to figure out!_
Daj#7482: You're making my point for me
bmk#1476: if you ask me, you can only define a well ordering ... |
Daj#7482: You're basically saying "Wow there's a hard problem here" and I'm responding "Yes! We should try to solve it!"
Daj#7482: If we fail, we fail
bmk#1476: is there already literature on well ordered but non scalar utility?
bmk#1476: i dont want to reinvent the wheel
Daj#7482: But it seems weird to say "Wow this p... |
Daj#7482: Sure, once the philosophy is deconfused, that's what's needed next
Daj#7482: If you have a comparative advantage for math, for heaven's sake do math
Daj#7482: I was unfortunately born with the severe cognitive defect called "Being an Artist"
Daj#7482: Trying my best to treat it with code therapy hah
Chlorokin... |
Chlorokin#6581: Dear god, the thought of that as an actual kink just popped into my head.
Daj#7482: ~~tbf, I also (mostly ironically) think some LW people should be bullied lol~~
bmk#1476: I'm going to unrefine my earlier search query: is there literature on the perspective of utility as assigning values to *world stat... |
AI_WAIFU#2844: Although I think it's more useful to think of utility over programs that generate the world rather than over world states.
Daj#7482: > @Sid I don't tbh, does that make me even nerdier than the baseline in here
@bmk No, me and Sid are just the artists among autists
bmk#1476: Ah
Sid#2121: lmao
Sid#2121: I'... |
@bmk Ah yes, Poe's Law
Daj#7482: (this is a joke)
FractalCycle#0001: i have a friend into philosophy, he's probably read Schmidhuber
bmk#1476: A la "best way to figure out how to do something on Linux is to say Linux can't do it"
Daj#7482: > But postironically.
@AI_WAIFU :bigbrain:
Daj#7482: Schmidhuber is pretty decen... |
To work on this stuff: _Just read tons of stuff_. You'll develop a "taste" and a list of people and ideas you're interested in
Daj#7482: > wait is Schmidhuber real? i thought that was fake
@FractalCycle :berk:
FractalCycle#0001: aaaaaaaaaa
FractalCycle#0001: https://en.wikipedia.org/wiki/J%C3%BCrgen_Schmidhuber
Daj#748... |
thenightocean#6100: "According to The Guardian,[35] Schmidhuber complained in a "scathing 2015 article" that fellow deep learning researchers Geoffrey Hinton, Yann LeCun and Yoshua Bengio "heavily cite each other," but "fail to credit the pioneers of the field", allegedly understating the contributions of Schmidhuber a... |
thenightocean#6100: whats with his company btw (Nnaisense)? I would expect great things based on his reputation
Daj#7482: I said _mostly_ correct
Daj#7482: lol
AI_WAIFU#2844: I think their buissness model was to get bought out but that never happend.
AI_WAIFU#2844: Because they're not cool
Daj#7482: I am trying to lobb... |
Daj#7482: I studied at the same uni as him though
AI_WAIFU#2844: IDSIA
Daj#7482: He used ot be
Daj#7482: > >mario says "A.G.I. will take 1,000 years to make and it will require semantic Bayesian networks complexity to mimic the human sacredness."
>
> >luigi says "If it's not at scale, it's going to fail!"
@FractalCycle... |
bmk#1476: this sounds a lot like inner alignment
AI_WAIFU#2844: It's very related to inner alignment.
Sid#2121: > eleuther has a massive meme footprint
@bmk unironic opinion: good alignment memes could be more effective to alignment as a whole than our alignment debates here
bmk#1476: to be exact, it sounds like a stri... |
Daj#7482: Now you're thinking with MIRI
bmk#1476: and there's a chance something like it could be made to work
AI_WAIFU#2844: I mean, that's just constructing a utilty function.
bmk#1476: or what if you weight each world-state by its likelihood
bmk#1476: yeah, but im constructing it from just a well ordering
bmk#1476: ... |
bmk#1476: that's what i'm saying
Daj#7482: I also think ordering is way more important than scalars btw
AI_WAIFU#2844: ^
bmk#1476: is it possible to build a reasonable enough utility function that doesnt break xyz DT given only an ordering
bmk#1476: my first reaction was "of course not"
bmk#1476: my second reaction was... |
Daj#7482: For the record: I think formalizing suffering may be the highest impact thing we as a species could accomplish for alignment
AI_WAIFU#2844: So you can still get actionables even if your information is incomplete
bmk#1476: also another advantage of doing partial orderings vs utility functions: each person defi... |
AI_WAIFU#2844: ok i gotta go
Daj#7482: See ya! Always fun to talk
bmk#1476: cya
Louis#0144: weird feeling when microsoft excel is literally the biggest piece of software i have on my computer by a long sholt
Louis#0144: 🤷♂️
chirp#4545: me just now on my side project:
> wow, my model outputs really suck. i'm gonna lo... |
bmk#1476: and if we ever need to collect more images it's quite easy
bmk#1476: just get one of them streetview cars and drive around for a few months
bmk#1476: that would actually be an incredibly fun job, streetview car driving
gwern#1782: @zphang the data scaling law so far is sublinear, and I'd argue you could easil... |
Noa Nabeshima#0290: Yes, that seems like a big problem.
Noa Nabeshima#0290: Paul Christiano's iterated amplification proposal seems to rely on lots of things that seem like they might intuitively work, so this seems like a good thing to check.
kindiana#1016: would be an interesting experiment for one of the big ML conf... |
but I expect to have more https://cdn.discordapp.com/attachments/729741769738158194/766048766402756628/unknown.png
Noa Nabeshima#0290: https://arxiv.org/pdf/1806.07572.pdf
StellaAthena#3530: @Noa Nabeshima the \cdot is a placeholder symbol for the input of an implicitly defined function.
StellaAthena#3530: the kernel ... |
Deleted User#0000: for the NTK i find the presentation in Jascha et al's paper easier to read
Deleted User#0000: also i made some notes for the Jacot paper some tiem ago -- maybe they are helpful https://github.com/damaru2/ntk/blob/master/notes/Neural_Tangent_kernels___Jacot_et_al.pdf
Deleted User#0000: im also down to... |
bmk#1476: Also lots of cool things emerged from LW which is also well past its prime at this point too
gwern#1782: (and people on SL4 missed the golden age of Extropy ML)
AI_WAIFU#2844: Are there any archives of Extropy ML? Back when I was young I read SL4, but I was never able to find the "extropians" that they referr... |
bmk#1476: ~~also he's said the phrase *less wrong* on multiple occasions~~
bmk#1476: oh
bmk#1476: that's a good thing then
bmk#1476: still, having it as a political force isn't a bad thing
bmk#1476: i wonder how many people in this server would count as vaguely transhumanist
AI_WAIFU#2844: Yup. The other thing is that ... |
bmk#1476: probably wont be reading it then
AI_WAIFU#2844: IMO it's better to just keep up with the alignment forum/do your own research.
bmk#1476: good idea
bmk#1476: ~~i still need to read the af sequences dammit~~
AI_WAIFU#2844: ~~don't forget the comments~~
bmk#1476: ~~***EEEEEEEEEEE***~~
AI_WAIFU#2844: Ok I need sl... |
Daj#7482: This is why normies don't associate with us lol
Deleted User#0000: i actually told Anders about furry GANs before arfa did it, and he loved the idea xD
and then he posted some tweets defending furries too xD
Deleted User#0000: > Haha you haven't seen?
> https://youtu.be/5XKTdx5BWic
@Daj no i havent seen it, ... |
Deleted User#0000: ah didnt know he gave a talk with virtual rooms after
Deleted User#0000: but yeah conversations with him are pretty awesome
Daj#7482: I talked to him once or twice then he never answered my follow up emails haha
Daj#7482: (this is totally fine, I know how professors are haha)
Deleted User#0000: ah no... |
bmk#1476: I'm not surprised
Daj#7482: https://cdn.discordapp.com/attachments/729741769738158194/767073521628479498/28c.png
Deleted User#0000: > Damn, I'm really missing out too, I don't think there are any transhumanist groups up here, let alone transhumanist furry groups
@bmk lets make a furry transhumanist group in ... |
Daj#7482: My theory moreover is that this stuff is super common, but only people with autism have the lack of social skills to admit to it publicly, so it becomes associated with internet/spectrum culture
Daj#7482: (to be clear: this is shitposting, not an actual argument)
Daj#7482: I remember what shocked me most abou... |
bmk#1476: unfortunately, yes
zphang#7252: it's more deviant than perversion, imo
but it's also a genre marker, and once a word become that it takes on a life of its own (like word "fantasy")
bmk#1476: i chose "perversion" because it sort of has a bit more of a connotation that it's "against the natural order of things"... |
Daj#7482: (btw I remember reading somewhere a statistic that the average age of furries is 19, which explains everything about their culture tbh)
gwern#1782: well, autism rates keep going up over time...
Daj#7482: > well, autism rates keep going up over time...
@gwern Do you actually believe this is true? I've heard lo... |
StellaAthena#3530: This doesn’t seem like a necessary fact, but rather a contingent one about the profitable ventures of major companies.
StellaAthena#3530: @miguelos that’s because words like “de facto surveillance state” and “make billions by freeloading on work done by non-employe people” are not popular words to sa... |
miguelos#7956: I could analyze how food affects my productivity/mood/energy/symptoms.
miguelos#7956: I'm not sure what aspects of a human are static and can be definitely learned. For example, I might have enough data to model my digestive system or my sleep requirements, but these might change over time due to externa... |
3dprint_the_world#6486: hah. But joking aside, benefits to the group aren't usually that much of an evolutionary forcing function. Benefits to one's immediate siblings are far more important.
miguelos#7956: Then wouldn't rates of autism and homosexuality rise to appear in every family?
3dprint_the_world#6486: well it w... |
3dprint_the_world#6486: exactly.
miguelos#7956: Who do you think wrote those records? Probably not normies.
StellaAthena#3530: Actually, our oldest records are *exceptionally* mundane
3dprint_the_world#6486: yeah
miguelos#7956: What's the purpose of classifying autism/homosexuality as evolutionary useful? I can't see a... |
bmk#1476: > @bmk so do many other factors like GDP per capita and quality of health services.
@3dprint_the_world also hold up did you just admit that *healthcare causes autism* (conspirasy theory intensifies)
3dprint_the_world#6486: exactly
3dprint_the_world#6486: lol
3dprint_the_world#6486: I guess I did
miguelos#7956... |
StellaAthena#3530: The US is one of those first-world countries
bmk#1476: @miguelos i think it would be best if you first list which "special privileges" you are referring to
miguelos#7956: Not being discriminated against is one of them.
StellaAthena#3530: https://en.wikipedia.org/wiki/List_of_U.S._jurisdictions_bannin... |
StellaAthena#3530: Congrats?
guac#4716: what's the deciding factor in off-topic vs. general discussion?
StellaAthena#3530: Things move to off-topic when people get asked to kick a convo out of general
bmk#1476: Go to #off-topic to continue politics discussions or I will have to kick you from the server
bmk#1476: @migue... |
@StellaAthena imo politics isn't *necessarily* a mindkiller, but it very often becomes one
bmk#1476: My favorite semi rationalist thing relating to this is waitbutwhy's "story of us" series
bmk#1476: It really dissects (one perspective of) why this happens
3dprint_the_world#6486: Perhaps the causal relationship is the ... |
ERROR: type should be string, got "https://github.com/huggingface/transformers\nkindiana#1016: huh that image is too big to inline, but there is actually a decent flowchart for transformers `https://raw.githubusercontent.com/javismiles/X-Ray-Transformer/master/images/xray-transformer-14000px-c.gif`\n3dprint_the_world#6486: that's a great infographic but pretty meaningless to someone who isn't already familiar with transformers\nNik McFly#3288: @kindiana thanks!\nkindiana#1016: yeah I agree haha\nNik McFly#3288: I'm a fast-learner 🙂\n3dprint_the_world#6486: maybe start with http://jalammar.github.io/illustrated-transformer/\n3dprint_the_world#6486: although that article, as good as it is, only describes *what* transformers do, not *why*. Which is a far more complicated thing to understand.\n3dprint_the_world#6486: but it's a start\ncfoster0#4356: In terms of tech stack, transformers are made of embarrassingly-simple components, so they should be fairly portable to different stacks. But at GPT-3 scale you're talking CUDA running on NVIDIA GPUs, with some kind of numerical library such as PyTorch on top, and probably a model-serving setup if you're doing it in production\nbmk#1476: GPT3\n\nML complexity: 1%\nEngineering complexity: 99%\n3dprint_the_world#6486: > In terms of tech stack, transformers are made of embarrassingly-simple components, so they should be fairly portable to different stacks. But at GPT-3 scale you're talking CUDA running on NVIDIA GPUs, with some kind of numerical library such as PyTorch on top, and probably a model-serving setup if you're doing it in production\n@cfoster0 lol that's not even at GPT-3 scale. At our work we do that for vision models which are like 40M parameters.\n3dprint_the_world#6486: yeah engineering complexity is the big thing. But if you just want to understand the core ideas without implementing it yourself, you can just focus on the maths.\nNik McFly#3288: in fact, I'm thinking about the whole architecture with implementation\n3dprint_the_world#6486: if you want to implement your *own* GPT-3.... well good luck\n3dprint_the_world#6486: lol" |
Nik McFly#3288: not just a concept as a sum of ideas
3dprint_the_world#6486: make sure you have millions of dollars of cash on hand
bmk#1476: :guilty:
Nik McFly#3288: it's not SO important to have them for thinking 🙂
bmk#1476: > if you want to implement your own GPT-3
> make sure you have millions of dollars of cash ... |
Nik McFly#3288: surely, we all can have that in our dreams
3dprint_the_world#6486: well it's still true isn't it
Nik McFly#3288: > EleutherAI: :guilty:
@bmk haha
bmk#1476: fact check we do not have millions of dollars
Nik McFly#3288: that's my question to this community also
Nik McFly#3288: true
3dprint_the_world#6486:... |
3dprint_the_world#6486: hence
> or have *very* good relations with someone at google
@3dprint_the_world
Nik McFly#3288: afaiu the main thing about tech is about people, so humane tech go first to deliver resources for the idea
bmk#1476: not here
bmk#1476: the main thing here is compute
bmk#1476: compute compute and mor... |
bmk#1476: i have poor opinions of art on the blockchain
Nik McFly#3288: now grasping the gpt-3 concept
bmk#1476: i have poor opinions of fine art in general but especially nft art
Nik McFly#3288: that's the industry i know it's growing this moment
cfoster0#4356: 👀
Nik McFly#3288: i know people making money there
guac#... |
Nik McFly#3288: i just want to do art and throw there and see what happens
guac#4716: is there blockchain for music?
bmk#1476: can you give me the elevator pitch for nft art, assuming im already not a fan of fine art as a whole
Nik McFly#3288: no illusions
cfoster0#4356: there are multiple @guac
bmk#1476: i'm into cryp... |
@bmk it's okay too. i'm cynical on the tech at all. but you wouldn't believe what's the project with blockchain application i'm involved into
bmk#1476: i cant parse that
bmk#1476: youre cynical of.. the tech? or applications?
Nik McFly#3288: of the tech at all
Nik McFly#3288: all of the tech
bmk#1476: you're cynical of... |
Nik McFly#3288: yep
bmk#1476: huh
bmk#1476: that's the exact opposite of me tbh
Nik McFly#3288: haha
bmk#1476: i think the tech is solid and people are building castles of shit on top
Nik McFly#3288: 🙂
bmk#1476: but yeah to each their own
Nik McFly#3288: i'll try to find my favorite meme about it
guac#4716: i don't th... |
Nik McFly#3288: society coding
Nik McFly#3288: society is a tech also
Nik McFly#3288: language is a software
shgidi#0284: Hi, is there a status doc or something for this project? will be happy to know where is stands now...
Daj#7482: Hey @shgidi ! We have https://docs.google.com/document/d/1yOnxEMlU57M8YFlQC3XNOvyMVX2E... |
FractalCycle#0001: sam altman ssc meetup date released: https://joshuafox.com/ssc-online-meetups/ 11/08 (although it appeared to be later last i checked, so i guess they moved it sooner)
StellaAthena#3530: @Sid what causes the break between sequences? The fact that monotonicity breaks?
Sid#2121: @StellaAthena the input... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.