35 ms·

GPT-3

291 points by cocoflunchy 6y ago

Sniffnoy 6y ago
Something seems to be wrong with the footnotes on this page... Edit: Nevermind, my browser seems to just be screwing up on Gwern's footnotes in general at the moment.
- renewiltord 6y ago
  On my desktop, I hover over and they pop up a little window with the footnote in it. That looks like the intended behaviour but it must be failing on your platform.
throwaway4666 6y ago
>Like the genomics revolution where a few far-sighted seers extrapolated that the necessary n for GWASes would increase exponentially & deliver powerful PGSes soon, while sober experts wrung their hands over “missing heritability” & the miraculous complexity of biology & scoff about how such n requirements proved GWAS was a failed paradigm, the future arrived at first slowly and then quickly. Yet, here we are: all honor to the fanatics, and shame and humiliation to the critics! No. Actual geneticists are still pretty skeptical about GWASes because they tell us almost nothing about, well, the genetics behind complex traits. It's all good and well running GWASes for literally anything (see: twitter.com/sbotgwa or that dude who got a pretty good PGS from correlating a country's GDP with the genotypes of Arabidopsis thaliana available in that country) but that's virtually useless for serious research or if you want to know how genes work. Actually figuring out to what extent a trait is genetically determined usually involves much more complex methods (e.g. mendelian randomization) and knock-out experiments on animal models, which is all terribly expensive and tedious. But that's how actual genetics works, not waving a magic wand of +15% heritability.
- jhrmnn 6y ago
  Agreed. To put it simply, if you model a complex system with another complex system, of which you have a similar level of understanding, then even if the model is very faithful, you gain very little. Unless the model is transferable, so you can predict, which AFAIK is not the case here.
- dash2 6y ago
  I think the point is not that GWAS can do everything, but that many people 5-10 years ago thought that they had a fundamental flaw - where was all the heritability? It has turned out so far that the heritability is there and discoverable with big enough sample sizes. But of course, you are right that GWAS don't tell us much about the very large space between strands of DNA and social or behavioural outcomes. They are a tool like others. (Mendelian randomization is hardly a panacea - how credible are those exclusion restrictions, usually?)
- esyir 6y ago
  Unless something has drastically changed in the past 5 odd years, Actual Geneticists (tm) happily use GWAS in many investigations of complex traits. They're often an early step in the overall pipeline in which they trawl for potential candidates for that target of interest, after which they go on to said more complex methods. Why? You've already said so yourself. Those are expensive and tedious, and searching across the entire pool of human genetics with them is an exercise in futility.
  - throwaway4666 6y ago
    The important part in your post is >They're often an early step I think we're in agreement here, I'm just arguing that "woah, look at all those correlations" isn't a breakthrough or 'genomic revolution' in any sense of the word as far as our understanding of human genetics is concerned.
  - hyperbovine 6y ago
    Seemingly, most of the complex traits for which GWA is an effective methodology (a few SNPs of large effect size) have already been discovered. More and more these days, I’m seeing association studies that fail to yield any hits. Whether this is due to sample size, polygenicity, or some other model failure remains to be seen.
- folli 6y ago
  Can you post the link about the GDP/A. thaliana correlation? That sounds hilarious, and Google is failing me to find it myself.
  - throwaway4666 6y ago
    Found it: https://mobile.twitter.com/pastramimachine/status/1263937115684892674 https://mobile.twitter.com/pastramimachine/status/1263937115... Edit: if it doesn't work try this i guess? https://www.google.com/url?q=https://mobile.twitter.com/pastramimachine/status/1263937115684892674&sa=U&ved=2ahUKEwjyz5eCo5rqAhUnxIUKHTl-BBwQFjAAegQIARAB&usg=AOvVaw1EqJAtVmJyjAEa3TiP3EFe https://www.google.com/url?q=https://mobile.twitter.com/past...
    - ralfd 6y ago
      Throws a Twitter error for me.
- bobcostas55 6y ago
  Who cares how genes work? The important part is the real predictive power of the PGS.
  - throwaway4666 6y ago
    'Predictive' is just a glorified word for 'correlating'. Staring at a bunch of correlations only gets you so far and doing only that certainly isn't science. If you have no idea how genes work you're going to be dismayed when all your fancy correlations don't work anymore, as people gathering genomic data from outside the European biobanks are starting to find out.
- Gatsky 6y ago
  I agree, holding up GWAS as some kind of lesson about scientific progress seems pretty silly. I don't think the 'sober experts' deserve shame and humiliation considering that PGSes have had minimal impact on human health.
- zxcb1 6y ago
  GWAS have been shown to be practically meaningless. How should this paragraph be interpreted, that transformers can make sense of DNA patterns?
- gwern 6y ago
  Nope. You're wrong. Where do you think your MRs or your sibling comparisons are coming from? You're not getting anywhere with weak instruments from a few significant hits. You're also wrong that the only thing of value is inferring "how genes work", and that is the sort of extremely blinkered mechanism-centric attitude which blinded people to GWASes working, because gosh, it would be awful if polygenicity was true, because how would we build any 'scientific theories' on this? (Cue Turkheimer.) And yes, those PGSes are useful for all sorts of things like selective sweeps, enrichments, and clinical instruments, because of incremental validity. (By the way, what does '15% heritability' refer to? I sure hope that, since you're claiming to be an expert here, you aren't confusing heritability with PGS power, like so many supposed human geneticists insist on doing...)
  - throwaway4666 6y ago
    I didn't say GWASes were useless, just that it's absurd to consider them to be a 'revolution'. The actual revolution would be second- and third-generation sequencing which enabled GWASes and a bunch of much more useful things. GWAS is, in effect, just a bunch of correlations. It's just the very starting point to an actual scientific analysis, because 'you have to start somewhere'. If you don't go beyond and investigate, you're not doing science. Everyone in the genomics community agrees to this, and literally every paper that investigates the causes of genetic diseases goes in the introduction like 'GWASes sure look nice but we still have no idea how things work with them so in this paper I present a method to do...' I notice you failed to address many of the spurious correlations drawn by the GWAS bot or the A. thaliana vs. GDP prediction. That it doesn't raise any red flag to you doesn't speak well as to your ability to approach the field of genomics. >You're also wrong that the only thing of value is inferring "how genes work" Yes it is, that's literally what genetics is about. Otherwise you're back to making a bunch of correlations. If you want a deep understanding of disease, design effective drugs, or even do proper gene editing, you have to understand what genes do. It seems ridiculous to have to say it. >By the way, what does '15% heritability' refer to? It is what we in the less-rationalist circles refer to as a 'joke'.
fock 6y ago
Well, I take the the opposite stand: GPT-3 = Give a 3rd grader wikipedia and some paper. While it's certainly fun how it's relatively coherent in writing text, I couldn't see yet, how it links syntactically correct text to actual facts. Which in my opinion is the difference between 1e100 apes with typewriters and aforementioned 3rd grader.
- throwaway4666 6y ago
  I think gwern's point (whose opinion on deep learning I respect a lot more than his on genetics) is that GPT-3 is nearly not the end of the story. OpenAI released a GPT pretty much every year and each one is a spectacular improvement on the previous one with no sign of that trend ever stopping. If size is really what matters at all, there's no telling what GPT-4 or GPT-5 might be capable of, let alone a giant GPT run by state-sized actors.
  - ipsum2 6y ago
    The machine that trained GPT-3 is one of the largest GPU clusters in the world at the moment. (10k v100 GPUs hooked up with fast interconnect). For comparison, Nvidia's "Selene" supercomputer only has 2k a100 GPUs and is #7 fastest supercomputer in the world: https://www.hpcwire.com/2020/06/22/nvidia-nabs-7-spot-on-top500-with-selene-launches-a100-pcie-cards/ https://www.hpcwire.com/2020/06/22/nvidia-nabs-7-spot-on-top... a100 GPUs are around 3x faster than v100 GPUs. It would take some time for governments to catch up :)
    - qeternity 6y ago
      1. This is a list of publicly known supercomputers (it’s likely that the intelligence community have larger secret assets). 2. Most of the largest super computers are owned by states.
      - throwawaygh 6y ago
        Lots of large companies don't include their clusters on those supercomputer lists. Including Google. IDK about DoD, but I wouldn't be surprised if they have top 10 clusters we don't know about. The list of largest supercomputers is definitely only a list of the largest disclosed supercomputers. That said, only so many actors could've secretly sourced 10K+ v100s with fast interconnect...
skrause 6y ago
The GUID Partition Table is already at its third version?
- anticensor 6y ago
  No, Generative Pre-Trained Transformer.
  - skrause 6y ago
    I know, it's just a stupid acronym because it's already used for something else (the partition table).
    - Smaug123 6y ago
      Oh boy, do I have bad news for you: Wikipedia lists ten overloaded acronyms starting with "AA" alone!
      - skrause 6y ago
        But they are usually in unrelated fields.
Tepix 6y ago
I'm happy to see it being released. What is the minimum hardware required to run this locally? Or what's the cheapest way to run this model (even at barely acceptable performance) at a cloud provider?
- bjourne 6y ago
  Probably out of reach, but you can run GTP-2 on Google Colab: https://colab.research.google.com/github/ilopezfr/gpt-2/blob/master/gpt-2-playground_.ipynb https://colab.research.google.com/github/ilopezfr/gpt-2/blob...
- p1esk 6y ago
  Any server with 512GB of RAM. So basically a couple of bucks per hour, to generate a few words per minute.
  - Tepix 6y ago
    You don't need fancy GPUs with lots of RAM?
    - p1esk 6y ago
      If you’re ok with generating a few words per minute, no.
dgrabla 6y ago
What book should I read to be able to understand this text? No papers please. I don't understand what scaling a model means.
- mkl 6y ago
  This is about new research, which mostly lives in papers and articles (like this) about the papers. It won't show up in introductory books for a while, so if you're unwilling to read or even look at papers, you won't be able to understand details of new research. Scaling a model is just like it sounds: more data fed into a bigger network with more parameters. The gist of what this article is saying about scaling is that there's no sign of diminishing returns yet in terms of what the network can do and how well it generalises as the number of parameters is increased: the "more parameters = better performance" trend continues up to the enormous size of the full GPT-3 model, with no indication that even bigger models won't have even better performance. Here is the GPT-3 paper: https://arxiv.org/pdf/2005.14165.pdf https://arxiv.org/pdf/2005.14165.pdf If you really want to understand, skim this, and focus especially on the graphs, as they show the scaling. The x axis is usually model size, and the y axis is mostly accuracy or "loss" (~error).
chillee 6y ago
Personally, as a ML researcher, I find GPT-3 very unsatisfying. There aren't any novel architectural details, it doesn't improve our "fundamental" understanding of the field, and it requires the type of computation I have no chance of getting. As a fan of the field, however, it is undoubtedly very cool. I still think there's tasks that GPT-3 has no chance of tackling (say, generating code to solve a novel programming task), but the bitter lesson is a bitter one indeed...
- FeepingCreature 6y ago
  It's kind of LHC-like, isn't it? Same physics "but larger".. "will it work if we scale it up? Apparently: yes." I think the big point, and what scares me a bit, is that we have yet to discover any sort of fundamental conceptual limit to the Transformer architecture.
  - sooheon 6y ago
    Each attention block in the Transformer models a fully connected graph (with the attention heads being learned edge attributes). A graph is the most general data structure possible, so yeah, I don't think there's really a fundamental limitation to them, just a computational one. Latest papers from ICLR explore how they fully model CNNs and RNNs, for example, and I'm sure papers on their theoretical equivalence with GNNs are coming.
    - contravariant 6y ago
      > I'm sure papers on their theoretical equivalence with GNNs are coming. Not exactly a paper, but you mean something like this article: https://graphdeeplearning.github.io/post/transformers-are-gnns/ https://graphdeeplearning.github.io/post/transformers-are-gn... ?
      - sooheon 6y ago
        Pretty much that, yeah. Thanks for the link. That kind of article sets the intuition, and papers like these give it a better theoretical treatment: - https://openreview.net/pdf?id=HJlnC1rKPB https://openreview.net/pdf?id=HJlnC1rKPB - https://openreview.net/pdf?id=ByxRM0Ntvr https://openreview.net/pdf?id=ByxRM0Ntvr
nl 6y ago
Winograd schemas falling at 10T parameters is interesting. That's probably only 5 years off. If we can build something capable of passing Winograd schema, then it can probably write working non-trivial computer programs from plain text. Google's PEGASUS summarization model[1] has learned to count up to five (which is amazing!!). That's "only" 568M parameters. It's be interesting to see GPT-3 fine tuned against the PEGASUS objective function. [1] https://ai.googleblog.com/2020/06/pegasus-state-of-art-model-for.html https://ai.googleblog.com/2020/06/pegasus-state-of-art-model...
- LordDragonfang 6y ago
  For those curious the "counting" is at the end of the article, and really is quite impressive: >Following this post is an example article from the XSum dataset along with the model-generated abstractive summary. The model correctly abstracts and paraphrases four named frigates (HMS Cumberland, HMS Campbeltown, HMS Chatham and HMS Cornwall) as “four Royal Navy frigates”, something an extractive approach could not do since “four” is not mentioned anywhere. Was this a fluke or did the model actually count? One way to find out is to add and remove ships to see if the count changes. >As can be seen below, the model successfully “counts” ships from 2 to 5. However, when we add a sixth ship, the “HMS Alphabet”, it miscounts it as “seven”. So it appears the model has learned to count small numbers of items in a list, but does not yet generalize as elegantly as we would hope. Still, we think this rudimentary counting ability is impressive as it was not explicitly programmed into the model, and it demonstrates a limited amount of “symbolic reasoning” by the model.
  - nl 6y ago
    It's one of the most amazing and surprising things I've seen in the last 12 months in machine learning (I follow and work in the field). It surprises me a lot more than the excellent performance of GPT-3 on text generation for example. GPT-3 is amazing but looking at GPT-1 -> GPT-2 -> GPT-3 it isn't surprising. Counting on the other hand is something I wouldn't have expected from a summarizer.
    - FeepingCreature 6y ago
personjerry 6y ago
I'm failing to connect the anecdotes with the conclusion. He's claiming the the ML is scaling well but then gives the data on how GPT-3 is expensively brute-forcing its way to "success". To me it just seems like what supercomputing is to normal computing: It makes the computationally expensive stuff do-able in a reasonable amount of time, or gives diminishing returns on existing algorithms. But it doesn't magic in any real advancements. The problem in AI/ML and the concept of "AI winter" to me was always the barrier of the fact that we're just doing predictions, with no deep meaning or comprehension of features. The layman thinks there's magic, but there's not, so when that truth is revealed there will be problems. There's nothing intelligent about our Artificial Intelligence; we're just doing statistics with big data and added sprinkles. OpenAI just proved they could do statistics with even bigger data and more expensive sprinkles. Has their work really shown we can get past that core problem? Personally, I don't see it.
- LordDragonfang 6y ago
  >we're just doing statistics with big data and added sprinkles. I mean, you could pretty much say that's how the human brain works, couldn't you?
  - personjerry 6y ago
    That comparison is like comparing gardening my backyard to agricultural farming. Technically the core concept is similar, but I think we both know that the scale and depth makes them not at all the same.
    - H8crilA 6y ago
      There's a continuity of small changes between your backyard and industrial farming. Just like there's a continuity of small changes between you and your cat.
      - p1esk 6y ago
        There’s a continuity of small changes between any two things.
SonOfLilit 6y ago
Just so we're all on the same page, this is a random GPT-3 sample (I clicked "Random" twice, the first one was a short Wikipedia-like article, this was the second): https://read-the-samples.netlify.app/sample_1749 https://read-the-samples.netlify.app/sample_1749 As far as I'm concerned, I would probably pick this over a random fanfic in a Turing Test.
- mkl 6y ago
  That's amazing. The only problems I caught are that there's a spurious double quote right at the start, and Murphy seems to have two children, a son and a daughter, but then later refers to "the boys". To really evaluate these samples, we need a way to search for phrases in the training data, to see how much is just learned and copied. I've tried Google and DDG, but not found anything.
  - hyperman1 6y ago
    I also wonder how much is invented, and how much is simply copied. I saw this one by GPT3 https://read-the-samples.netlify.app/sample_2007 https://read-the-samples.netlify.app/sample_2007 and compare it to e.g. https://www.bbc.com/news/world-europe-35869985 https://www.bbc.com/news/world-europe-35869985 GPT-3 has quite a lot of correct facts, manages to mix fouad belkacem and Sharia4Belgium into it, etc... I suspect this is what happens when there is not enough training data: You get a lookalike, not something new. To be clear: It's impressive what GPT-3 is capable of, and it's very probable something better is comming in the near future. (It's also quite amusing how a lot of these terrorists get first killed by the police and then thrown into jail for 20 years. That'll teach their corpses! )
    - gomox 6y ago
      > I also wonder how much is invented, and how much is simply copied Well, the average human probably does not excel at this metric either
  - taneq 6y ago
    I don't know much about these kinds of neural nets but this is the thing I've wondered about - with billions of parameters and enormous training sets, it seems more than likely that it's squirreled away fair chunks of text inside itself that it can regurgitate. Edit: I pasted the random sample above into Grammarly's plagiarism detector and it says it "detected plagiarism" but I'm not paying to sign up just to find out what it said (also they make you sign up with your email address before mentioning that it costs money... rude!) Maybe someone with a Grammarly subscription can try it? Edit 2: Getting off-topic but wow, Grammarly is a nest of dark patterns. Doesn't tell you you need to create an account until you've pasted your text in. Doesn't tell you you need to pay a subscription until you've created an account. Sends you a follow-up email a few minutes later if you don't subscribe. Puts the account deletion link in a different colour and location to the rest of the account settings so it looks like a footer. Swaps the styling for the 'delete account' and 'keep account' links when you do find the delete button to try and steer you away from deleting your account. After that lot I'm never giving them a cent.
andreyk 6y ago
"What should we think about the experts? Projections of failure were made by eminent, respectable, serious people. They spoke in considered tones of why AI hype was excessive and might trigger an “AI winter”, and the fundamental flaws of fashionable approaches and why brute force could not work. These statements were made routinely in 2014, 2015, 2016… And they were wrong. I am aware of few issuing a mea culpa or reflecting on it. It is a puzzling failure, and I’ve reflected on it before." Who are these experts? Where are records of these routine statements? Seriously, I am an AI researcher, who said this?
- p1esk 6y ago
  Gary Marcus?
  - andreyk 6y ago
    I mean, he and one or two other AI-research adjacent people basically have a hobby of bringing up limitations of Deep Learning etc, sure. But if that's all that is meant by this post, this is a weak point indeed...
    - p1esk 6y ago
      After IBM Watson won Jeopardy! game Noam Chomsky said: "Watson understands nothing. It's a bigger steamroller". I wonder if he still holds this view after reading GPT-3 samples.
- nl 6y ago
  Here's a good example from 2018: "Deep learning (does not) scale... We can't just scale up AlexNet and get respectively better results - we have to fiddle with specific architectures, and effectively additional compute does not buy much without order of magnitude more data samples, which are in practice only available in simulated game environments." https://blog.piekniewski.info/2018/05/28/ai-winter-is-well-on-its-way/ https://blog.piekniewski.info/2018/05/28/ai-winter-is-well-o... But Gary Marcus is the gold standard. Check out his 2018 take-down attempt, which is clearly mostly wrong: https://arxiv.org/ftp/arxiv/papers/1801/1801.00631.pdf https://arxiv.org/ftp/arxiv/papers/1801/1801.00631.pdf
cannabis_sam 6y ago
“ but an idiot savant, we should remember, is only a genetic mutation or bit of brain damage away from a normal human.” I’m not sure this comparison holds.. seems like a chain of fuzzy implications taken as necessary fact.
roca 6y ago
I don't know much about ML but I wonder what the need for so much training data (apparently about 500 billion words for GPT-3) means for this approach. Humans achieve their performance levels while only ever observing a tiny fraction of that training data --- at least in word form. I see only two possibilities: -- Human brains somehow are able to learn from all forms of sensory input and world interaction in ways that pay off in word tests. (Even then, would the information processed by an average human match the GPT-3 training corpus?) -- The human brain has a very different architecture that lets us learn vastly more efficiently from small amounts of training data. Is there another possibility I'm missing? If the latter is true then that gives humans a durable advantage over GPT-3-like approaches.
- TomMarius 6y ago
  Whether humans have a genetic language facility is not yet known.
  - sooheon 6y ago
    It is also kind of a moot point. Whether or not there is a genetic component, it is relevant only as far it is cashed out in physical brain matter and its arrangement.
- the8472 6y ago
  Probably a combination of factors not a single one. In addition to the ones you mentioned: GPT is also much smaller than a brain (fewer parameters), so it can't form higher-level concepts that might help it to organize things better. And the human brain has priors about the world baked in through evolution something that GPT sort of avoids by working on arbitrary token sequences.
  - roca 6y ago
    I considered the "pretrained via genetics" possibility but I don't see how that could make much difference. It doesn't look like the genome has enough information to encode really large amounts of brain structure details.
losvedir 6y ago
When I read the output of this model I'm really quite impressed. However, given the sheer size of it and huge training corpus, to what extent is it just regurgitating source text?
- justanotherhn 6y ago
  I guess we could ask the same question about ourselves. How much of what we say is just a regurgitation of what we hear/read every day?
- the8472 6y ago
  This is a common question and the authors answer this in their paper by checking the training set when evaluating it on particular tasks. E.g. for the addition task they search for existing samples of addition in the training data and find that not all are represented.
nonesuchluck 6y ago
There's a lot of mixed metaphors here, but just to attack one I immediately recognize: the 1940 Scientific American article "Don't worry, it can't happen" is not (as Gwern appears to be implying?) stating that nuclear explosions are impossible. Instead, the article explains why nuclear chain reactions eventually stop, and do not continue to explode all available matter. S.A. is not saying Trinity Test would not explode, merely that it would not turn Earth into a new sun, and you can get some sleep about it.
- nkurz 6y ago
  Gwern's typographical conventions are idiosyncratic, and you might not have noticed that he linked the original Scientific American article in his text. Reading the article now, I don't see how your interpretation is defensible. It doesn't mention exploding the earth, and while there is a little ambiguity, as Gwern does imply, they are describing a recent paper that concludes that large fission reactions are simply impossible: "They found instead, that instead of building up to a grand climax, it runs down and stops like an unwound clock." The final line of the caption is 'Readers made insomnious by "newspaper talk" of terrific atomic weapons held in reserve by dictators may now get sleep'. At least superficially, that sure sounds to be more about atomic weapons being impossible than about whether the chain reaction would consume the entire earth. I think you are confusing the actual linked article with Edward Teller's later argument that a nuclear fission explosion might ignite the atmosphere: https://www.realclearscience.com/blog/2019/09/12/the_fear_that_a_nuclear_bomb_could_ignite_the_atmosphere.html https://www.realclearscience.com/blog/2019/09/12/the_fear_th....
deskamess 6y ago
Love the document format and how hovering on links presents enough detail inline (example Bitter Lessons, Sutton). I appreciate not being taken to another page/tab. Wonder if a CSS style sheet is easily available. Had never heard of MuZero before. Its impressive that it can reach AlphaZero levels in Go without knowing the rules.
- davnn 6y ago
  Just look at the page source for CSS and JS used to create it.
yters 6y ago
Reminds me of explorers where the alien talks in soundbites from earth TV. All hail our future plagiarist AI overlords!
abj 6y ago
What human like abilities would a scaled up version of GPT-3 have? "Would it be worthwhile, even if it represented another large leap in AI capabilities, to spend up to 10 milli-Manhattan-Projects to scale GPT-3 100x to achieve human-like performance in some domains? Many researchers feel that such a suggestion is absurd and refutes the entire idea of scaling machine learning research further, and that the field would be more productive if it instead focused on research which can be conducted by an impoverished goatherder on an old laptop running off solar panels. Nevertheless, I think we can expect further scaling." [1] [1] https://www.gwern.net/newsletter/2020/05#gpt-3 https://www.gwern.net/newsletter/2020/05#gpt-3
- nl 6y ago
  Solving Winograd schemas would be a pretty interesting and significant step forward.
nl 6y ago
Given that Image-GPT[1] which was based on the GPT-2 architecture has been shown to do very good image completion, I think in the next 12 months we'll see a unified Image/Language GPT. I think it will be able to (imperfectly) do things like the following: - OCR from images - Textual descriptions of images It may start to make some progress towards things like: - Generating images from a textual description - producing structured documents (eg HTML) from document images It'd be interesting to see how far along they already are with this. [1] https://openai.com/blog/image-gpt/ https://openai.com/blog/image-gpt/
jefftk 6y ago
"GPT-3 is an extraordinarily expensive model by the standards of machine learning: it is estimated that training it may require the annual cost of more machine learning researchers than you can count on one hand (~$5m), up to $30 of hard drive space to store the model (500–800GB), and multiple pennies of electricity per 100 pages of output (0.4 kWH). Researchers are concerned about the prospects for scaling: can ML afford to run projects which cost more than 0.1 milli-Manhattan-Projects? Would it be worthwhile, even if it represented another large leap in AI capabilities, to spend up to 10 milli-Manhattan-Projects to scale GPT-3 100x to achieve human-like performance in some domains? Many researchers feel that such a suggestion is absurd and refutes the entire idea of scaling machine learning research further, and that the field would be more productive if it instead focused on research which can be conducted by an impoverished goatherder on an old laptop running off solar panels. Nevertheless, I think we can expect further scaling."
minimaxir 6y ago
I released a large number of GPT-3 demos yesterday: https://github.com/minimaxir/gpt-3-experiments https://github.com/minimaxir/gpt-3-experiments
Der_Einzige 6y ago
All this advancement in text generation - but it you want to use it to represent sentences or documents in a semantic space you still have to use horrifically bad techniques like average or max pooling. When is someone going to advance pooling techniques? We desperately need improvement!
mark_l_watson 6y ago
I retired a year ago from managing a deep learning team. While I am a fan of the technology, I really yearn for more research aimed at hybrid AI systems. Even given one-shot (few shot?) learning, transfer learning, etc., I keep coming back to watching my grandchildren back when they were infants. They could see a picture of a new animal in a picture book, and really "get" why the animal looked different form others, etc. Deep learning will not get us there. I think we get to general AI by using prior knowledge, exploit deep learning, and also build systems that can develop their own models of the world that they can maintain, modify, discard, and combine. As always, a great write up by gwern!
gwern 6y ago
Incidentally, I've also been working on a selection of GPT-3 generated creative writing (primarily, but far from limited to, poetry): https://www.gwern.net/GPT-3 https://www.gwern.net/GPT-3
akeck 6y ago
Did Gwern write parts of this with GPT-3? It has a certain... flavor.

291 points by cocoflunchy 6y ago

Something seems to be wrong with the footnotes on this page... Edit: Nevermind, my browser seems to just be screwing up on Gwern's footnotes in general at the moment.

renewiltord 6y ago
On my desktop, I hover over and they pop up a little window with the footnote in it. That looks like the intended behaviour but it must be failing on your platform.

throwaway4666 6y ago

>Like the genomics revolution where a few far-sighted seers extrapolated that the necessary n for GWASes would increase exponentially & deliver powerful PGSes soon, while sober experts wrung their hands over “missing heritability” & the miraculous complexity of biology & scoff about how such n requirements proved GWAS was a failed paradigm, the future arrived at first slowly and then quickly. Yet, here we are: all honor to the fanatics, and shame and humiliation to the critics! No. Actual geneticists are still pretty skeptical about GWASes because they tell us almost nothing about, well, the genetics behind complex traits. It's all good and well running GWASes for literally anything (see: twitter.com/sbotgwa or that dude who got a pretty good PGS from correlating a country's GDP with the genotypes of Arabidopsis thaliana available in that country) but that's virtually useless for serious research or if you want to know how genes work. Actually figuring out to what extent a trait is genetically determined usually involves much more complex methods (e.g. mendelian randomization) and knock-out experiments on animal models, which is all terribly expensive and tedious. But that's how actual genetics works, not waving a magic wand of +15% heritability.

jhrmnn 6y ago
Agreed. To put it simply, if you model a complex system with another complex system, of which you have a similar level of understanding, then even if the model is very faithful, you gain very little. Unless the model is transferable, so you can predict, which AFAIK is not the case here.
dash2 6y ago
I think the point is not that GWAS can do everything, but that many people 5-10 years ago thought that they had a fundamental flaw - where was all the heritability? It has turned out so far that the heritability is there and discoverable with big enough sample sizes. But of course, you are right that GWAS don't tell us much about the very large space between strands of DNA and social or behavioural outcomes. They are a tool like others. (Mendelian randomization is hardly a panacea - how credible are those exclusion restrictions, usually?)
esyir 6y ago
Unless something has drastically changed in the past 5 odd years, Actual Geneticists (tm) happily use GWAS in many investigations of complex traits. They're often an early step in the overall pipeline in which they trawl for potential candidates for that target of interest, after which they go on to said more complex methods. Why? You've already said so yourself. Those are expensive and tedious, and searching across the entire pool of human genetics with them is an exercise in futility.
- throwaway4666 6y ago
  The important part in your post is >They're often an early step I think we're in agreement here, I'm just arguing that "woah, look at all those correlations" isn't a breakthrough or 'genomic revolution' in any sense of the word as far as our understanding of human genetics is concerned.
- hyperbovine 6y ago
  Seemingly, most of the complex traits for which GWA is an effective methodology (a few SNPs of large effect size) have already been discovered. More and more these days, I’m seeing association studies that fail to yield any hits. Whether this is due to sample size, polygenicity, or some other model failure remains to be seen.
folli 6y ago
Can you post the link about the GDP/A. thaliana correlation? That sounds hilarious, and Google is failing me to find it myself.
- throwaway4666 6y ago
  Found it: https://mobile.twitter.com/pastramimachine/status/1263937115684892674 https://mobile.twitter.com/pastramimachine/status/1263937115... Edit: if it doesn't work try this i guess? https://www.google.com/url?q=https://mobile.twitter.com/pastramimachine/status/1263937115684892674&sa=U&ved=2ahUKEwjyz5eCo5rqAhUnxIUKHTl-BBwQFjAAegQIARAB&usg=AOvVaw1EqJAtVmJyjAEa3TiP3EFe https://www.google.com/url?q=https://mobile.twitter.com/past...
  - ralfd 6y ago
    Throws a Twitter error for me.
bobcostas55 6y ago
Who cares how genes work? The important part is the real predictive power of the PGS.
- throwaway4666 6y ago
  'Predictive' is just a glorified word for 'correlating'. Staring at a bunch of correlations only gets you so far and doing only that certainly isn't science. If you have no idea how genes work you're going to be dismayed when all your fancy correlations don't work anymore, as people gathering genomic data from outside the European biobanks are starting to find out.
Gatsky 6y ago
I agree, holding up GWAS as some kind of lesson about scientific progress seems pretty silly. I don't think the 'sober experts' deserve shame and humiliation considering that PGSes have had minimal impact on human health.
zxcb1 6y ago
GWAS have been shown to be practically meaningless. How should this paragraph be interpreted, that transformers can make sense of DNA patterns?
gwern 6y ago
Nope. You're wrong. Where do you think your MRs or your sibling comparisons are coming from? You're not getting anywhere with weak instruments from a few significant hits. You're also wrong that the only thing of value is inferring "how genes work", and that is the sort of extremely blinkered mechanism-centric attitude which blinded people to GWASes working, because gosh, it would be awful if polygenicity was true, because how would we build any 'scientific theories' on this? (Cue Turkheimer.) And yes, those PGSes are useful for all sorts of things like selective sweeps, enrichments, and clinical instruments, because of incremental validity. (By the way, what does '15% heritability' refer to? I sure hope that, since you're claiming to be an expert here, you aren't confusing heritability with PGS power, like so many supposed human geneticists insist on doing...)
- throwaway4666 6y ago
  I didn't say GWASes were useless, just that it's absurd to consider them to be a 'revolution'. The actual revolution would be second- and third-generation sequencing which enabled GWASes and a bunch of much more useful things. GWAS is, in effect, just a bunch of correlations. It's just the very starting point to an actual scientific analysis, because 'you have to start somewhere'. If you don't go beyond and investigate, you're not doing science. Everyone in the genomics community agrees to this, and literally every paper that investigates the causes of genetic diseases goes in the introduction like 'GWASes sure look nice but we still have no idea how things work with them so in this paper I present a method to do...' I notice you failed to address many of the spurious correlations drawn by the GWAS bot or the A. thaliana vs. GDP prediction. That it doesn't raise any red flag to you doesn't speak well as to your ability to approach the field of genomics. >You're also wrong that the only thing of value is inferring "how genes work" Yes it is, that's literally what genetics is about. Otherwise you're back to making a bunch of correlations. If you want a deep understanding of disease, design effective drugs, or even do proper gene editing, you have to understand what genes do. It seems ridiculous to have to say it. >By the way, what does '15% heritability' refer to? It is what we in the less-rationalist circles refer to as a 'joke'.

fock 6y ago

Well, I take the the opposite stand: GPT-3 = Give a 3rd grader wikipedia and some paper. While it's certainly fun how it's relatively coherent in writing text, I couldn't see yet, how it links syntactically correct text to actual facts. Which in my opinion is the difference between 1e100 apes with typewriters and aforementioned 3rd grader.

throwaway4666 6y ago
I think gwern's point (whose opinion on deep learning I respect a lot more than his on genetics) is that GPT-3 is nearly not the end of the story. OpenAI released a GPT pretty much every year and each one is a spectacular improvement on the previous one with no sign of that trend ever stopping. If size is really what matters at all, there's no telling what GPT-4 or GPT-5 might be capable of, let alone a giant GPT run by state-sized actors.
- ipsum2 6y ago
  The machine that trained GPT-3 is one of the largest GPU clusters in the world at the moment. (10k v100 GPUs hooked up with fast interconnect). For comparison, Nvidia's "Selene" supercomputer only has 2k a100 GPUs and is #7 fastest supercomputer in the world: https://www.hpcwire.com/2020/06/22/nvidia-nabs-7-spot-on-top500-with-selene-launches-a100-pcie-cards/ https://www.hpcwire.com/2020/06/22/nvidia-nabs-7-spot-on-top... a100 GPUs are around 3x faster than v100 GPUs. It would take some time for governments to catch up :)
  - qeternity 6y ago
    1. This is a list of publicly known supercomputers (it’s likely that the intelligence community have larger secret assets). 2. Most of the largest super computers are owned by states.
    - throwawaygh 6y ago
      Lots of large companies don't include their clusters on those supercomputer lists. Including Google. IDK about DoD, but I wouldn't be surprised if they have top 10 clusters we don't know about. The list of largest supercomputers is definitely only a list of the largest disclosed supercomputers. That said, only so many actors could've secretly sourced 10K+ v100s with fast interconnect...

skrause 6y ago

The GUID Partition Table is already at its third version?

anticensor 6y ago
No, Generative Pre-Trained Transformer.
- skrause 6y ago
  I know, it's just a stupid acronym because it's already used for something else (the partition table).
  - Smaug123 6y ago
    Oh boy, do I have bad news for you: Wikipedia lists ten overloaded acronyms starting with "AA" alone!
    - skrause 6y ago
      But they are usually in unrelated fields.

Tepix 6y ago

I'm happy to see it being released. What is the minimum hardware required to run this locally? Or what's the cheapest way to run this model (even at barely acceptable performance) at a cloud provider?

bjourne 6y ago
Probably out of reach, but you can run GTP-2 on Google Colab: https://colab.research.google.com/github/ilopezfr/gpt-2/blob/master/gpt-2-playground_.ipynb https://colab.research.google.com/github/ilopezfr/gpt-2/blob...
p1esk 6y ago
Any server with 512GB of RAM. So basically a couple of bucks per hour, to generate a few words per minute.
- Tepix 6y ago
  You don't need fancy GPUs with lots of RAM?
  - p1esk 6y ago
    If you’re ok with generating a few words per minute, no.

dgrabla 6y ago

What book should I read to be able to understand this text? No papers please. I don't understand what scaling a model means.

mkl 6y ago
This is about new research, which mostly lives in papers and articles (like this) about the papers. It won't show up in introductory books for a while, so if you're unwilling to read or even look at papers, you won't be able to understand details of new research. Scaling a model is just like it sounds: more data fed into a bigger network with more parameters. The gist of what this article is saying about scaling is that there's no sign of diminishing returns yet in terms of what the network can do and how well it generalises as the number of parameters is increased: the "more parameters = better performance" trend continues up to the enormous size of the full GPT-3 model, with no indication that even bigger models won't have even better performance. Here is the GPT-3 paper: https://arxiv.org/pdf/2005.14165.pdf https://arxiv.org/pdf/2005.14165.pdf If you really want to understand, skim this, and focus especially on the graphs, as they show the scaling. The x axis is usually model size, and the y axis is mostly accuracy or "loss" (~error).

chillee 6y ago

Personally, as a ML researcher, I find GPT-3 very unsatisfying. There aren't any novel architectural details, it doesn't improve our "fundamental" understanding of the field, and it requires the type of computation I have no chance of getting. As a fan of the field, however, it is undoubtedly very cool. I still think there's tasks that GPT-3 has no chance of tackling (say, generating code to solve a novel programming task), but the bitter lesson is a bitter one indeed...

FeepingCreature 6y ago
It's kind of LHC-like, isn't it? Same physics "but larger".. "will it work if we scale it up? Apparently: yes." I think the big point, and what scares me a bit, is that we have yet to discover any sort of fundamental conceptual limit to the Transformer architecture.
- sooheon 6y ago
  Each attention block in the Transformer models a fully connected graph (with the attention heads being learned edge attributes). A graph is the most general data structure possible, so yeah, I don't think there's really a fundamental limitation to them, just a computational one. Latest papers from ICLR explore how they fully model CNNs and RNNs, for example, and I'm sure papers on their theoretical equivalence with GNNs are coming.
  - contravariant 6y ago
    > I'm sure papers on their theoretical equivalence with GNNs are coming. Not exactly a paper, but you mean something like this article: https://graphdeeplearning.github.io/post/transformers-are-gnns/ https://graphdeeplearning.github.io/post/transformers-are-gn... ?
    - sooheon 6y ago
      Pretty much that, yeah. Thanks for the link. That kind of article sets the intuition, and papers like these give it a better theoretical treatment: - https://openreview.net/pdf?id=HJlnC1rKPB https://openreview.net/pdf?id=HJlnC1rKPB - https://openreview.net/pdf?id=ByxRM0Ntvr https://openreview.net/pdf?id=ByxRM0Ntvr

nl 6y ago

Winograd schemas falling at 10T parameters is interesting. That's probably only 5 years off. If we can build something capable of passing Winograd schema, then it can probably write working non-trivial computer programs from plain text. Google's PEGASUS summarization model[1] has learned to count up to five (which is amazing!!). That's "only" 568M parameters. It's be interesting to see GPT-3 fine tuned against the PEGASUS objective function. [1] https://ai.googleblog.com/2020/06/pegasus-state-of-art-model-for.html https://ai.googleblog.com/2020/06/pegasus-state-of-art-model...

LordDragonfang 6y ago
For those curious the "counting" is at the end of the article, and really is quite impressive: >Following this post is an example article from the XSum dataset along with the model-generated abstractive summary. The model correctly abstracts and paraphrases four named frigates (HMS Cumberland, HMS Campbeltown, HMS Chatham and HMS Cornwall) as “four Royal Navy frigates”, something an extractive approach could not do since “four” is not mentioned anywhere. Was this a fluke or did the model actually count? One way to find out is to add and remove ships to see if the count changes. >As can be seen below, the model successfully “counts” ships from 2 to 5. However, when we add a sixth ship, the “HMS Alphabet”, it miscounts it as “seven”. So it appears the model has learned to count small numbers of items in a list, but does not yet generalize as elegantly as we would hope. Still, we think this rudimentary counting ability is impressive as it was not explicitly programmed into the model, and it demonstrates a limited amount of “symbolic reasoning” by the model.
- nl 6y ago
  It's one of the most amazing and surprising things I've seen in the last 12 months in machine learning (I follow and work in the field). It surprises me a lot more than the excellent performance of GPT-3 on text generation for example. GPT-3 is amazing but looking at GPT-1 -> GPT-2 -> GPT-3 it isn't surprising. Counting on the other hand is something I wouldn't have expected from a summarizer.
  - FeepingCreature 6y ago

personjerry 6y ago

I'm failing to connect the anecdotes with the conclusion. He's claiming the the ML is scaling well but then gives the data on how GPT-3 is expensively brute-forcing its way to "success". To me it just seems like what supercomputing is to normal computing: It makes the computationally expensive stuff do-able in a reasonable amount of time, or gives diminishing returns on existing algorithms. But it doesn't magic in any real advancements. The problem in AI/ML and the concept of "AI winter" to me was always the barrier of the fact that we're just doing predictions, with no deep meaning or comprehension of features. The layman thinks there's magic, but there's not, so when that truth is revealed there will be problems. There's nothing intelligent about our Artificial Intelligence; we're just doing statistics with big data and added sprinkles. OpenAI just proved they could do statistics with even bigger data and more expensive sprinkles. Has their work really shown we can get past that core problem? Personally, I don't see it.

LordDragonfang 6y ago
>we're just doing statistics with big data and added sprinkles. I mean, you could pretty much say that's how the human brain works, couldn't you?
- personjerry 6y ago
  That comparison is like comparing gardening my backyard to agricultural farming. Technically the core concept is similar, but I think we both know that the scale and depth makes them not at all the same.
  - H8crilA 6y ago
    There's a continuity of small changes between your backyard and industrial farming. Just like there's a continuity of small changes between you and your cat.
    - p1esk 6y ago
      There’s a continuity of small changes between any two things.

SonOfLilit 6y ago

Just so we're all on the same page, this is a random GPT-3 sample (I clicked "Random" twice, the first one was a short Wikipedia-like article, this was the second): https://read-the-samples.netlify.app/sample_1749 https://read-the-samples.netlify.app/sample_1749 As far as I'm concerned, I would probably pick this over a random fanfic in a Turing Test.

mkl 6y ago
That's amazing. The only problems I caught are that there's a spurious double quote right at the start, and Murphy seems to have two children, a son and a daughter, but then later refers to "the boys". To really evaluate these samples, we need a way to search for phrases in the training data, to see how much is just learned and copied. I've tried Google and DDG, but not found anything.
- hyperman1 6y ago
  I also wonder how much is invented, and how much is simply copied. I saw this one by GPT3 https://read-the-samples.netlify.app/sample_2007 https://read-the-samples.netlify.app/sample_2007 and compare it to e.g. https://www.bbc.com/news/world-europe-35869985 https://www.bbc.com/news/world-europe-35869985 GPT-3 has quite a lot of correct facts, manages to mix fouad belkacem and Sharia4Belgium into it, etc... I suspect this is what happens when there is not enough training data: You get a lookalike, not something new. To be clear: It's impressive what GPT-3 is capable of, and it's very probable something better is comming in the near future. (It's also quite amusing how a lot of these terrorists get first killed by the police and then thrown into jail for 20 years. That'll teach their corpses! )
  - gomox 6y ago
    > I also wonder how much is invented, and how much is simply copied Well, the average human probably does not excel at this metric either
- taneq 6y ago
  I don't know much about these kinds of neural nets but this is the thing I've wondered about - with billions of parameters and enormous training sets, it seems more than likely that it's squirreled away fair chunks of text inside itself that it can regurgitate. Edit: I pasted the random sample above into Grammarly's plagiarism detector and it says it "detected plagiarism" but I'm not paying to sign up just to find out what it said (also they make you sign up with your email address before mentioning that it costs money... rude!) Maybe someone with a Grammarly subscription can try it? Edit 2: Getting off-topic but wow, Grammarly is a nest of dark patterns. Doesn't tell you you need to create an account until you've pasted your text in. Doesn't tell you you need to pay a subscription until you've created an account. Sends you a follow-up email a few minutes later if you don't subscribe. Puts the account deletion link in a different colour and location to the rest of the account settings so it looks like a footer. Swaps the styling for the 'delete account' and 'keep account' links when you do find the delete button to try and steer you away from deleting your account. After that lot I'm never giving them a cent.

andreyk 6y ago

"What should we think about the experts? Projections of failure were made by eminent, respectable, serious people. They spoke in considered tones of why AI hype was excessive and might trigger an “AI winter”, and the fundamental flaws of fashionable approaches and why brute force could not work. These statements were made routinely in 2014, 2015, 2016… And they were wrong. I am aware of few issuing a mea culpa or reflecting on it. It is a puzzling failure, and I’ve reflected on it before." Who are these experts? Where are records of these routine statements? Seriously, I am an AI researcher, who said this?

p1esk 6y ago
Gary Marcus?
- andreyk 6y ago
  I mean, he and one or two other AI-research adjacent people basically have a hobby of bringing up limitations of Deep Learning etc, sure. But if that's all that is meant by this post, this is a weak point indeed...
  - p1esk 6y ago
    After IBM Watson won Jeopardy! game Noam Chomsky said: "Watson understands nothing. It's a bigger steamroller". I wonder if he still holds this view after reading GPT-3 samples.
nl 6y ago
Here's a good example from 2018: "Deep learning (does not) scale... We can't just scale up AlexNet and get respectively better results - we have to fiddle with specific architectures, and effectively additional compute does not buy much without order of magnitude more data samples, which are in practice only available in simulated game environments." https://blog.piekniewski.info/2018/05/28/ai-winter-is-well-on-its-way/ https://blog.piekniewski.info/2018/05/28/ai-winter-is-well-o... But Gary Marcus is the gold standard. Check out his 2018 take-down attempt, which is clearly mostly wrong: https://arxiv.org/ftp/arxiv/papers/1801/1801.00631.pdf https://arxiv.org/ftp/arxiv/papers/1801/1801.00631.pdf

cannabis_sam 6y ago

“ but an idiot savant, we should remember, is only a genetic mutation or bit of brain damage away from a normal human.” I’m not sure this comparison holds.. seems like a chain of fuzzy implications taken as necessary fact.

roca 6y ago

I don't know much about ML but I wonder what the need for so much training data (apparently about 500 billion words for GPT-3) means for this approach. Humans achieve their performance levels while only ever observing a tiny fraction of that training data --- at least in word form. I see only two possibilities: -- Human brains somehow are able to learn from all forms of sensory input and world interaction in ways that pay off in word tests. (Even then, would the information processed by an average human match the GPT-3 training corpus?) -- The human brain has a very different architecture that lets us learn vastly more efficiently from small amounts of training data. Is there another possibility I'm missing? If the latter is true then that gives humans a durable advantage over GPT-3-like approaches.

TomMarius 6y ago
Whether humans have a genetic language facility is not yet known.
- sooheon 6y ago
  It is also kind of a moot point. Whether or not there is a genetic component, it is relevant only as far it is cashed out in physical brain matter and its arrangement.
the8472 6y ago
Probably a combination of factors not a single one. In addition to the ones you mentioned: GPT is also much smaller than a brain (fewer parameters), so it can't form higher-level concepts that might help it to organize things better. And the human brain has priors about the world baked in through evolution something that GPT sort of avoids by working on arbitrary token sequences.
- roca 6y ago
  I considered the "pretrained via genetics" possibility but I don't see how that could make much difference. It doesn't look like the genome has enough information to encode really large amounts of brain structure details.

losvedir 6y ago

When I read the output of this model I'm really quite impressed. However, given the sheer size of it and huge training corpus, to what extent is it just regurgitating source text?

justanotherhn 6y ago
I guess we could ask the same question about ourselves. How much of what we say is just a regurgitation of what we hear/read every day?
the8472 6y ago
This is a common question and the authors answer this in their paper by checking the training set when evaluating it on particular tasks. E.g. for the addition task they search for existing samples of addition in the training data and find that not all are represented.

nonesuchluck 6y ago

There's a lot of mixed metaphors here, but just to attack one I immediately recognize: the 1940 Scientific American article "Don't worry, it can't happen" is not (as Gwern appears to be implying?) stating that nuclear explosions are impossible. Instead, the article explains why nuclear chain reactions eventually stop, and do not continue to explode all available matter. S.A. is not saying Trinity Test would not explode, merely that it would not turn Earth into a new sun, and you can get some sleep about it.

nkurz 6y ago
Gwern's typographical conventions are idiosyncratic, and you might not have noticed that he linked the original Scientific American article in his text. Reading the article now, I don't see how your interpretation is defensible. It doesn't mention exploding the earth, and while there is a little ambiguity, as Gwern does imply, they are describing a recent paper that concludes that large fission reactions are simply impossible: "They found instead, that instead of building up to a grand climax, it runs down and stops like an unwound clock." The final line of the caption is 'Readers made insomnious by "newspaper talk" of terrific atomic weapons held in reserve by dictators may now get sleep'. At least superficially, that sure sounds to be more about atomic weapons being impossible than about whether the chain reaction would consume the entire earth. I think you are confusing the actual linked article with Edward Teller's later argument that a nuclear fission explosion might ignite the atmosphere: https://www.realclearscience.com/blog/2019/09/12/the_fear_that_a_nuclear_bomb_could_ignite_the_atmosphere.html https://www.realclearscience.com/blog/2019/09/12/the_fear_th....

deskamess 6y ago

Love the document format and how hovering on links presents enough detail inline (example Bitter Lessons, Sutton). I appreciate not being taken to another page/tab. Wonder if a CSS style sheet is easily available. Had never heard of MuZero before. Its impressive that it can reach AlphaZero levels in Go without knowing the rules.

davnn 6y ago
Just look at the page source for CSS and JS used to create it.

yters 6y ago

Reminds me of explorers where the alien talks in soundbites from earth TV. All hail our future plagiarist AI overlords!

abj 6y ago

What human like abilities would a scaled up version of GPT-3 have? "Would it be worthwhile, even if it represented another large leap in AI capabilities, to spend up to 10 milli-Manhattan-Projects to scale GPT-3 100x to achieve human-like performance in some domains? Many researchers feel that such a suggestion is absurd and refutes the entire idea of scaling machine learning research further, and that the field would be more productive if it instead focused on research which can be conducted by an impoverished goatherder on an old laptop running off solar panels. Nevertheless, I think we can expect further scaling." [1] [1] https://www.gwern.net/newsletter/2020/05#gpt-3 https://www.gwern.net/newsletter/2020/05#gpt-3

nl 6y ago
Solving Winograd schemas would be a pretty interesting and significant step forward.

nl 6y ago

Given that Image-GPT[1] which was based on the GPT-2 architecture has been shown to do very good image completion, I think in the next 12 months we'll see a unified Image/Language GPT. I think it will be able to (imperfectly) do things like the following: - OCR from images - Textual descriptions of images It may start to make some progress towards things like: - Generating images from a textual description - producing structured documents (eg HTML) from document images It'd be interesting to see how far along they already are with this. [1] https://openai.com/blog/image-gpt/ https://openai.com/blog/image-gpt/

jefftk 6y ago

"GPT-3 is an extraordinarily expensive model by the standards of machine learning: it is estimated that training it may require the annual cost of more machine learning researchers than you can count on one hand (~$5m), up to $30 of hard drive space to store the model (500–800GB), and multiple pennies of electricity per 100 pages of output (0.4 kWH). Researchers are concerned about the prospects for scaling: can ML afford to run projects which cost more than 0.1 milli-Manhattan-Projects? Would it be worthwhile, even if it represented another large leap in AI capabilities, to spend up to 10 milli-Manhattan-Projects to scale GPT-3 100x to achieve human-like performance in some domains? Many researchers feel that such a suggestion is absurd and refutes the entire idea of scaling machine learning research further, and that the field would be more productive if it instead focused on research which can be conducted by an impoverished goatherder on an old laptop running off solar panels. Nevertheless, I think we can expect further scaling."

minimaxir 6y ago

I released a large number of GPT-3 demos yesterday: https://github.com/minimaxir/gpt-3-experiments https://github.com/minimaxir/gpt-3-experiments

Der_Einzige 6y ago

All this advancement in text generation - but it you want to use it to represent sentences or documents in a semantic space you still have to use horrifically bad techniques like average or max pooling. When is someone going to advance pooling techniques? We desperately need improvement!

mark_l_watson 6y ago

I retired a year ago from managing a deep learning team. While I am a fan of the technology, I really yearn for more research aimed at hybrid AI systems. Even given one-shot (few shot?) learning, transfer learning, etc., I keep coming back to watching my grandchildren back when they were infants. They could see a picture of a new animal in a picture book, and really "get" why the animal looked different form others, etc. Deep learning will not get us there. I think we get to general AI by using prior knowledge, exploit deep learning, and also build systems that can develop their own models of the world that they can maintain, modify, discard, and combine. As always, a great write up by gwern!

gwern 6y ago

Incidentally, I've also been working on a selection of GPT-3 generated creative writing (primarily, but far from limited to, poetry): https://www.gwern.net/GPT-3 https://www.gwern.net/GPT-3

akeck 6y ago

Did Gwern write parts of this with GPT-3? It has a certain... flavor.