Large language models are having their Stable Diffusion moment
simonwillison.net
simonwillison.net
The big difference between it and Stable Diffusion which caused the latter to go megaviral is a) it can run on a typical GPU that gamers likely already have without hitting a perf ceiling and b) it can run easily in a free Colab GPU. Although Hugging Face transformers can run a 7B model on a T4 GPU w/ 8-bit loading, but with its own caveats too.
There's a big difference between "can run" and "can run well". VQGAN + CLIP had a lot of friction too and that's partially why AI image generation didn't go megaviral then.
Unless it's relicensed I don't expect LLaMA to be a long-term foundation model. But it's shown that yes, you can run a GPT-3 class model on an M1 Mac with 8GB of RAM (or maybe 16GB for the 13B one?)
I fully expect other models to follow, from other organizations, with better capabilities and more friendly licensing terms.
We're all still waiting to hear about (non-shovel-selling) successes in this space.
1 - some guy won hundreds of dollars in an art contest from AI generated art (and this made big news, so it should be easy to find)
2 - one person reported using midjourney's images as a starting point for images that wound up being used in a physical magazine
3 - another artist has used midjourney images that they modify to sell in all sorts of contexts (like background images on stock illustration sites)
You'd probably find many other examples in midjourney's #in-the-world discord channel.
I'd also be shocked if stock image sites, clipart sites and freelance design/illustration sites weren't already flooded with AI generated images that have been sold for money.
That being said, because high questly AI-generated images are so easy to make, the value of images of all types is likely to plummet soon if it hasn't already.
The price might, but the marginal value having them or not has to other products which incorporate them likely won’t as much. So, say, if you are a TTRPG publisher, the value of art to your business doesn't change, but the cost of moving farther along the no art -> scattered sketches & tangentially relevant public domain works -> ideal quantity of art, specific to the product, goes down. So, you are, in effect, making money on SD.
Making money with it in the non-shovel-selling sense is going to be commercial artists (or other people with visual image needs, perhaps avoiding the need for commercial artists) incorporating it in production workflows. But except when they are also selling shovels (e.g., Pixar with Renderman) there’s not a whole lot of incentive to trumpet workflow components rather than product in that space.
Maybe it's just me, but that reeks of the archetypal race to the bottom.
"We didn't have to pay professional X to do task Y as we and others have been doing for decades, instead we used modern replacement Z funded by cheap money which costs peanuts at present".
Q: Is there really a business to be made selling Z?
(Of course, yes, if there are businesses to be enhanced with Z, even if Z is free in the sense that open source software is free – which is exactly or approximately the case with much of the Stable Diffusion ecosystem – there will be businesses also to be made “selling” Z, or, more accurately, selling specialized services relating to Z like bespoke enhancements, support, hosting, and training. But those are very much the selling-shovels services that the question upthread asked to exclude; the using-shovels businesses are actually producing products incorporating images.)
How about:
1) Model weights are heavy for just experimentation, although quantizing it down to 4-bit might make them on par with SD FP16.
2) Requires extreme CLI shenanigans (and likely configuration since you have to run make) compared to just running a Colab Notebook or a .bat Windows Installer for the A1111 UI.
3) Again hardware: a M1 Pro or a RTX 4090 is not super common among people who are just curious about text generation.
4) It is possible the extreme quantization could be affecting text output quality; although the examples are coherent for simple queries, more complex GPT-3-esque queries might become relatively incoherent. Particularly with ChatGPT and its cheap API (timely!) out now such that even nontechies have a strong baseline on good output already. The viral moment for SD was that it was easy to use and it was a significant quality leap over VQGAN + CLIP.
I was going to say inference speed since that's usually another constraint for new LLMs but given the 61.41 ms/token cited for the 7B model in the repo/your GIF, that seems on par with the inference speed from OPT-6.7B FP16 in transformers on a T4.
Some of these caveats are fixable, but even then I don't think LLaMA will have its Stable Diffusion moment.
I'm not too worried about CLI shenanigans, because of what happened with whisper.cpp - it resulted in apps like https://goodsnooze.gumroad.com/l/macwhisper - wouldn't be at all surprised to see the same happen with llama.cpp
A regular M1 with 8GB of RAM appears to be good enough to run that 7B model. I wonder at what point it will run on an iPhone... the Stable Diffusion model was 4GB when they first released it, and that runs on iOS now after some more optimization tricks.
For me though, the "Stable Diffusion" moment isn't necessarily about the LLaMA model itself. It's not licensed for commercial use, so it won't see nearly the same level of things built on top of it.
The key moment for me is that I've now personally seen a GPT-3 scale model running on my own personal laptop. I know it can be done! Now I just need to wait for the inevitable openly-licensed, instruction-tuned model that runs on the same hardware.
It's that, but also the forthcoming explosion of developer innovation that a local model will unleash. llama.cpp is just the first hint of that.
I hate to pooh-pooh it for everyone, but this was possible before LLaMa. GPT-J-125m/6b have been around for a while, and are frankly easier to install and get results out of. The smaller pruned model even fits on an iPhone.
The problem is more that these smaller models won't ever compete with GPT-scale APIs. Tomorrow's local LLaMa might beat yesterday's ChatGPT, but I think those optimistic for the democratization of chatbot intelligence are setting their hopes a bit high. LLaMa really isn't breaking new ground.
There are all kinds of things I want to be able to do with a LLM that are a lot tighter than general chatbots.
I'd love to see a demo of GPT-J on an iPhone!
The best models will be by API but there is a threshold of utility for LLMs, below which they're just bad story telling toys, that hasn't been available for home users until recently, with FlanT5s and LLaMas.
I’m sure someone did the work you suggest to let folks run on this on more reasonable setups- but I never saw it on the front page here.
Did they achieve GPT-3 performance on anything at that model size? Happy to be corrected, but my understanding was llama does in fact outperform GPT-3 even at the smaller parameter-counts on certain downstream tasks due to using the chinchilla training regime (and other tricks).
Now that I've seen that LLaMA can work I'm confident someone will release an openly licensed instruction-tuned model that works on the same hardware at some point soon.
I also expect that there are prompt engineering tricks which can be used to get really great results out of LLaMA. I'm hoping someone will come up with a good prompt to get it to summarization, for example.
For a model running on my own laptop I'm OK taking more risks. I'd like it to be able to obey simple instructions like "Summarize this text" or "Extract the names of everyone mentioned in this article" - I don't care as much about the stuff ChatGPT has to get right.
Enough programmers want this badly enough that its going to happen. Inference at 8 GB and fine tuning at 24 GB, just like stable diffusion, on a 13B model.
I understand azure has its own chatgpt api.
IMO, Apple's habit of cheaping out on ram is maddening. Although perhaps this is a sign for me to pick up a m2 pro mini with extra ram.
The Base mac studio with 32 gigs of ram can be a good value from a ram perspective.
Let’s see if the Mac studio is discontinued or updated.
I do hope that it's updated. I could see them reserving the top end chip for the Mac Pro, though. Or maybe make it a dual socket?
In any case, I much prefer the ability to run an Apple computer with a proper cooling setup. I know that cooling pads for laptops exist, but ultimately it's a bit of a janky solution compared to actually adequately cooling the system properly.
The Mac studio is surprisingly quiet and well cooled. I haven’t had to install a fan control utility as of yet.
I thought I’d be trying one of these out and sell it if my setup didn’t work, but it’s the first desktop I’m starting to consider especially if I want to leave a workload running on it instead of a decade of carrying a lot of horsepower around with me.
Instant on for 3 monitors, no need for docking station or USB hubs, everything is plugged in and works, I can run a virtual cam obs setup easily.
The studio is a great upgrade on Mac mini, which might cannibalize the Mac pro sales and why it might not get toasted or refreshed. I’ll F the max studio was a way to put a dent in hackkntoshes, it makes sense.
The M2 mini looks super decent but quickly goes up in price when ram and ssd is setup to match a studio.
For my purposes the ultra wasn’t that much faster than the M1 max on because the software has to be optimized to benefit from it. Maybe some of this ML stuff will be shortly.
Apple's optimization here is nice for the people who own their hardware, but it's totally silly to read through the comments promising the end of CUDA.
I think there might be a different use case for me but to compare with your economics - I picked up a base mac studio (m1 max with 32 gb) for about $1370 USD with 8 months warranty remaining. It’s letting me test my daily setup to see if everything can run on Apple silicon yet without intervention.. if not, I’ll be able to get rid of it at little to no loss, and decide if I want to carry that much computing power in a laptop or not, or head on to other options.
My interest currently is computational power, per watt. Not much comes close to Apple Silicon.
The cost of a loaded pc can quickly outstrip other options with electricity costs included so it needs to be useful in any case.
The integrated speed of the Apple silicon, ram, and ssd is a little astonishing. More than I expected to admit. I don’t know if there’s anything like it on PC. If Apple silicon supported eGPUs it would interest me.
Comparable PCs have described are generally power inefficient. Still, the system you’re laying out is interesting, especially the ram, mind sharing a bill of materials?
I guess my PoV is with a GPU + inexpensive RAM, I can do some things more quickly, and other things are at least possible. But granted, hardware nuts and bolts is a hobby for me too, I've always enjoyed researching and building systems, there is a huge community of people and companies, and I much prefer KDE to MacOS, Android to IOS.
If Apple released a reasonably priced system with lots of RAM, or that could take a GPU, I'd not have as much to stand on, but they don't seem inclined to do that.
Here's BoM, $2170:
$700 for 192GB https://www.corsair.com/us/en/Categories/Products/Memory/VEN...
About $200 for a LGA1700 board with four slots (currently Asus and Gigabyte support 4 X 48GB memory modules on consumer hardware)
$200+ for a CPU
$120 for an 850 watt PSU
$700 for a used 16GB quadro, or a new 12GB A2000
$150 for 2TB NVME storage
$50 for some kinda case (I use an XTIA open frame, it is the opposite of a Macbook)
These are "best prices," You'd have to pay around 10% more if you're not into finding deals.
If it’s 400-800w 24 hours a day, op costs quickly pile up.
It looks really reasonable for everything here. I am interested more in eGPUs for interchangeability but have enough on my plate.
The GP's info is outdated. Apparently they don't realize how quickly LLaMA is being developed on. It's just like the early days of SD.
running 4-bit LLaMa 13B on my MacBook Air M2:
"possible resolutions to the Fermi paradox, the question of where is the evidence for extra-terrestrial civilizations: - "
answer:
possible resolutions to the Fermi paradox, the question of where is the evidence for extra-terrestrial civilizations:
- 1. We are looking in the wrong direction and in the wrong frequency bands for signals from intelligent life in the galaxy.
- 2. We are looking in the right frequency bands, but the technological level of intelligent life in the galaxy is not high enough to be a threat to us.
- 3. We are looking in the right direction and in the right frequency bands, but the probablity of interstellar travel is so small that we are not likely to meet civilizations.
- 4. We are looking in the right direction and in the right frequency bands, and the probability of interstellar travel is not small, but we are unlikely to meet civilizations at our level of technological development.
- 5. We are looking in the right direction and in the right frequency bands, and the probability of interstellar travel is not small, and we are likely to meet civilizations at our level of technological development.
- 6. We are looking in the right direction and in the right frequency bands, and the probability of interstellar travel is not small, and we are likely to meet civilizations at our level of technological development.
I would be inclined to think that option 5 or 6 are the most likely. Given the nature of the universe, I think that we are more likely to encounter civilizations in the earliest stages of their development, than those that are further developed.
I would think that the big question of the next 10 years is whether SETI will be able to detect an ET signal. I think that the chances of being able to do so are less than even.
The reason is that a lot of the technical issues that SETI has been addressing are becoming less of a problem, because the Earth-based radio telescopes are being joined by many smaller radio telescopes that are being sent into space. So, from a technical standpoint, the ability to look for ET signals is becoming easier and easier.
The problem is that any signals that are sent to Earth that come from technologically advanced civilizations are likely to be only a few percent of the power of the ET signals that are coming from sources at our level of technological development. So, for example, SETI@home is looking forI wonder if there are learnings for how to measure intelligence, just from proliferation of these AI models. Maybe even learnings for how to more efficiently access biological intelligence.
A human brain carries a lot more baggage with it. Long story short, our biological brain hardware is much superior to current AI hardware, but our biological brain software (where we learn the culture that surrounds us) is much inferior to current AI machine learning. GPT-3 trained 175B on basically the whole internet in 15 days. That's insane. And now there are rumors that ChatGPT is executed on an optimized subset of this 175B model, which may be a bit more than 1/10 that size.
I think going forward we'll see 20-30B parameter models perform much better than an average human in typical tasks. Especially after we evolve the topology of an AI to not be a straight directed acyclic graph, as it is now, but allow structures to loop internally, to invoke other structures with input (as subroutines) and so on.
And also there's new hybrid analog/digital hardware in the pipeline, which makes AI execution 100 times more energy efficient and compact. So above-human level AI on your phone is on the horizon, I'd say.
We're only getting started.
You literally couldn’t get GPT to come up with a single novelty if you tried. It’s all remixing existing content, and again, doing so in a way to fit the average of the dataset.
When you realize this you realize it has no intelligence as we most typically define it (novel solutions to novel problems).. its not AI. Call it what it is: a beautifully advanced way to regurgitate the exact most popular (mundane) reply you’d expect given a huge dataset.
It’s sort of good for studying what already exists. It won’t even really ever show you the edges though so it’s actually almost dangerously deceptive as evidenced by this absurd rounding up people are doing. If you want to learn the gist of anything, ask GPT. If you want to know anything in depth, GPT in fact will only mislead you towards genericity, platitudinous mediocrity.
You're also not right that GPT can't produce anything novel. Here's a tiny, modest example:
I won't waste my time, as I see you didn't read what you responded to the first time. But you need to start paying attention. You denial & empty confidence in what's happening will only get you so far.
I know many people who work in ML share this concern - people thinking it’s smart, it’s more than it is.
It’s very useful no doubt, but your comment is an example of extrapolation from maladjusted priors.
Your example proves my point, a very uninteresting looking game. I don’t even doubt GPT will find all sorts of interesting stuff by giving averaged out things. There’s plenty of gaps in the averages humans have missed. But nothing new, and until you can show that let’s just call it what it is.
On niche subjects, it fails often. But you can get it to correct itself by making it think logically by spelling out axioms, which seems to fit your definition of novelty.
All the arguments I've read about it not being intelligent seem to stem from some form of essentialism. Right now it's dumb intelligent but I think giving it an adjustable recursive depth, similarly to what BulgarianIdiot suggests, will go a long way to improving it given its base capabilities.
This demonstrates then when you instruct it "break down the problem and solve it iteratively" it comes out with superior solutions. More accurate, more novel. Ergo, then if it was allowed to iterate in this way internally, it'd produce better answers by default.
This thread sums up the same idea: https://twitter.com/ItIsFinch/status/1634730897520951296
https://openreview.net/forum?id=_VjQlMeSB_J
https://openreview.net/forum?id=5NTt8GFjUHkr
https://openreview.net/forum?id=qFVVBzXxR2V
I'm also looking forward to explicit long-term memory mechanisms that don't rely on the main architecture's weights nor on CoT.
As for "we just need a way to include iterations": isn't this what attention is supposed to do (kind of dynamically updating weights). The usual way to really update weights are variations on gradient descent. Can you link a paper outlining how we integrate your proposal into the current framework (which took around 20years to mature)? Otherwise your statements are Sci-Fi (at the flying cars-level).
Oh so it's not learning by example, it's just primed by a sample.
Oh so there's no iteration of thought, just a loop of prediction.
OK.
Also, the papers you asked about: https://news.ycombinator.com/item?id=35115563
BTW, people are already implementing recursive/iterative queries on the ChatGPT API, and getting promising results.
https://github.com/hwchase17/langchain
The general (non-technical) guideline is that the LLMs can "answer" anything you just gave them the answer for. So you give it a problem, ask it how to solve it, tell it to use that method and explain the data it needs, give it that data, and then show it everything at once: "With this data you requested and summarized, use this technique to answer this question".
How can you possibly know what's in its training set? Its training set is so large that any "guess" about what is or is not in it is a fool's errand. Given that we know how the technology works, there's no reason to claim emergent behavior when Occam's razor would say the items are likely in the training set.
It's quite easy to invent a sentence that's never been uttered before, it's also fairly easy to hit the limits of the internet's knowledge in pretty much any complex discipline that's not software engineering.
And it's apparent when you speak to ChatGPT, he will hallucinate a lot to compensate for niche topics it doesn't know much about.
Occam's razor wouldn't point to ChatGPT having omniscient knowledge of all things ever thought of and that will ever be invented.
https://acoup.blog/2023/02/17/collections-on-chatgpt/
> If you know nothing about either book, this answer almost sounds useful (it isn’t). Now this is a trivial research task; simply typing ‘the limits of empire review’ into Google and then clicking on the very first non-paywalled result (this review of the book by David Potter from 1990) and reading the first paragraph makes almost immediately clear the correct answer is that Isaac’s book is an intentional and explicit rebuttal of Luttwak’s book, or as Potter puts it, “Ben Isaac’s The Limits of Empire offers a new and formidable challenge to Luttwack.”
> A human being who understands the words and what they mean could immediately answer the question, but ChatGPT which doesn’t, cannot: it can only BS around the answer by describing both books and then lamely saying they “intersect in some ways.”
> The information ChatGPT needed was clearly in its training materials (or it wouldn’t have a description of either book to make a lossy copy of), but it lacks the capacity to understand that information as information (rather than as a statistically correlated sequence of words). Consequently it cannot draw the right conclusion and so talks around the question in a convincing, but erronous way.
It immediately cuts to the heart of the matter - an LLM does not understand things. It does not know anything but which word follows another. It can't do anything but regurgitate things it has read in an incredibly lossy manner.
But it's a neat party trick that seems to fool a lot of people into thinking that there's something legitimately useful, here.
I don't think that blog post's excercise is a particularly useful demonstration of anything.
This is still just deterministic except for a bit of noise to vary the output. There is no "thinking" going on. What you get is still just the most statistically likely output based on the training data.
Actually it turned out it didn't invent any of the games in that article. You can just Google them and find out.[0][1]
Reading is not believing.
0: https://www.digitaltrends.com/gaming/sumplete-chatgpt-ai-gam...
1: https://www.novelgames.com/en/labyrinthsudoku/
> I won't waste my time, as I see you didn't read what you responded to the first time. But you need to start paying attention. You denial & empty confidence in what's happening will only get you so far.
Maybe he didn't read what you wrote, but you didn't even bother Googling the claims in the article you posted.
> “Can you code a game called Sumplete?” Despite protesting about its inability to make games earlier, ChatGPT immediately started spitting out fully formed HTML, Javascript, and CSS code. Sure enough, it had once again created a version of Tait’s game – only this time, we never discussed what it actually was beforehand.
I guess "a version of Tait's game" leaves some wriggle room about exactly what it did.
I don't want or need to anthropomorphize AI (yet?), but this reaction by us is a tad too similar to rather uncomfortable parts of our history.
Your second link is a game different than Sumplete's rules, so not sure why you linked it.
You have one source, the Digital Trends article. It speaks about a mobile game being similar. The link here is at best circumstantial. The app page doesn't describe the rules of the game. Instead it says this:
"It is very similar to classical game of Kakuro"
I checked Kakuro, and it has similar, but distinctly DIFFERENT rules, as well. So, what's the theory here? GPT downloaded the game and ran it on its phone? No. Maybe the description of this game is somewhere. Maybe. But I couldn't find it. Where is it?
The excuse "it's somewhere, so it didn't invent it" will always be used for why an AI can't produce a "novel idea". While the reality of novel ideas is very simple: it's like conservation of matter & energy. You can never make something or destroy it. You can only rearrange it. New ideas are a rearrangement of other existing ideas. It's the case for everything we ever made. And everything we ourselves are.
It's one of those questions about the nature of us that's on the level of "is there a ghost in the machine"; "do we have actually novel ideas?".
I'm kind of leaning the way you are; I don't think we do. What seems to be happening that we _call_ novelty is just an advanced form of recursive synthesis - we take information in and remix it. But we're capable of levels of abstraction, so we can remix it really well.
I think that getting the abstraction part into these models might just be as simple as wiring them together with a meta-model. A network responsible for identifying similarity between outputs of different types, or one that serves as a connective layer, distributing tokens through other networks and then synthesizing the results.
What I'm almost certain of, is that the important changes are going to be architectural.
Three things I'm noticing:
1. We're reinventing structured programming in AI in fast forward. First it's a plain Markov chain. Then it's an "attention" directed acyclic graph. Then we realize we need loops. Then we realize we need to jump to different points in the loop. Then we realize it's useful to recursively call yourself or parts of yourself as a subroutine, parametrized with specific input. Etc.
2. Even before we fully realize this framework of thought into a model, I'm almost sure the model EVOLVES some of these structures during training. In the form of crude unrolled loops etc. Simply because it's inevitable for processing certain types of input data.
3. In order to preserve pragmatic outcomes, I'd bet the future is not one giant monolithic model for AI, but many medium-sized models, communicating in a meta network, like meta neurons, sending meta (high-level) messages to each other.
Essentially, we need to make neural networks more like a fractal. I have this rule of thumb that always works somehow: "no concept definition is complete, until it's made recursive". Neural networks will get there.
Mechanics work well for an AI generated game. No idea if it's novel or not, but it's a good take.
> These nets regurgitate the most average possible output given inputs. By definition.
You seem to be confusing the training objective with the capabilities.
For example:
> Why Can GPT Learn In-Context? Language Models Secretly Perform Gradient Descent as Meta-Optimizers
But note that this is happening at inference time, so it's not just generating the average of the input the model is trained on.
This constant reach to see human cognition as somehow unique (without evidence) is tiring and frankly an ego-extension / "made in God's image" type thought hole.
The person making the claim that this is as or more intelligent than humans should at a minimum show it do something as impressive as humans. I’ll wait for a great piece of art or science until I for no reason diminish humans, that’s all. It’s not tiring to me to see the beauty of human intelligence.
People seem to be able to draw outside the lines. It’s cool, it’s beautiful. I shouldn’t have to point to any specific amazing achievements because there are too many. I haven’t seen one from an LLM yet. I’ll know it when I see it! Not hating, just clarifying.
I'm saying that human intelligence isn't some mystic process above replication or emulation ... and that if an LLM-type computation can do what it does now I see not reason it can't match or surpass human intelligence in the future.
It may produce that intelligence in a different way to our organic squishglobes but I don't see that it matters.
We know so far that human cognition and sense of self is still largely a mystery and no evidence exists to show that it's just mechanical remixing of absorbed things. Contextually, many humans can indeed create novel suggestions, arguments and ideas, and they can self direct towards these in ways that no AI can so far. This does indeed make our cognition visibly unique without even having to mention anything religious
ChatGPT on the other hand is literally just a well-engineered design for creating coherent phrases from a huge training set of human information. As even its very creators admit, it doesn't consciously think, consider for real or perform literal AGI in any way that's sentient. For you and others here to call this equal to human cognition is absurd and not based on measurable evidence. It seems more guided by emotional awe at something presented in a new way than guided by sober reasoning. And you talk about human-centric ego arguments being tiring?
I'm not saying that ChatGPT is comparable in either capability or design to a human's cognition. What I am saying is that a human brain is just input,process and output. There's no reason a model of sufficient size can't emulate human thought in such a way that there's not much difference in capability (even if the mechanics differ).
I also think you're overstating the uniqueness of "novel" suggestions from humans and that "novel" ideas are some kind of high-water mark of intelligence rather than a fuzzy mechanism with induced feedback.
I can think of a novel idea by randomly mixing concepts I know. What matters is if this idea is good, and that test is done via interacting with the world. This mechanism can be implemented in an LLM now.
So if I ask chatGPT to make up a poem about the power of friendship, and it must involve a unicorn and a jackelope, it will give me that. What's not novel about it? Isn't it pretty much what any person would do, mash up some related words?
What does novelty mean in this context?
> Isn't it pretty much what any person would do, mash up some related words?
Yes, that’s exactly what chatgpt does, and it’s what many humans do. But to be analogous to your initial example, there should be another person who actually came up with the instruction specifying which words to mash into a poem. The word-masher, whether human or chatgpt, is just following the instructions, not coming up with them.
If I ask a random person I meet to write a poem with the same instructions as were given above I'd be surprised if they were able to come up with something as good as ChatGPT.
The novelty was not in the prompt.
That's an interesting claim I don't know how to measure.
>> Hey, ChatGPT, could you invent five words that mean hungry?
> Sure, here are five words that mean hungry:
> 1. Cravacious
> 2. Rumble-hungry
> 3. Chewyemptiness
> 4. Famishy
> 5. Voracihunger
Is that novel? I don't know. It feels a lot like the word games Lewis Carroll or Roald Dahl played, though.
Unfortunately none of this matters, because "no true novel AI" is at this point just the "no true Scotsman" fallacy over and and over again.
We'll need to learn the hard way. Thanks for the creative test.
Have you actually tried? You'd be surprised.
The fruit flys brain is capable of piloting the body of the fruit fly on a precision and energy efficiency that's simply unmatched in any kind of aerial machine we have developed. And only with those many connections, half a million? Neurons are very complex computation machines that can module their response very carefully.
The human brain does so much more than simply being able to process language. It's capable of one shot learning and to course correct incredibly easy. These features are very important for survivability and I would say are the "magic sauce" that we just haven't been able to replicate.
So I'm actually not so sure this is simply a parameters game. There could be important structures and specialisation that we are missing. In the fruit fly there are 93 different types of neurons, in the human brain iirc it might be hundreds.
You may think I don't know what "one shot learning" means, but I do know, I'm just saying it's a problem of our perspective, not of the model's capability. Aside from it having no long term memory (by design) from the chat. Yet.
Our high-level learning is absolutely nothing like how we learned as toddlers. It took us DECADES to understand complex concepts from "few shots". Only pretrained models can do that, and you're also one, in your adulthood.
What you say about extra complexity from neuron types and neurotransmitters combining with one another to modulate even simple networks is true of course. But it's also true you don't need to replicate all those organic details in a NN, because they're equivalent to just a few more "regular" parameters. Think of an artificial neuron in a network as a Turing machine. A Turing machine can do everything with just some more tape, and a neural network can do everything with just some more parameters.
And yet, our models show that we don't need AS many parameters as we thought. And that's exciting.
The whole point of machine learning is to do all that "decades" of training in a few hours.
In terms of learning material, it takes LLM a few orders of magnitude more material to learn than it does a human. After all, a human child trained with very few books (by age 7 a child would have seen only a few books) still outperforms CHatGPT trained with the same material, because ChatGPT trained on 4 books is basically useless.
So, sure, they can learn fast after all that. But so can GPT, which also one-shots a lot.
...
> I cannot phantom the amount of data and analyses that passes a typical human child in a day.
You're counting every second of every image that a child sees as a separate "image" that is learned. I'm counting the number of objects that a child sees, because people aren't doing frame-capture and analysis on every image, they're doing lazy interpolation (which is why all your peripheral vision in in monochrome but you perceive it as full color).
A 6 month old baby who has only ever seen close family (parents, siblings, pets) is quite able to discern the same features on strangers.
Nothing we have in AI right now can take 3 example models of humans, and detect hair, nose and mouth in a brand new never before seen model of a human.
Not really the same thing as training on millions of different images and videos. The number of distinct and different objects in a 7 yo child's training set is rarely the same as "every single youtube video ever uploaded".
Going back to the original point: ChatGPT needs millions of books to learn to regurgitate information as well as a semi-intelligent young adult, while typical toddlers are quite able to regurgitate what they've just seen.
Humans and animals get by with millions of times less training data than AI models need.
Then it sounds like he is slightly overestimating the dataset (since some of the frame is monochrome) and you are massively underestimating it.
However, if you substitute abstract shapes (blocks, balls, etc.), your point absolutely stands up.
In some ways, though, isn't it more interesting to have an intelligence that learns in a completely different way from humans, as opposed to one that learns through essentially the same method?
Like temporarily, or permanently, allocate some part of the brain for a certain configuration, beyond what experts can do and closer to what a processor can do.
Not especially long ago humanity knew basically nothing. The epitome of technology was the stone age. Literally crush two rocks together and then poke things with the sharp bits that flake off. Somehow, in the blink of an eye, we went from that to putting a man on the moon.
So let's do the same with these sort of LLM models and see what happens. Train one with the entirety of expressible human knowledge from the stone age. Where will it send you? I think the answer is largely pretty much where you are. In any case, certainly not to the moon. It seems we're just building natural language search.
This isn't shifting the goal posts. The problem is we keep intentionally making meaningless goal posts in pursuit of AI, because if you don't then the goal seems impossibly far off. But that's because the reality is that it probably is!
LLMs can make social consequences for us, from creating and maintaining state about opinions of us individually, and we will want to coexist in that reality. That's enough to adjust to and... respect. It doesn't seem to matter what arbitrary threshold of "intelligence" there is, in the face of an entity that creates much less arbitrary and dynamically generated - but predictable - outcomes. I don't have to understand what a human was trained on, or how smart they are, to understand the same thing. The same with an animal.
On HN intelligence is often seemed to be defined as more about new and novel thought / invention - it’s only intelligence if it thinks a thought that has never been thought/written before.
Off HN ‘intelligence’ seems to be closer defined to “is it smart enough to replace my job as a knowledge worker”. Pharmacists would say it has intelligence if it could automatically check prescriptions, and data analysts would say it has intelligence if it can answer a data request without their intervention.
IMO most people seem to define intelligence as whatever the part of their job is that currently requires analytical skill, because people are worried about the impact this could have on the middle and entry tier of knowledge work, rather than defining AI as needing to invent something fundamentally new.
These goalposts have literally been sent to space.
[1] https://twitter.com/LalwaniVikas/status/1634648646535770113
Tweet: "I hate it when my phone battery dies." Sentiment: Negative ### Tweet: "My day has been " Sentiment: Positive ### Tweet: "This is the link to the article" Sentiment: Neutral ### Tweet: "This new music video was incredibile" Sentiment:
and it's output was:
Tweet: "I hate it when my phone battery dies." Sentiment: Negative ### Tweet: "My day has been 90% balls." Seentiment: Negative ### Twitter: "i don't have a favourite type of music i just hate country" Seentiment: Negative ### Twitter: "c'mon first sunshine of the year...why the rain already?!" Seentiment: Negative ### Twitter: "Dear twitter, are you getting tired of me yet? Seentiment: Negative ### Twitter: "I don't care about the race, I just care about the president being a psych
Doesn't feel like that did the trick, but at least you got me to setup the shell command in a way where I pipe long-form prompts in via a separate file.
Rip open the code and make sure it has top_k 40, temperature 0.7, no top_p, and a repetition penalty of 1/0.85.
Proof: https://twitter.com/theshawwn/status/1632569215348531201?s=6...
Implementation: https://github.com/shawwn/llama
It feels basically identical to davinci. I think no one realizes it yet because Facebook released a horrible sampler that everyone copied.
Edit: This fork ( https://github.com/beiller/llama.cpp/tree/feature/repeat_pen... ) adds repetition penalty via two values, repeat_last_n = 64, repeat_penalty = 1.300000.
Playing with those now.
I wish I could somehow force it to output longer texts. It just stops at arbitrary points, often in the middle of sentences.
./main -m ./models/13B/ggml-model-q4_0.bin -t 7 -b 10 --top_p 0.0 -n 5000 --temp 0.7 --repeat_last_n 64 --repeat_penalty 1.3 --prompt "You are a question answering bot that is able to answer questions about the world. You are extremely smart, knowledgeable, capable, and helpful. You always give complete, accurate, and very detailed responses to questions, and never stop a response in mid-sentence or mid-thought. You answer questions in the following format:
Question: What’s the history of bullfighting in Spain?
Answer: Bullfighting, also known as "tauromachia," has a long and storied history in Spain, with roots that can be traced back to ancient civilizations. The sport is believed to have originated in 7th-century BCE Iberian Peninsula as a form of animal worship, and it evolved over time to become a sport and form of entertainment. Bullfighting as it is known today became popular in Spain in the 17th and 18th centuries. During this time, the sport was heavily influenced by the traditions of medieval jousts and was performed by nobles and other members of the upper classes. Over time, bullfighting became more democratized and was performed by people from all walks of life. Bullfighting reached the height of its popularity in the 19th and early 20th centuries and was considered a national symbol of Spain. However, in recent decades, bullfighting has faced increasing opposition from animal rights activists, and its popularity has declined. Some regions of Spain have banned bullfighting, while others continue to hold bullfights as a cherished tradition. Despite its declining popularity, bullfighting remains an important part of Spanish culture and history, and it continues to be performed in many parts of the country to this day.
Now complete the following questions:
Question: What happened to the field of cybernetics in the 1970s?
Answer: "
I'm curious if this part actually does anything to improve the output, my intuition says it doesn't help much. I don't have it running locally to test yet, though.
> # repetition penalty from CTRL paper (https://arxiv.org/abs/1909.05858)
Now I’m assuming a base rate of knowledge for this to help, but in general I find diving into the code paths of open source models, usually is a good use of time.
if you troll HN for AI related stuff and just read read read, progressively googling the terms you don't understand, eventually the haze will lift and things will start to make some sense.
source: in the middle of doing that right now
Do you mean "troll" or "trawl"?
(I'm thinking of an old saying that the best way to get an answer isn't to ask a question, but to say the thing is impossible. Can't remember it well enough to google for the exact form, however).
Trolling is done by hand with a fishing pole, sometimes on a dock.
Trawling is from the back of a boat, generally with a net.
I have a hazy recollection of folks in the early days of 4chan describing 'trolling for idiots' - as in, feigning ignorance to get a rise out of folks. Like, you post a dumb, obviously stupid but infuriating comment and watch people get all worked up about it. You know, for the lulz.
I feel like since then it's changed to just actually being a dumb asshole and enjoying when people are mad at you. Or maybe I've just gotten smarter. In any event at the time it certainly didn't seem like the kind of thing racist lunatics would go out of their way to do. Perhaps more the fool I, as I was a teenager at the time and as such that opinion cannot really be trusted.
https://paperswithcode.com/paper/most-language-models-can-be...
My work got cited by Gwern in this article! https://gwern.net/gpt-3
Looks like you don't have a 'most common/simple words' vocab restriction? That's a pretty fun one too. https://twitter.com/jonathanfly/status/1144345934857474048
I think this vocabulary filtering technique will pick up steam again now that good quality open source LLMs are proliferating. I've always wanted to try these techniques with ChatGPT, and the inability to do so has caused me to somewhat publicly complain about their lack of access to ChatGPT's vocabulary probability distribution (they don't give more than the top 10 most likely continuations).
As an aside - do you have any academic references for masking / constraining based next step transitions? It's a trick I've been messing with for years, but I probably picked this up from some older poetry generation work. Haven't been able to find any direct reference for simple 1 step masking, let alone more involved things like constraining based on the unrolled, modified markov sequence or things bigger than 1 step.
https://huggingface.co/blog/constrained-beam-search
This technique has been quietly available within HF for awhile, and has a lot of real world use-cases.
I've also been reading your paper on these topics, based on previous comments on HN. So thanks for that! Always interested in work about constrained generation.
I'd really like to see some version of the techniques from this palindrome generation https://www.ijcai.org/Proceedings/15/Papers/353.pdf revisited in modern contexts. Or some of the other Sony CSL work on this topic (e.g. https://www.francoispachet.fr/wp-content/uploads/2021/01/pap...) which I will admit I (still) do not understand yet.
This is insane
Vanilla Ubuntu desktop with a used server-grade NVIDIA GPU with lots of GPU RAM is probably your best bang for buck for messing with large model inference right now.
Mac has unified memory architecture that no other mainstream desktop can offer right now. This is the source of having up to 96GB VRAM on a laptop.
On a separate note I never understood why NVIDIA was always so stingy on RAM.
Give us a GPU with 256GB of RAM already. It doesn't need to be that fast, it just needs to be big enough to hold a huge model.
Instead they're still selling RTX4090Ti with 24GB of RAM while the free GPU I got from them at a raffle a few years ago has 32GB. WTF? I would have expected the RTX4090Ti to have 64GB minimum, being the highest-end consumer GPU.
But to me, Mac's architecture is the most inevitable final solution: the GPU becomes part of the CPU, and there's no CPU RAM and GPU RAM, just RAM. Will be great for AI too.
Very often multi-GPU training is maxing out RAM on each GPU and isn't making full use of compute capacity.
I suspect the future is tightly integrated CPU+GPU+RAM, but we may STILL have DIMM slots to acts as a buffer between "fast" RAM and disk. So your swap file goes from "fast RAM" to "DIMM RAM" before it has to go to disk, basically.
If the definition of RAM is DIMMs then it's really just a matter of merging the CPU and GPU into one unit with a massive L2 cache (or maybe we have an L3 cache) of e.g. 32GB that's able to hold a massive model without cache misses.
I've had success running the 7B model on a 16GB M1 Mac Mini (purchased in 2020!). I didn't bother to try but pretty sure the 13B model works as well.
I suspect you could grab one of those for less than $1000 these days.
It's going to perform like a brilliant person with brain injury. 8 bit is far better and it would be interesting to see if they can get it working on any reasonable hardware.
So a GTX 1060[1] or GTX 1650[1]?
A proper gamer would have a 4080, or at least a 3070, with the requisite 8gig of video ram.
Gamers with RTX 3070 or above are a small count, at least according to Steam. RTX 4080 gamers are a staggering 0.20% of those surveyed.
Today's price for 1TB of consumer RAM is around 3,000 USD; since it's paramount to be able to run these models locally, especially as they get better the privacy concerns will become even greater, it's time for hardware to step up to the challenge: x10 price reduction in 10 years, a new curve for Gordon Moore's observation.
I have a dim sense of the hardware industry, but I know that disk storage has gotten 10-100x cheaper, CPU/GPU power increases quite steadily, but RAM is still quite expensive.
Is it even possible for RAM to get much, much better than it is currently? Are there physical limitations, or is there just not the same amount of market interest in more RAM?
Beyond the supply/demand economics and also corporate greed [1], there are fabrication issues, just look at the current mess [2], but as new US-based fabs come online from 2025 and beyond [3], they will be solved to some extent.
Theoretically, the maximum amount of RAM a 64-bit machine can address is around 16.8 million terabytes [4]: as always with computing, the journey hasn't even begun.
[1] Easier and more money in manufacturing RAM for smartphones, https://fossbytes.com/flagship-phones-ram-vs-pc-ram
[2] "EU tells Taiwan to forget about a bilateral investment pact even as bloc seeks more chips", https://www.scmp.com/news/china/article/3213163/eu-tells-tai...
[3] "U.S. Semiconductor Renaissance: All the Upcoming Fabs" (2022), https://www.tomshardware.com/news/new-us-fabs-everything-we-...
[4] https://superuser.com/questions/65191/what-is-the-maximum-am...
AI models that require a huge amount of memory will be done in the cloud.
Will memory demand increase by 2033? Of course. Will demand reach 1TB of RAM for a consumer level computer/mobile device? No.
Speculative Sampling allows to run a smaller model locally and only check up with a larger model in the cloud from time to time, once every few tokens.
On the other hand, yes, everyday use (web, mail, video/media consumption) don't require today much more than 8 or 16GB of RAM. If we go a bit creative, a PC with Linux can run very smoothly on 4GB alone, and surely someone here can point to their one anecdote of a machine with 2 GB or even 1 GB sporting a nice desktop environment, or 128mb CLI-only machine.
Edit: also memory has to improve its bandwidth and data transfer rates to keep up with faster processors, so it could also improve over 10 years time without much focus on storage capacity. Or maybe they focus on latency instead, or a mix of all three. Point is that it's not a single metric to improve.
DRAM currently consists of a small capacitor per bit (cell), which can hold a small amount of charge (1) or not (0). The amount of charge it can hold, its capacitance, depends on the area and the distance between the conductors[1].
So, as you try to cram more cells into the same area, the amount of capacitance goes down. This is problematic because of leakage[2], causing the charge to leave the capacitor. This is why DRAM cells are refreshed periodically, hence the "dynamic" in DRAM. Too much leakage, and you spend all the time refreshing rather than accessing the memory.
To raise the capacitance back up, you can move the conductors closer. However, if the insulating material stays the same, this will lead to increased leakage. Thus a better insulating material is needed.
While the current DRAM cells are quite impressive, it seems they're starting to hit fundamental road blocks with the current technology[3]. There's a limit to just how little amount of charge a cell can hold and still be reliably read, and finding better insulators is getting harder and harder.
It may well be a new approach is the solution to continued scaling. Moving away from individual capacitors as the basis for a memory cell might be a way forward[4].
[1]: https://en.wikipedia.org/wiki/Capacitance#Capacitors
[2]: https://en.wikipedia.org/wiki/Leakage_(electronics)#In_capac...
[3]: https://www.techinsights.com/blog/dram-scaling-trend-and-bey...
[4]: https://spectrum.ieee.org/new-type-of-dram-could-accelerate-...
The unified memory model doesn't mean it's faster for inference. Nvidia's GPUs run circles around Apple Silicon in training and inference right now. Though a part of that is software optimization.
If anything, the reason to have unified memory SoC is to increase general efficiency, not to provide faster AI inference.
I think chip makers will instead start making their NPU (such as Apple's neural engine) really massive. It'll start to take more transistors away from the CPU and GPU. By 2033, perhaps you're buying an NPU with a CPU and GPU attached to it, not the other way around like today.
Also, 2/4/16TB of RAM will not have a 10x price reduction in 10 years. As far as I know, there's nothing in the horizon to do this. Prices will go down but it might stay relatively flat for the next 10 years.
See this chart for $/GB: https://aiimpacts.org/wp-content/uploads/2020/04/memorytrend...
Memory stopped getting exponentially cheaper some time around 2011. This is why the MacBook Air still starts with 8GB of RAM.
The one thing that maybe could have helped was Intel's Optane tech, which they abandoned last year. It's memory that sits in between the speed of DRAM and SSDs.
Sure, plenty of room for optimizations.
"there's nothing in the horizon to do this"
Sure, the prediction is that a new curve is laying just a bit beyond the horizon.
MacBook Air still starts with 8GB of RAM for the same reason they still start with a 256 GB disk and for the same reason an 8 TB SSD from Apple means +$2,200†: corporate greed since it's good enough. The argument is it won't be good enough in the near future.
† whereas a Samsung 870 8TB SSD is around 700 USD
[1] "Telehealth startup Cerebral shared millions of patients’ data with advertisers", https://news.ycombinator.com/item?id=35094866
There must be a real necessity for that to happen, and I do not see how LLMs provide substantial benefits to push for a change.
LLMs are merely a technical detail, a stepping stone, the major breakthrough will be the first algorithm of a synthetic agent. Do we really want the boring dystopia HAL, Rehoboam, and so forth, or our own personal Star Trek Data.
IDK about DRAM density breakthroughs though, in fact I kinda worry scaling will slow down over the next decade.
Seems it now also supports AVX2 for x86 architectures too. https://twitter.com/ggerganov/status/1634588951821393922
I still think that open source LLMs models have to be much smaller than 200GB and to be much better than ChatGPT to be more accessible and highly disruptive to OpenAI.
It is a great accident needed thanks to Meta. For now one can use it as a service and make it as a SaaS rather than depend fully on OpenAI. Open source (or even free binary only) LLMs will eventually disrupt OpenAI's business plans.
Mostly using https://github.com/oobabooga/text-generation-webui/, the AUTOMATIC1111 of textgen.
OpenAI limits prompts to 4096 tokens.
If there was no limit, could the LLM be fed a 100 page document in the prompt and then answer questions about it?
You can potentially recursively summarize/create index to breakdown larger texts into smaller chunks.
>Despite recent success in large language model (LLM) reasoning, LLMs struggle with hierarchical multi-step reasoning tasks like generating complex programs. For these tasks, humans often start with a high-level algorithmic design and implement each part gradually. We introduce Parsel, a framework enabling automatic implementation and validation of complex algorithms with code LLMs, taking hierarchical function descriptions in natural language as input. We show that Parsel can be used across domains requiring hierarchical reasoning, including program synthesis, robotic planning, and theorem proving. We show that LLMs generating Parsel solve more competition-level problems in the APPS dataset, resulting in pass rates that are over 75% higher than prior results from directly sampling AlphaCode and Codex, while often using a smaller sample budget. We also find that LLM-generated robotic plans using Parsel as an intermediate language are more than twice as likely to be considered accurate than directly generated plans. Lastly, we explore how Parsel addresses LLM limitations and discuss how Parsel may be useful for human programmers.
just checkout Haystack tutorials. I started looking into it after getting introduced to the concept by articles mentioning OpenAI embeddings and GPT 3 api, but it can be done using open source models.
Langchain docs are helpful, and it would be even better if you published an end-to-end notebook using a popular dataset. Definitely looking forward to try langchain as I dive deeper into this.
1. https://haystack.deepset.ai/blog/how-to-build-a-semantic-sea...
So asking a simple one sentence question to a model that has a 81,920 token limit would be a collosal waste of resources.
A set of data structures to augment LLM's with your data: https://github.com/jerryjliu/gpt_index
What does that mean?
For each token in your input or inference output it requires the model to have some understanding of what the position of the word means.
So there is the word position embedding matrix that contains a vector per position. The matrix has "only" 1024 entries in it for GPT2 or 4096 for GPT3. The size of each entry varies as well, containing a vector from 768 for GPT2 small and up to 12,288 for GPT3.
So the WPE (word position embeddings) for GPT2 is (1024x768) and for GPT3 (4096x12288)
Inference requires info from this vector to be added to the word tokens embedding for each token in the original prompt + each generated token.
As often is the case with these large models, you can change it with some finetuning on longer context samples from the same dataset, with what is really a small amount of compute invested compared to the million hours spent on training the thing.
I just recently tried out one of the most popular [0] Stable Diffusion WebUIs locally, and I'm positively surprised at how different it is to the rest of the space around ML research/computing. I consider myself to be a competent software engineer, but I still often find it pretty tricky to get e.g. HuggingFace models running and doing what I envision them to do. SpeechT5 for instance is reported to do voice transformations, but it took me a good bit of time and hair-pulling to figure out how to extract voice embeddings from .wav files. I'm sure the way to do this is obvious to most researchers, maybe to the point of feeling like this needs not a mention in the documentation, but it certainly wasn't clear to me.
The community around Stable Diffusion is much more inclusive, though. Tools go the extra effort to be easy to use, and documentation for community created models/scripts/tools is so accessible as to be perfectly usable by a non-technical user who is willing to adventure a little bit into the world of hardcore computing by following instructions. Sure, nothing is too polished and you often get the feeling that it's "an ugly thing, but an ugly thing that works", but the point is that it's incredibly accessible. People get to actually use these models to build their stories, fantasy worlds, to work, and things get progressively more impressive as the community builds upon itself (I loved the style of [1] and even effortlessly merged its style with another one in the WebUI, and ControlNet [2] is amazing and gives me ideas for integrating my photography with AI).
I think the general interest in creating images is larger than for LLMs with their current limitations (especially in current consumer-available hardware). I do wonder how much this community interest will boost the spaces in the longer run, but right now I can't help but be impressed by the difference in usability and collaborative development between image generative and other types of models.
[0] https://github.com/AUTOMATIC1111/stable-diffusion-webui
Ooba's textgen webui runs LLaMA 4bit on 8GB VRAM consumer video cards.
KoboldAI (https://github.com/henk717/KoboldAI/) by the creator the AI Horde of stable diffusion fame also supports LLaMA and has forked the Stable Horde for textgen. Currently horde is running half a dozen LLaMA models accessible at http://lite.koboldai.net/
Go to 4chan right now, and poke around their technology and video game boards. There's so much chatter about LLaMa. The last time I saw that much chatter about a technology was when eth was 3 dollars a coin. The communities exist, the general public just isn't aware of them.
George Hotz was so confident that he was riding the wave with his Python implementation: https://github.com/geohot/tinygrad/blob/master/examples/llam.... But I guess not, pure C++ seems better.
That basically the whole AI revolution is powered by CPython of all things (not even PyPy) is the 100 megaton nuke that should end language warring forever.
That the first AGI will likely be running under a VM so inefficient that it refcounts even integers is God laughing in the face of all the people who've spent the past decades arguing that this language or that language is "faster". Amdahl was right: only inner loops matter.
And a lot of new AI tooling such as tokenization has been developed for Python using Rust (pyo3)
If mainstream AI moves beyond just deep learning and start to encompass other approaches and disciplines (for example, differentiable simulators/renderers), suddenly the language (and the compiler) will be much more important than before.
I can see fake accounts, phishing and disinformation at a whole new scale, too.
Stable Diffusion kicked off a wave of interest which hasn't stopped building yet - albeit massively accelerated by ChatGPT's release in November.
What do you even mean!
It is being used right now on all sorts of creative industries for things like ui/Ux generation, concept art, vfx production etc
I myself use it for concepting and color correction on video games pipeline production
Before Stable Diffusion democratized image generation, Very Serious People across the industry claimed that unless they and their friends exercised strict control over the kinds of images people generated using AI, society would collapse, the sky would fall, and it would all be over. After that, the public gained access to diffusion models for imagine generation, and the sky did not in fact fall. These prognosticators of doom are wearing egg on their faces.
Likewise, today, the intelligentsia is predicting DOOM! DOOM I SAY! if LLMs become available outside the close supervision (and censorship control) of San Francisco trust and safety teams.
Democratizing LLMs will further delegitimize claims that some central authority needs to control this new technology. This is great! When the printing press came out, it took a hundred years to democratize it. It'll take us ten months!
The single biggest benefit of projects like llama.cpp is annihilating claims on the past of gatekeepers that for our own safety we need to let them control how we use math.
Not math. The magical sauce is not math, it's the raw text used to train the model. Language data has this property that it can take a raw initialised neural net and turn it into LLaMA and chatGPT.
The same language data turns babies into modern empowered adults instead of cave men or worse. We're not that smart without language. A single human can't rediscover it, it is a self replicating system evolving in parallel with us.
I wouldn't order figures as Elon Musk or Peter Thiel as intelligentsia. They are just greedy operators for the MIC.
I'm not so sure this is true
Have you tried it?
If they were using techniques like GPTQ, presumably it would work much better.
On top of that, you may want to do fine-tuning with reinforcement learning. But the above is “all you need” to get into chat mode with a LLM (or any text generation model, really).
Prior to LLaMA + llama.cpp you could maybe run a large language model locally... if you had the right GPU rig, and if you really knew what you were doing, and were willing to put in a lot of effort to find and figure out how to run a model.
My hunch is that the ability to run on a M1/M2 MacBook is going to open this up to a lot more people.
(I'm exposing my bias here as a M2 Mac owner.)
I think the race is now on to be the first organization to release a good instruction-tuned model that can run on personal hardware.
And given how early the cpp port is, there is likely plenty of performance headroom with more m1/m2-specific optimization.
I suspect people saying it's not good are prompting it like ChatGPT, not realizing how much trickier a raw model is to prompt. Getting the hyperparameters for good sampling is another stumbling block. The models are very good if you do everything properly.
Personally I hope we quickly get to the stage that there's a real open llm like SD is to DALL-E. It sucks to have to bother with Facebook's core model, and give it more attention than it deserves, just because it's out there.
If facebook had actually released it as an open model, I would have said that all the credit should go to them. But instead people are doing great open source work on top of their un-free model just because it's available, and in the popular conception they're going to get credit that they shouldn't
Llama 7B is much better that something like GPT-Neo at text generation.
Kudos to Yann lecun for getting his revenge for Galactica
https://blog.google/products/pixel/introducing-google-tensor...
TPUs were released in 2016, Tensor Cores in 2017. But Edge TPUs (that is, the ones that exist outside of Google's Cloud infrastructure) weren’t until 2018. So, Nvidia had their cores available for people to buy first, but not before TPUs existed.
Some video cards/GPUs have become just that, becoming more and more geared towards non-graphics workloads ...
My money is on something like processing-in-memory enabling inference on gigantic models with rock bottom energy consumption, and rolling out to the masses as a PCIe device. Lots of research on this direction, see https://mythic.ai for an example.
Neuromorphic hardware
Artificial neural networks with backprop is all we need for AI
Same thing is likely to happen here : artificial neural networks directly in hardware might have extra possibilities for optimization ?
No. Biological neural networks run many orders of magnitude more energy efficiently, and learn substantially more data efficiently and continuously.
We've got a lot of catching up to do with nature. In my opinion, backprop on today's compute architecture won't ever compete (I can elaborate if you like) - we need some fundamental changes in our approach.
We're currently running up against the Von Neumann bottleneck in ML. It's a great architecture for general purpose compute, but not suited for efficient operation of neural networks: a ton of energy is spent shuttling weight and activation values between GPU RAM and ALUs. Biological networks appear to be able to process a large amount of information per unit energy by exploiting the static physical structure of the network. A small amount of energy in sparse neural activations can transform a large volume of information stored in the structure and synapse strengths.
Here's some very rough napkin math: assume the brain has 100 billion neurons, each with 1000 synapses, and is sparsely activated at ~5% activation, and processes at 30Hz (gamma wave frequency, roughly). This means every "tick", 5 billion neurons must interact with 5 trillion other neurons. If you tried to implement this in a Von Neumann architecture, even if you only calculate for the sparsely activated neurons, and even if you quantize your weights values to 1 byte, this means processing ~150 TB/s of weight data. That's an insane amount of memory bandwidth. An A100 is 2 TB/s at 300W, but our brain only uses 20W.
My two cents is that the future will involve a hardware architecture that let's us avoid moving these weights around during inference. Whether this will mean a "mortal computer" as Hinton has recently discussed (https://www.youtube.com/watch?v=sghvwkXV3VU) or whether the weights will be loaded and fixed at init time (as with Mythic's or other neuromorphic approaches), time will tell.
Well, we have dedicated coprocessors for them (Google TPUs, Nvidia Tensor Cores, Apple Neural Engine, etc.)
There are dedicated cards, though some systems come with them on the mainboard and some videocards have them as well.
Can anyone explain how llama.cpp works? On a very high level, I'm interested in how it's possible to run it on CPU.
Also do I have to bother with their crappy driver module that doesn't support most GPUs?
If your CPU supports AVX2, the output will be "avx2". If your CPU does not support AVX2, the command will not produce any output.
If you want to run on GPU, use the official Python-based stack, which BTW takes about 10GB of runtime binaries on disk, and only supports nVidia GPUs because CUDA.
The python + GPU one can be found on the official facebook repo: https://github.com/facebookresearch/llama (Presumably GP thought this was already known to everyone so they pasted the other link)
1. Do you have a link which proves the implementation being discussed is CPU-only?
2. Do you have a link to the official Python-based implementation?
When I wrote the above answer, I was only aware about the first interpretation.
Do we still need to fill out a form and say we are a researcher?
Not there yet?
Does this still use the gpu?
If it used the GPU/ANE and was a true large language model then it would only work on M1 systems because they're unified memory (which nothing except an A100 can match.)
LLaMA's GPT-3 175B level model, LLaMA-13B, only requires 8GB of VRAM (and no RAM) to run using pre-quantized 4bit weights. So it's hardly a job for an A100.
Even the largest model, LLaMA-65B, is only 30GB and inference can be split between two graphics cards with almost no effect on performance.