Academics can’t put themselves on a pedestal for being revolutionaries who care about the ultimate truth the most and at the same time whine when an advance comes along.
These CS academics need to understand that this is what it feels like to be in other fields like physics or bio where you can’t do jack unless you have costly equipment. This is what people in developing countries deal with all the time.
And people will build the next GPT or whatever and even that’ll get boring. The people in big tech have to pivot and do whatever the economy demands. Whereas the academics can go back to writing papers like nothing changed.
I don't think they are putting themselves on a pedestal. It's more that they made the choice to remain in academia, because they believed that were trading a higher salary for the academic freedom and the chance to work with really cool problems.
And then it turns out, industry are the ones really working with the really cool problems. (Disregarding the fact that ~85% of machine learning academics do not have what it takes to be hired by OpenAI anyway).
I think your response is very narrow in defining what a "really cool problem" is. If in the 1970s you imagined that the cool problem in ECE/CS was "building ever-more-powerful computer processors" then you basically got smoked by Intel and other industry labs.
But there's a different way to look at it. Because industry went off and solved the now-"boring" problem of building ever-more-powerful computer processors, the people left behind in academia got to invent machine learning, public-key cryptography, modern coding theory, distributed systems... and so on. What makes academic research valuable is not the freedom of "I get to plan my own day", but the freedom of "I get to work on the weird problems that industry doesn't even realize are problems."
Public key cryptography was invented in the 70s, commercialized in the 80s.
In machine learning we already had back-propagation in the 1970, but had to wait another 40 years for Intel and Nvidia to create ever more powerful processors to tackle useful problems.
The "cool problem" is cracking the mystery of intelligence, of consciousness. Building a general AI. Until recently, this was mostly an academic pursuit, and people expected it to take decades or more. But now, suddenly, it's highly likely the problem will be cracked within a decade, and it will be done by corporate R&D teams, by means of scaling up the transformer architectures.
So I get how they feel - one of the coolest problems ever instantly went from being something anyone could hope to contribute to, to something almost no one can. You can't match the compute to do the things OpenAI & Google & friends do. And if you happen to stumble on something related that can be explored without access to obscene amounts of capital, guess what, the corporate research teams will notice it, and they can do it better than you, and then they can the idea much further than you ever could.
Has industry "cracked the mystery of consciousness"? Has anyone at OpenAI even made a plausible start at explaining what the hell is going on inside these LLMs that enables them to produce such human-like comprehension? To me these are the really exciting questions.
From what I can tell, the team at OpenAI (following after Google/Deepmind etc.) are simply mashing the pedal to the floor to get bigger and better models from their existing techniques, and then tuning the resulting black box to make it produce more "useful" answers. And that's fine! That's precisely what an industry lab is expected to do: they have the resources to do the training and the need to get products in front of paying customers as quickly as possible to justify it. And frankly with top AI engineers getting paid millions of dollars and Google/Meta tight behind you, emphasizing results is the most viable strategy. If "turn the needle on the box to the right" is giving you good answers, why would you waste a $1-$5m-salary engineer on academic questions like "why does the box do that?"
And yet, asking questions like "why does the box do that?" is the reason technology didn't stop at the steam engine. I suspect that finding the answer to those questions won't immediately sell enterprise licenses, but will be very important. And the answers will probably fall to someone who's making $30k/year in a graduate program.
ETA: Of course, it may turn out that "turn the needle on the box to the right" is enough to obsolete all human researchers, in which case I'll be wrong about this. But it'll hardly matter in that case ;)
> Has anyone at OpenAI even made a plausible start at explaining what the hell is going on inside these LLMs that enables them to produce such human-like comprehension?
I'm only an amateur in the field, so my uneducated high-level understanding is that, in a sufficiently high-dimensional latent space, there's more than enough dimensions to assign to any single semantic relationship people ever thought of, which is what the training process effectively does, which reduces an important part of thinking - working with concepts and their relationships - entirely to vector adjacency search.
I'm only beginning to study the details, and I don't know how much of specific understanding of this exists, but at the very least this high-level model explains why scaling makes qualitative difference here.
Now, I agree they have strong commercial incentives to push their models as far as possible as fast as possible, but honestly, if I were a researcher working on these models, even if I was somehow unconcerned with any kind of commercial viability and had access to more compute, I'd absolutely keep scaling those models up and up, all the way until I hit the limit of available compute, or the models stop qualitatively improving with scale.
Basically, there's no reason[0] to stop now and try to fully comprehend how GPT-2 works, when GPT-3 was a qualitative jump, and GPT-4 even more so, and GPT-5 is around the corner, and GPT-6 might be a year away from now. All those steps yield important new insights into how the whole architecture works, and if at some point the scaling breaks, that would be even more important knowledge to have. And this doesn't even take into account the fact that, starting with GPT-3, those models are increasingly useful in accelerating both research and scaling alike.
----
[0] - Except, of course, that if transformer models are the road to generic AI, then we'll just blindly race straight into a point of no return.
>> But now, suddenly, it's highly likely the problem will be cracked within a decade, and it will be done by corporate R&D teams, by means of scaling up the transformer architectures.
I'll counter that. In the end we need AI that can do training AND inference on edge devices out in the real world. A good (and possibly profitable) example would be robotic pets that can learn (even to understand words) and interact with their owners like real animals, but don't need to go to the vet or eat and poop. Big companies relying on huge compute resources are not even aiming at this type of thing. They're too busy using their "scale" to even bother looking at smaller but interesting methods or solutions.
That's fair, and I'd agree in any other case, but This Time Is Different™. If (when) someone cracks this problem, they will be able to apply compute to every other problem. Some marketer convinces some C-suite that the world needs smart robotic pets? Why, here's access to ChatGPT-8, make it so. Just don't burn more than half a billion dollars in a week on it.
That's of course if, by this point, we aren't in a middle of a futile scramble to avoid getting extincted by ChatGPT-7 that someone left in self-play mode and forgot to turn off before going on vacation.
Point being, general AI is general. Even at extreme expenditure of resources, the closer the corporations get to it, the more problems they can put it to - including, eventually, the problem of optimizing itself. Already today people are using current-gen models to assist in developing next-gen models; this trend will only continue, until at some point you'll be able to let the model self-improve, mostly unsupervised. I imagine the compute costs per AI value delivered will drop like a stone then.
OK, I agree with you. But this sort of struggle that you highlight is common in many aspects of research. As a researcher you are conflicted between your curiosity instinct to explore new paths, and your desire to also have big results. Ideally you would have a balance between the two, although you can definitely make a career out of only one.
As someone in this space I can attest that AI teaching in most (UK) universities is generally poor on detail, abstract and behind industry by at least 3-5 years.
Not to mention that there is zero appetite from undergrads or postgrads to get into the nitty-gritty of it. To learn CNNs at the deep-dive level you need calculus, at least differentiation and integration. Calculus or even pre-calculus doesn't form part of the degree programme for most compsci BScs any more, because it is 'too hard'.
The way most students 'learn' AI is to use a method out of a Python library with near-zero understanding of how it works, and regurgitate it for an assessment.
Professorial research staff in most UK universities are light-years from AI within industry, and there's no clear path to that gap tightening, especially while universities are being run like second-rate consulting houses (don't get me started on THAT).
lmao even in my university (the serbian uni), we have at least calculus+linear algebra before any nn course. also to "learn" what's a cnn you just need gradients not integrals (unless you use some kind of non-lipschitz function as activation?), plus the idea of what a convolution is... but even Mobius knew it back in the 800s.
anyway i think your statement that industry is light years away from unis is just misleading. i think the two are trying to answer different questions:
1. how can i achieve a "somewhat" decent chatbot that gets me rich albeit not even knowing what it does [industry in case you wondered]
2. try to understand, quantify and measure how well a model works, is it stable? does it converge if we have small datasets? and so on so forth.
just my two cents, to conclude i think a good analogy to the current climate is the 700-800s with electromagnetism: plenty of people discovered "empirical" laws but didn't understand really the phenomenon.
You might be able to understand what a convolutional network "is" without calculus, but you'll be woefully unequipped to ask even obvious questions like "what if we put Fourier transforms around the convolutional layers" (a cursory search suggests it provides the expected speedup but is for some reason not a standard thing to do?). As someone outside of the industry, I'd also imagine any effort to explain what NNs are actually "learning" (or I suppose dually, how to design network architectures) is going to have a lot of fruitful overlap with signal processing theory, which is heavy on calculus, linear algebra, probability, etc.
> a cursory search suggests it provides the expected speedup
What do you mean? Processing a CNN layer takes an amount of time that does not depend on the input data, only the input/output sizes. Fourier transform is just a change of basis. Why should anything speed up?
Because convolution is an O(N^2) operation, but a Fourier transform (and its inverse) can be done in O(NlogN) and turns convolution into O(N) multiplication. So if you do FFT, multiply, Inverse FFT, you get convolution in O(NlogN). I would guess that you don't even need to do the inverse FFT and can just learn in frequency space instead, but maybe there's some reason why that doesn't work out.
Computing convolutions using FFTs is efficient for large kernels (or filters). Most convolutions in popular ML models have small kernels, a regime where it is typically more efficient to reformulate the convolution as a matrix multiplication.
I think your complexity argument is correct for N=pixels=kernel size. But typically, pixels>>kernel size.
Disclosure: I work at Arm optimising open source ML frameworks. Opinions are my own.
I see, the wording confused me. You don't really "put Fourier transforms around the convolutional layers" because you have to completely replace the convolutions.
This seems to be done in some cases. I guess it isn't done more widely because the "standard" convolution kernels are very small and the performance would actually be worse?
>just my two cents, to conclude i think a good analogy to the current climate is the 700-800s with electromagnetism: plenty of people discovered "empirical" laws but didn't understand really the phenomenon.
Sounds dead on. Do these large """language""" models actually even implement any concepts from linguistics? Or is the entire "language" part of the model merely derived from the fact that it's inherently part of the training data?
I don't fault Chomsky at all for being fed up with the hype here.
The entire field is also glossing over the fact that other languages which aren't English exist.
>> anyway i think your statement that industry is light years away from unis is just misleading. i think the two are trying to answer different questions: 1. how can i achieve a "somewhat" decent chatbot that gets me rich albeit not even knowing what it does [industry in case you wondered] 2. try to understand, quantify and measure how well a model works, is it stable? does it converge if we have small datasets? and so on so forth.
GP here is, IMO, confusing what the corporations want (1), with what corporate R&D people want (2). As long as the corps see good ROI on throwing infinite money at their AI R&D departments, then those corporate researchers are better positioned and better equipped to do actual, solid science, than academia ever can be. This has happened many times before, including in this industry. Research is best done by well-funded teams of smart people left to do whatever they fancy. When those conditions arise, progress happens, and it doesn't matter whether it's the government or industry that creates them.
(Conversely, the best hope for academia to become relevant again is that corporations lose interest in this research, and defund their departments. This could happen if e.g. transformers end up being a dead end, or compute suddenly becomes very expensive.)
> Do these large """language""" models actually even implement any concepts from linguistics? Or is the entire "language" part of the model merely derived from the fact that it's inherently part of the training data?
The latter. And guess what, they're not trying to solve the issue of linguistics. They started as tools to generate human-sounding text, but in the process of just throwing more data and compute at them, they not only got better, but started to acquire something resembling concept-level understanding.
It turns out that surprisingly many aspects of thinking seem to reduce well to proximity search in a vector space, if that space is high-dimensional enough. This result is both surprising and impactful well beyond the field of AI. It's arguably the first potential path we identified that the evolution could take to gradually random-walk itself from amoeabas to human brains.
i'm not saying LLM are modeling linguistics in any way lol. i only meant that there's some kind of phenomena related to scaling+attention that produces good enough result for most "human language stuff", which is kind of unexpected (i mean everyine knows that if you build a large enough model you can teach it any function, but cmon it is architecture+scaling that made it possibile not scaling alone). Moreover, the architectures used, from attention layer, even LSTM for that matter are not completely understood, are being used because "they works" just as in the old days of electromagnetism the empiral laws "just worked" for their usage.
btw, in other languages i guess it is decent although it depends on which language, at least gpt4.
> Calculus or even pre-calculus doesn't form part of the degree programme for most compsci BScs any more, because it is 'too hard'.
In the US I’ve never seen a BS in computer science that didn’t require calculus. I can’t speak for the UK, but it would surprise me that what you say is true.
It's not interesting. You can regard OpenAI as one of the most prestigious AI labs in the world and only the best of the best get a chance to work there.
It’s very interesting that you think most people with a PhD working in AI / ML / Data Science academically can’t get a job at OpenAI.
You can easily see a lot of people who work at OpenAI on LinkedIn. University AI labs are always left out because non-commercial products just aren’t as noticeable.
I highly doubt that many of these academics would struggle being amount candidates at DeepMind, OpenAI and FAIR.
Prestige is very subjective, which makes it also very broad thankfully.
> These CS academics need to understand that this is what it feels like to be in other fields like physics or bio where you can’t do jack unless you have costly equipment. This is what people in developing countries deal with all the time.
This was my thought too. It cost approximately $4.75 billion to build the Large Hadron Collider at CERN. My educational background is in physics, but I would not be the least bit upset if a big chunk of public research funding was shifted from physics to CS/ML/AI. It’s obviously far more important to the future of humanity than finding the Higgs boson.
I don't know, core physics leading to Fusion reactors and FTL travel seen much more important to me than AI that can make convincing robocalls and generate furry porn stories.
It may or may not. I mean it’d be pretty hard to increase yields of GPU fabs or data center sizes another 100x. There are logistical limitations. Unless some Apollo level mission is created by a superpower, we will hit bottlenecks. Algorithmic innovation is the only long term bet.
I thought that too, a year ago. But then chatGPT Turbo came out (ten times cheaper), and a slew of 30B and 10B models that are decent. Now I believe we will be able to run non-trivial AI on trivial hardware. Not to mention the Stable Diffusion revolution, hardware requirements went down pretty fast. Even an old GPU from 5 years ago can generate images quickly. Some LLMs run on iPhones.
The trick is always to offset the cost of inference with larger cost of training. You don't apply the Chinchilla scaling law, that's for academics and people who don't have to pay for inference. You pretrain the model 10x longer (like 1T tokens for LLaMA) to make it the best you can fit into an A100 or 4090 quantised to 4 or 3 bits. So everyone can have AI assistants running on their own toys.
And? ChatGPT doesn't give you logprobs. It doesn't allow you to alter the generation algorithm. They've left their lunch out on the table for someone else to eat.
These CS academics need to understand that this is what it feels like to be in other fields like physics or bio where you can’t do jack unless you have costly equipment. This is what people in developing countries deal with all the time.
And people will build the next GPT or whatever and even that’ll get boring. The people in big tech have to pivot and do whatever the economy demands. Whereas the academics can go back to writing papers like nothing changed.