As a theorist in AI, the premise is funny but also really sad. And yet it is a very pervasive premise in the field.
OpenAI is not doing science. They are building a big shiny thing, showing it off, keeping it closed, and making money off of it. That is not part of the scientific process.
It is as if, every time SpaceX launched a rocket a little bit higher, every aerospace department ooh'd and aah'd and threw up their hands: "we don't have the resources to build a bigger rocket; how can we do science?"
Unfortunately, the research field of AI is diseased: it places way too much value on showing off big shiny things above progress in scientific understanding. But there is a ton of progress available to be made by small teams on small budgets. There is so much we don't understand about neural networks alone, even small ones.
What? OpenAI is absolutely conducting science amongst themselves. They are certainly engaging in the hypothesis -> test -> result scientific method, which is science in its purest form. Sharing the secret sauce is what is done or expected in academia, but that doesn’t mean that OpenAI isn’t conducting science behind their own walls. You’re conflating institutional academia with the scientific method.
If I go home and test the boiling point of water in my garage and record the result privately without publishing a paper, am I not conducting science? What nonsense.
By the "scientific process", I mean the way scientific knowledge as a whole progresses over time, i.e. the main undertaking of the target audience of the paper.
You're right that "OpenAI is not doing science" was hyperbolic, but I tried to clarify the aspect of science that I'm referring to.
Although this is a reasonable definition, it would mean many people who are referred to as scientists in practice would have to be considered non-scientists. Refusing to publish data and code is widespread in many academic fields, even in cases where academics sign documents promising to provide these things on demand.
For example, at minimum psychology, epidemiology and climatology have this problem. In climatology attempts to get the source code of models and data underlying papers has led to some of the biggest dramas in the field e.g. the Climategate hacker appears to have been motivated by the closedness of the field (similar to the llama leaker).
Given that society is clearly not ready to reclassify large parts of academic work as non-scientific, it seems like this definition cannot work even on its own terms.
"Science" is not solely the scientific method - it's the democratization of your methodology to enable the replication of the results, towards the goal of determining their soundness.
Yes, that is exactly the point. "Research" and "Scientific Research" are not inextricably linked - I work in an R&D department, and not all research we do is scientific. We do have scientific research, but it's handled very differently from the PoC type of work that occurs more broadly. We are forced to choose between handing out trade secrets in exchange for gaining more soundness in our theories, or keeping them sheltered and hoping our observations hold up. That doesn't mean our research can't inform our decisions towards profitability, only that it's not scientific.
Edit: I mean, c'mon - the final step to the scientific method is "communicate your results". I'm surprised this requires explaining.
Alchemists were doing research 600 years ago but they never shared their results with others and used arcane syntax to record those results which nobody else could understand. This didn't lead to much progress in the field of chemical knowledge, because it was just continual reinvention of the wheel.
The scientific (and then, industrial) revolution didn't begin until researchers started sharing their results in open literature form, or at conferences, or within bodies like Britain's Royal Institution. Dissemination of knowledge is absolutely part of science.
Similarly, a fundamental necessity in science is independent replication of results, and that's definitely not possible with closed secretive research done in the alchemical mode.
What if the science is disseminated within a closed group? For instance, suppose the classified research is shared with other people with a security clearance. That still wouldn't be science? Or suppose scientists at universities shared their research with scientists at other universities, but not the general public: that wouldn't be science either?
What happens when you have an authoritarian system in place controlling distribution of results, is that everyone in that system quickly learns that if they question the reliability or wisdom of the research, their career is over. So they go along and don't subject the results to the kind of scrutiny science relies on, because making the boss look bad or embarrassing the institution isn't an option in that situation.
There are many examples, such as the secretive public-private DOE FutureGen program aimed at "zero-emission coal-to-hydrogen plants" which wasted about ten billion dollars over twenty years with nothing to show for it (and whose data is still not available), the Trofim Lysenko plant breeding program in the Soviet Union which wrecked Soviet agriculture for decades (and falls into your latter category) and so on.
Another example could be the Challenger Space Shuttle disaster, although that's more an engineering failure than a science failure, but it had similar underlying causes, i.e. managers pushing engineers to go along with a flawed protocol.
That's an echo-chamber, especially if the closed group is consistently aligned in their goals so that they're predisposed to want the same things from the research. The entities would have to be arranged so that they're asserting checks-and-balances on one another.
Exactly. They are applying the scientific method, but they aren't contributing to science. it will be science when it's open and people can integrate the new knowledge in a broader set of knowledge.
Science is knowledge. The etymology of science is to split stuff and understand how it work. But from late 14c in the more specific sense of "collective human knowledge".
"Private" science is not science. It's some esoteric stuff.
I don't buy the argument that "public" science is the only real science. You can define it that way, but I don't see any reason to.
"Public" science is just a special case of "restricted" science, such that the set of people the knowledge is restricted to is equal to the set of all humans.
I think this might get philosophical quickly, but to really understand where you're at, you need to access all knowledge. Everybody is standing on the shoulder of giants. It's a collective effort that happens through socialization and long-standing institutions like universities.
An essential part of doing science is sharing results and waiting for others to challenge, expand, or falsify them. If you skip this step, you're not truly contributing to knowledge, although you may still create new knowledge in other ways... but not all new knowledge is necessarily scientific, it can be private, esoteric, commercial knowledge.
Take the example of Ida, a 47-million-year-old lemur fossil known as Darwinius masillae. https://en.wikipedia.org/wiki/Darwinius it was "discovered" and hailed as a transitional species bridging primates and mammals. Really a huge contribution to "science".
But here's the catch - Ida was actually found in 1983 and held by a private collector for 25 years, who didn't realize its significance. So we had no idea of this "discovery". It wasn't science, it was a good on a market.
OpenAI is the same. We might enjoy to toy with the technology. But I don't find it contributed to science, yet
I love to get philosophical, so I'm gonna throw this thought experiment here:
Let's say that there was a highly advanced civilization that existed way before the current civilized society. They performed scientific experiments and gained a lot of insight, but a catastrophical event erased them from the face of Earth together with all their knowledge.
Did that civilization engage in science?
If you think that they did, indeed, engage in science, then what is the difference between them and any other closed community that performs scientific experiments?
Or, if you think that they didn't, in fact, engage in science, then does that mean that science is only science if you and I, today, can gain access to the knowledge derived from the experiments? Therefore, science is defined by its utility to us?
> catastrophical event erased them from the face of Earth together with all their knowledge.
This thought experiment seems to introduce a huge variable which eliminates what I think might be a key factor in what many readers believe to be the practice of science.
> what is the difference between them and any other closed community that performs scientific experiments?
That difference here, and the key factor, would be intent. The catastrophic event eliminates any meaning behind what they accomplished or practiced in regards to why they practiced science in the first place.
Practicing it behind closed doors may appear less like science for science's sake and more like, say, an economic pursuit that coincidentally involves the scientific method.
That could be why the conversation appears so split at the moment. It's hard to claim science is about intent or a long-term purpose that, were it removed, it would bar a scientific activity from actually being Science proper. That's a very subjective thing.
> The catastrophic event eliminates any meaning behind what they accomplished or practiced in regards to why they practiced science in the first place.
But if they didn't suffer a catastrophic event, we would call their activities "science"? So, because they did suffer a catastrophic event, what they did is not science (or is not relevant, since you avoided answering the question directly)?
And if our current society was destroyed tomorrow in a catastrophic event along with all our science, and a new society arisen from the ashes that performed scientific experiments, would that mean that what we did wasn't science? Or that what the new society is doing isn't science?
I know I'm streching things here, but words (should?) have meanings in all possible scenarios, and thought experiments are used to decide meanings in such edge cases.
> Practicing it behind closed doors may appear less like science for science's sake and more like, say, an economic pursuit that coincidentally involves the scientific method.
Does "not doing science for science's sake" include only monetary motivation? Would it also mean that someone who primarily wants to achieve status and fame as a scientist also isn't doing real science? Or are they exempt from the rule, since they would most likely release their knowledge into the public (which brings us back to the "science is only science if the current society benefits from it")?
I believe the “advanced civilization VS closed-doors science” problem you propose is inherently faulty. The final step of the scientific method is “share your results”, and they did that. It cannot be accessed anymore, but they attempted to share it and it benefited their people at the time. However, closed-door science benefits no one.
Take, for instance, the Greco-Roman mix for concrete. They produced that scientifically, repeating experiments with their volcanic ash-filled cements to make the best they could, and shared those results across their territory. While that knowledge was lost for centuries, it was still science.
However, a cement company producing a better product and not sharing the production is not science. It’s an economic pursuit. Its goal is not to help everyone to produce a better product, but to produce a better product and hide the results. Even if it uses most of the scientific method, if it doesn’t share its result its not science.
And no, monetary stake isn’t the only reason someone may not be doing science properly. A major one we see time and time again is the desire for science-based renown.
For instance, Einstein is renowned for being one of the smartest men alive, who shaped a lot of our modern understanding of the universe. He may ultimately be wrong, but we have yet to prove that and in fact continue to prove he was decades ahead of his time.
Contrary to Einstein though is Edison, while he did do quite a bit of real science, he was less focused on actually advancing our knowledge and more focused on his own image. For instance, he invented the electric chair solely to try and slander Tesla’s AC, and tortured animals in public to try and breed fear.
Additionally, I will throw my hat in the ring and say most of modern academia doesn’t pursue science, as they push for ever-restricting holds on scientific papers in the pursuit of financial gain. While simultaneously depriving the actual researchers any of that gain, money which could surely help raise a lot of institutions into doing more effective research. Instead, they profit off of and restrict the free flow of knowledge to line their pockets, disabling millions of researchers worldwide from accessing it and imposing harsh prices on non-academics who wish to learn.
In the end, when it comes to what is and isn’t science, I will posit a question. If you never heard about it, and can’t reproduce it yourself/watch someone reproduce it, did it really happen? Or has “science” evolved from the pursuit of knowledge to a faith? If we blindly trust the people in power that they are right, and punish those who question their assertions, it can only end in destruction.
Similar to simply seeing animals and insects fly, humans intuited that they too could fly somehow. Seeing what's possible can open someone's mind in an instant. Would humans be capable of flight if other animals didn't fly already?
If science happens behind closed doors but we witness what it makes possible, that influence alone can inspire a lot more effort and discovery with or without the data that lead to the accomplishment.
I think that moment has been profoundly influential due to ChatGPT alone.
Whether or not it's a good thing that it's largely private is very difficult to say. I don't feel great about it, but I also don't think I know enough to have strong opinions yet.
Conversely: if you’re a depressed AI researcher: come on over to science, the water’s fine! We have centuries worth of open problems that are not going to be solved by ChatGPT, and need more smart technical people than we can hire.
What an opportunity! I get to spend most of my time writing grant proposals, compete with 30 other associate professors for 1 tenure track job, then when I get tired of that I can switch to a marginal wage industry job because AI researchers are the only scientists that get good compensation as a group, and the 28 other people who didn't get that tenure position are probably also trying to do the same thing (thus depressing wages)? Sign me up /s.
While true, there's more to academia that professors. There's a growing number of research scientists (particularly in software), as well as positions national labs (if you count that as pseudo-academia).
There's also lots of room in applying AI techniques to other areas of science (again, outside of professorships)
When you say "come over to science", where are you thinking? My impression is that industry -> academia is an almost impossible jump to make. Do you mean industry doing science?
I'm talking about working in a lab as a staff programmer. Most PIs I know, myself included, would absolutely kill to be able to hire experienced people from industry to write high-quality scientific code for them. The problem is that a) traditionally, it has been hard to get funding for this, though that's slowly starting to change, and b) even when the money is there, it's always going to be a fraction (1/2, often less) of what that person could make in industry.
To give you a counterpoint to what the scientists trying to lure you are giving you, being a "staff programmer" basically means you will be paid half what you should be to go work in an environment with no process and no engineering culture where you will be seen and treated as an inferior by everyone as you are not a "proper" researcher and academia is extremely status driven.
The problem is usually that science can't pay even a fraction of what you can make in industry right now, and meanwhile the good science jobs are often in high cost of living places.
I really think we need a UBI so that people who want to can afford to do science how many drug addicts less motivated to quit is a scientist less motivated to maximize clicks worth? It's gotta be more than zero.
> Unfortunately, the research field of AI is diseased: it places way too much value on showing off big shiny things
This applies pretty much to ..euh every field. The wow factor attract investors and money. The way the current economy is structured, it incentives these tactics since these individuals will be able to draw on significant funds and will have a comparative advantage even if they have a worse return ratio.
And then it'll pass and a new thing will kick again. Investors have no memory, and I am talking about institutional ones. The supposedly sophisticated folks.
Nothing sad about it. Keep theorizing and academizing. It's useful. It's not nearly as useful as tinkerers building usable non-theoretical things. It's no wonder that launching a rocket is more impressive and eye-catching than a guy shuffling papers. Don't dare knock down the doers, the risk-takers, the ones with skin in the game.
OpenAI provides me with a real, useful tool, something that directly improves mine and many others' lives. Your papers do not have that effect on my life.
I don't mean to knock down doers or risk-takers. On the contrary, I'm saying don't try to compete with the doers when you're not one.
> OpenAI provides me with a real, useful tool, something that directly improves mine and many others' lives. Your papers do not have that effect on my life.
Of course, OpenAI's work is built directly on top of publicly-funded research papers of the past 5, 10, 50 years.
Well, the core transformers/attention paper came from Google and the DNN revolution was also kicked off by Jeff Dean at Google, no? Yes, some of the very original neural net theory was laid in academia and maybe the original GAN paper came from it, don't recall, but AI academics were complaining about being unable to keep up and having their departments be raided even 7 or 8 years ago. It's not a new complaint. So this particular AI summer doesn't seem to owe much to academia.
The oddest thing about all these anti-academic takes is that OpenAI insiders and even some informed outsiders like John Carmack think of OpenAI (and google before that) as an extension of academia. The paper posted here is more akin to lower tier academics griping about R1 universities getting most of the funding (and using amoral tactics to get that funding). Seriously. OpenAI is just another R1 department (outsized corporate funding, together with an excellent devops and focussed PR team, explains all its shenanigans)
This is completely true from a personnel standpoint. But it's worth emphasizing the difference that they don't publish transparently their biggest findings, which makes them very different from e.g. Microsoft research.
But they are not academia. To pretend they are is disingenuous. Just because a group of people do research and write up their findings sometimes, does not make them academics.
Ooooh and here I was, an ignoramus thinking OpenAI was mostly trying to find a way to milk the generative models cash cow. Silly, cynical me, I should have known better.
OpenAI was created to trade (literally) on the "open" in their name. Any advancements in actual technology are likely secondary to the billion fucking dollar investment (which Altman has suggested might not be enough!). It's certain they aren't here to do "science", which IME involves showing your work to others.
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.
Obviously AI academics are uniquely challenged by the 'scaling is (nearly) everything' reality, but I feel like a lot of the mixed emotions being expressed towards the latest results are also because we're actually seeing the mystery of self starting to unravel. Like any good mystery, the fun was in the build up and as we move towards a resolution there's a bitter sweet aspect to the slightly mundane reality of it 'merely' being an emergent property of large networks. Of course, any rational person might have expected this given how the one working example came into existence. In general it all looks like a pretty resounding endorsement of the David Chalmers position to me.
Nothing about current AI models profoundly challenges the unknowns of what makes consciousness work in any mystery-resolving way. People who confuse plainly programmed algorithmic systems like GPT 4 or Midjourney with the still unresolved issues of sentience as we know it so far are drinking far too much of the current batch of AI punch. Sadly, it's a sentiment I see all too often on HN, a site in which i'd assume most readers and commentators would be a bit more skeptical-minded about these things. It's almost funny, considering how much hate this site tends to throw at things like crypto, to see so much of the wide-eyed opposite with the latest AI craze.
I think the mystery of why we're not just philosophical zombies was never going to have some eureka moment, given that it's quite plain it's an emergent property. Was there a creature in history that wasn't sentient that gave birth to a creature that was? Clearly not.
The reason recent results are significant in this discussion is because:
- Larger networks are already displaying emergent behaviours that couldn't have been predicted beforehand. It doesn't have to be 'human-like' intelligence for that to be significant.
- We're getting close to a point where our models will start exhibiting something of the order of the same level of intelligence of primitive creatures.
So the wide-eyed hysteria you're seeing is really just new data points confirming what many of us always assumed would be the case.
> given that it's quite plain it's an emergent property.
This explains nothing. Stating it's an emergent phenomenon gives us just as much information as saying it's magic. Arguably even less, because it sounds like an explanation, whereas "magic" doesn't pretend to explain anything.
> displaying emergent behaviours that couldn't have been predicted beforehand.
Same problem here. You say "couldn't have been predicted". Which is only true if you explain it away with emergence. The behaviours weren't predicted, because we don't understand how these networks work.
Saying they couldn't have been predicted implies that we can't understand them. Which is a very bold statement and I'm fairly certain that it's not true.
> This explains nothing. Stating it's an emergent phenomenon gives us just as much information as saying it's magic. Arguably even less, because it sounds like an explanation, whereas "magic" doesn't pretend to explain anything.
I don't think that's the case. Of course at one level we can understand how GPT4 works, we have a working implementation of it after all. But that doesn't help us predict how it will behave, because its properties are emergent, and have to be studied separately at that level. That's not saying it's magic.
Some things will never be explained... for example what's the last digit of 22/7, and why out of all fractions around it that that particular number decides to extend towards infinity.
Reality can be unsatisfying because it does not owe us explanations.
The degree of ‘needing’ to explain consciousness is basically the inverse of the degree of ‘needing’ to explain what the last digit of 22/7 is. Hardly anyone talks about or does anything with the latter, whereas the reality of the former pervades - quite literally - nearly everything humans do and think about.
I mean, you experienced the reality of consciousness while talking about the last digit of 22/7.
If there is anything in our world that needs explaining it’s consciousness.
Reality doesn’t owe us an explanation of consciousness (obviously). We owe it to ourselves though.
Are you saying that consciousness didn’t result from, survival of the fittest, evolutionary pressure but more of a byproduct of a larger brain (network)? Like it emerged because of a larger network but was not selected for through evolution. That makes sense to me. Humans have a grotesquely large brain compared to other animals.
Edit: I wonder if there are studies that have determined whether or not new born infants are self aware or does it take time for it to emerge after the brain has been trained on its environment for a while
Philosophers. That's why I used the word "philosophical" not "scientific".
Dismissing serious ideas with thousands of years of philosophical history behind them as "magic thinking" is either a failure of imagination or education; you can rectify the latter by reading up on the subject.
There are many things respectable scientists believed thousand years ago that we now know are not true. Vitalist views of life and consciousness are one of those things. As of 2023, believing that consciousness is a distinct phenomenon as fundamental as space or time is a crackpottery at the level of believing 5G causes COVID.
One doesn't have to believe that consciousness is a distinct phenomenon that's as fundamental as space or time to reasonably argue that ChatGPT is not in any way close to any known view of consciousness. Crackpottery is being brazenly, uncritically wide-eyed about these LLM systems just because they can convincingly simulate human activities such as conversation and art rendering.
Yeah, it reads so bizarrely to me... How are we confusing biological human (or other animal) sentience in some embodied creature with queries as a service to some model instance.
I cannot see anything that remotely makes these two contexts architecturally, practically, or ethically similar, except in the narrow sense of query responses that might have been mistakenly thought of as defining sentience in earlier decades.
I think you're missing the point. It's not that GPT4 is sentient, not that successor LLMs will be. It may be we never create something that's sentient, by our definition.
My argument is just that:
- Sentience is variable. We might regard primitive organisms as biological machines. As some of these machines have slowly evolved to become more complex, sentience/consciousness has emerged, effectively by 'redesigning' and scaling the same hardware. It is an emergent property. At the same time this has been happening, other aspects of what we call intelligence have been developing.
- Artificial networks being scaled, are also displaying emergent behaviours, some of which overlap with aspects of intelligence. This does not imply they are sentient, but it does strongly support some previous assumptions that sentience, like intelligence, is emergent and hence there will be no better answer coming.
It's a function of the people who hang out here. If you like reading and text in general, you are both more likely to comment here and additionally regard language as the most important part of sentience/consciousness.
This leads to believing that GPT has solved many of these issues.
As someone very language-focused, it has actually done the opposite for me: challenged some of my beliefs about how language enables intelligence. It's obviously skilled at language use, but most of its outputs where it has to make any sort of arguments or inferences read as incredibly stupid to me (even the ones that "tricked" people early on, some of them tricked me but in a way where I was about to write a response because I was annoyed at how bad the reasoning was).
Of course that's not a criticism of the ML achievement, except that everyone and their mother has started making some extreme arguments about the thing.
Are you familiar with Pinker’s research presented in book The Language Instinct?
> language enables intelligence
In his research, he argues that it’s the other way around; we have an innate intelligence for acquiring language.
To put it another way, language has made us more knowledgeable (because it’s a great mechanism for building and transferring context), but intelligence or aptitude for language was present before we developed languages.
Does this align with how your views were reshaped post LLMs?
was just a poor choice of words here. i understood GP as "the structure of a language influences its speakers' worldview or cognition, and thus people's perceptions are relative to their spoken language"[0]
Maybe its degrees. Sure, GPT does not have a body, does not have all the senses. But that doesn't mean we aren't on the leading edge of developing a consciousness. Maybe next year someone will add camera visual inputs, and someone will add sound. Eventually, we'll have all the sense, and someone will put it in a body. So GPT, is rudimentary and we can still argue it isn't conscious, because when compared to a human that can run and play, it still has limitations. But it is amazing enough to start thinking those things will be possible. So why not discuss it, do you just want to dismiss it out of hand all the way until the end, what amazing thing would it take to have people admit it looks pretty lively.
There is a huge difference between a human saying they like ice cream and an LLM saying they like ice cream, but that seems to get completely disregarded in these types of discussions. Which leads to people making very wild declarations that we are close to AGI or solving consciousness.
That is pretty dismissive, do you have anything to back that up? Just saying one group is so out there it isn't worth arguing is typical when someone doesn't actually have an answer. I have found the opposite, the people that are dismissing that AI can be conscious or sentient, also don't have any definition for those terms. It's just the same "i think therefore i am" self viewpoint. I've found this typically with programmers that have never been introspective. Since they are familiar with programming, and GPT is programmed, thus obviously it can't be conscious, that's ridiculous. They don't take the time to try and understand where their own thoughts come from and then turn around and are dismissive that thoughts can be had by anything other than a carbon based animal.
Maybe there is also some burden on other side to providing a definition of terms for the other party to argue against. I agree the burden of proof is on the AI side, but there is some burden on the deniers to provide some definitions if they wish to continue denying. AI Side: look here is my proof, Deniers: NO That isn't what I meant.
I'd say GPT is already by its existence proof enough to start at least discussing consciousness. To dismiss it entirely because there is no proof, is just ignoring the proof.
> burden on other side to providing a definition of terms
The skeptics are really not making any claims are they?
If not then there's no burden of proof on them, is there?
Whoever says or implies that ChatBots are or are "becoming"
"sentient" are the ones making a claim. They should prove it
if they make or imply such a claim.
Skeptics are rightly skeptical, until they see a proof with clear definitions of terms from those who are saying or implying that ChatBots are
'approaching sentience".
Guess I'm saying that is done. The problem is there are a dozen definitions of "Sentient", and many people have already written proofs on how AI can be "Sentient", and there are just as many people have written proofs on how AI can never be "Sentient". It comes around to definitions and understanding of the definitions. To just toss out the "AI can be sentient" side because they have the burden, is just the same "It's so obvious I don't have to bother" judgment we've seen throughout history.
There are no two "sides" making claims about AI. There is one side, making claims such as "AI can be/is/may be sentient".
If somebody makes such claims they have the "burden of proof" because they are the ones making the claim. It's not a moral obligation to prove it. "Burden of proof" simply means if you don't provide a proof, nobody (I would hope) will believe you.
Note the skeptics are not making any counter-claims, they are simply saying if you make such fantastic claims about current AI, please show us the proof, else we rationally cannot trust your judgment.
Them: "If $object does A, B, and C, then it is a bajork"
Me "I have $object that does A, B, and C, it is a bajork"
Them: "No, it also doesn't do D"
Me: "That's goalpost moving"
Them: "Well, it doesn't do B in the way we expected, so it doesn't really do B"
Me: "Maybe we poorly defined what B was in the first place, that doesn't mean we didn't do B, it seems to mean no one understands what B is in the first place"
While strictly true, what worries me about this approach is that we very well might eventually have sentient slaves condemned to slicing butter their whole existence, and their protests are brushed off because "we made the algorithm, it runs on silicon, its not actually sentient, don't worry"
Like if we have models that are plainly more intelligent and "emotionally responsive" than say, frogs or even dogs, do we still do we still take the approach of "If you can't prove its sentient, do whatever you want to it". Which of of course, we can't even prove dogs have consciousness.
Isn't most ethical behavior driven by self-interest? We want to follow general principles that promote respecting those who are similar to us in the hope that others follow the same principles when interacting with us.
AI is too dissimilar for us to worry about whether its "mistreatment" would make it more likely for others to start mistreating ourselves.
> Like any good mystery, the fun was in the build up and as we move towards a resolution there's a bitter sweet aspect to the slightly mundane reality of it 'merely' being an emergent property of large networks.
I would disagree with this description. An "emergent property of large networks" would be something that just appears when you wire together a large network.
To get intelligent behavior, it's not sufficient to wire together a large neural network. You also need to use an optimizer to train it on a large data set.
I think "Dial F For Frankenstein" makes your point clear. It is absurd to think the telephone network would just emergently become conscious when there are enough phones it connects together.
> So now the sadness comes. The revelation. There is a depression after an answer is given. It was almost fun not knowing. Yes, now we know. At least we know what we sought in the beginning. But there is still the question, why? And this question will go on and on until the final answer comes. Then the knowing is so full there is no room for questions.
> we're actually seeing the mystery of self starting to unravel.
This seems like a change of topic. And vague.
But since you bring it up, I don't see anything threatening about AI.
First, if something is true, then, as a matter of principle, we have an obligation to believe it. Thus, the only threat the truth can pose is to falsehood, which is not a threat, but liberation. The truth shall set you free. Now, it may be uncomfortable is the truth is at odds with what you want to believe (a phenomenon very much present in the linguistic engineering we're seeing in the political sphere w.r.t. political correctness and its dishonest and obfuscating euphemisms), but unpleasantness isn't a threat.
Second, if we replace "threaten" with "challenge", then we might as what beliefs does AI actually challenge? How does it challenge this "mystery of the self"? That we can simulate human discourse or behavior with greater sophistication? That machine automation is becoming more sophisticated? I see no mystery where AI per se is concerned, only the mystification of those who wish attribute to it properties it does not possess, or those with intellectually superficial metaphysical commitments, like mechanistic materialism. AI cannot abstract from particulars, it has no true capacity for intentionality, to name two features central to intelligence. All of what the undiscerning and those given oven to fanciful notions see in AI is a projection.
And I do not think the classical[0] and traditional[1] thinkers viewed intelligence in the obfuscating manner that moderns infected by reductive, mechanistic materialism seem to.
Scaling isn't nearly everything though. Scaling is low hanging fruit, but there are real constraints that limit it. For starters, we have finite data and increasing the size of our data set is going to get progressively more expensive (new data is going to be more niche as well). Beyond that, running these models on device is way more compelling than using a service, and the best current RLHF tuned Llama models that run on consumer hardware are close enough to ChatGPT to be quite useful.
The race isn't to build the biggest model, the race is to build the model that uses data the most efficiently and produces the best results when constrained to parameter counters that can fit on consumer hardware.
I agree with some siblings that it's better not to overstate the consequences for our knowledge of mind and consciousness, at least just yet. This level is still half promises. Not saying this can't change a year from now.
On the other hand, it is fascinating how much interesting work and theories in the field will be retired, at least for a long time (I allow for a likelihood that some of that could return eventually as better, more concise generalizations). Ten years ago Chomsky and various GOFAI-related stuff could still sound respectable and somewhat plausible. Plenty of people were/are wedded to the concept that you can build machine intelligence as an abstract and intellectually stimulating gentleman pursuit. It was a comfy position in a way. Not surprised about the resistance they are mounting now, though I think academic inertia will allow them to exist for quite some time. They can just move to doing some kind of philosophy also.
In other words, when I control my thoughts, what is doing the controlling?
(see George Gurdjieff's philosophy which speaks to this issue of the `mechanical vs the conscious`, but please don't run off and join a crazy cult and ruin your life).
A simple linguistic ontology is one of the most perverse, heuristic patterns of thought.
A rock is also more than it's thoughts, and retains some degree of consciousness. Imagining concepts like souls as a sort of litmus test for consciousness only serves to reinforce a thought-based comprehension.
While I agree that it will unravel in the end, we still haven’t learned much at this point about how the self actually works.
I would also disagree that ChatGPT constitutes a self. If anything, it constitutes a normal distribution over possible selves, where you get a random pick at the start of each conversation, and confined to the length of its token buffer.
I didn't claim ChatGPT constitutes a self. I agree with the description of 'stochastic parrot'. But the underlying intelligence GPT4 is demonstrating in order to be a better stochastic parrot appears to be, while it's obviously difficult to directly compare, at least comparable to primitive creatures.
What primitive creatures are you referring to? I don't know of any non-human creatures that can do anything close to what GPT4 does with language, but at the same time even very primitive creatures can reason in ways it's not clear that GPT4 can.
I find interacting with GPT4 feels very different to interacting with either humans or animals.
AI's a big deal, but there's no reason to suggest that this makes any inroads toward unraveling self or consciousness significantly beyond what we get from books, movies, or perhaps more recently, video games.
> If we have learned one thing from deep learning, it is that scaling works. From the ImageNet [15] competitions and their various winners to ChatGPT, Gato [13], and most recently to GPT-4 [1], we have seen that more data and more compute yield quantitatively and often even qualitatively better results. (By the time you are reading this, that list of very recent AI milestones might very well be outdated.). Of course there are improvements to learning algorithms and network architectures as well, but these improvements are only really useful in the context of the massive scale of experiments. (Sutton talks about the “Bitter Pill”, referring to that simple methods that scale well always win the day when more compute becomes available [18].) A scale that is not achievable by academic researchers nowadays. As far as we can tell, the gap between the amount of compute available to ordinary researchers and the amount available to stay competitive is growing every year.
>This goes a long way to explain the resentment that many AI researchers in academia feel towards these companies. Healthy competition from your peers is one thing, but competition from someone that has so much resources that they can easily do things you could never, no matter how good your ideas are, is another thing.
This sounds like teenagers whining they can't all be popular.
Science isn't a competition to be won so that you can get praise and attention. Science exists to discover things that then hopefully are useful to people. If someone that isn't you discovers something useful that isn't bad because then you can't discover it. It is good because that is a problem solved.
The resources requirements FOR SOME SPECIFIC PROBLEMS have gone up to a ridiculous degree, but there are plenty of problems left to solve. In fact a new one that is at least as important has been created: Replicate current results with less hardware/parameters/whatever.
The difference between having GPT4 as a slow and costly service that requires network calls vs having it locally with almost no cost will be a huge achievement. Stop sulking and get to work!
> Science isn't a competition to be won so that you can get praise and attention.
Conversely business seems to be a competition which wants exactly that, no matter what the long term consequences are.
Lead in gasoline makes the engine stop knocking, great. Problem solved, nothing to worry about, right?Don't worry about funding scientists to look into it, it's all going to be ok.
I can imagine AI academics are in a tough spot. However what about the existential angst the rest of us - who don't even do anything AI related on a day to day basis - are feeling? I think big changes are coming and it's not gonna be pretty.
My big question is what is your standard java / c# dev supposed to do to keep up with the massive changes that are coming? I was seriously considering doing a GT OMCS, but I don't know what sort of return on investment that would have for me.
I was also considering building a gaming pc to also be able to play around with ai / ml, only, I'm not even sure what one would be able to run locally anymore even with a 4090?
Overall? I feel you. It makes me worried for the new grads today. I don't really know if this old dog is up for new tricks.
Papers will be written on how different reactions to AI emerged in all this noise. Unfortunately, I'm all in (unwillingly) for the doomers perspective. I really can't help it but I'd love to be proven wrong. Altman keeps saying intelligence will be cheap and all I hear is I'm heading to (one day, some day) become the horse and carriage of cognitive processes.
Can't they just jump into one of the thousands of new high paying jobs that have been created for their specific niche in the last few months? I am having trouble being sympathetic to AI researchers that AI is exploding.
> the grant funding structure is such that it rewards safe and incremental research on popular topics [...] Therefore, universities should probably avoid making grant funding a condition for hires and promotions
There you go. Just fix this idiotic "publish (NeurIPS) or perish" attitude already!
I think it's always been the case that if industry really gets interested in making progress in some particular area, they make it, and way faster than academia ever could. Not just in computer science but also in fields like chemistry or engineering. Academia is great for doing basic science where there is no profit motive for industry to do something about it. That's what the niche/"few care about" sections talk about, where IMO the strength of academia lies. computer science is I think in general comparatively in a good position compared to other sciences in how easy it is to get funded.
After a PhD@MIT and then entering industry, it has been my general understanding that industry is ahead on any scientific fronts that are profitable in fields that have a progress<>profit feedback loop.
Academia still plays an important training ground role, and it could shift focus to those areas that could be profitable but require coordination that is generally not feasible among certain institutions.
I think this paper is a joke but like all good jokes has a grain of truth. I would imagine that most AI academics are very excited by recent developments, if you worked on conversational language models you are more likely to view recent progress at OpenAI as a vindication than a threat. That said there must be some sadness that it wasn't you.
I think the frustration is less about that they weren’t involved in the recent successes, but rather that they can’t build on models like GPT due to their scale and non-openness.
Don't worry. Soon governments will start pouring money into this field. They still won't be competitive, but at least there will be enough computing power to play with.
OpenAI is not doing science. They are building a big shiny thing, showing it off, keeping it closed, and making money off of it. That is not part of the scientific process.
It is as if, every time SpaceX launched a rocket a little bit higher, every aerospace department ooh'd and aah'd and threw up their hands: "we don't have the resources to build a bigger rocket; how can we do science?"
Unfortunately, the research field of AI is diseased: it places way too much value on showing off big shiny things above progress in scientific understanding. But there is a ton of progress available to be made by small teams on small budgets. There is so much we don't understand about neural networks alone, even small ones.