The AI Scaling Hypothesis
lastweekin.ai
lastweekin.ai
GPT-3 175B is trained with 499 Billion tokens[1]. Let's assume token = word for the sake of this argument[2]. The average adult person reads at a rate of 238wmp[3]. Then a human who reads 24hrs/day from birth until their 18th birthday would read a total of 2.2B billion words[4], or 0.45% of the words GPT-3 was trained on.
Human's simply do much more with much less. So what gives? I don't disagree that we still haven't reached the end of what scaling can do, but there is a creeping suspicion that we've gotten something fundamentally wrong on the way there.
1. https://lambdalabs.com/blog/demystifying-gpt-3/
2. GPT-based models use BPE and while we would dive into the actual dictionary of tokens and make a word-token relationship, we both agree that although this isn't a 1-to-1 relationship it won't change the conclusion https://huggingface.co/docs/transformers/tokenizer_summary
3. https://psyarxiv.com/xynwg/
4. 238*60*24*365*18=2,251,670,400
If you think about it, neutral networks have roamed the earth for millions of years - including generic algorithm for optimizing the hardware. And yet only extremely recently something like humans happened. Why?
The amount of training and processing power which happened naturally through evolution beats current AI research by several orders of magnitude. Yes, evolution isn't intelligent design. But the current approach to AI isn't intelligent design either.
GPT-3 is learning everything it knows about the entire universe just from text.
Imagine we received a 1TB information dump from a civilization that lives in an alternate universe with entirely different physics. How much could we learn just from this information dump?
And from our point of view, it could be absurdly exotic. Maybe their universe doesn't have gravity or electromagnetic radiation. Maybe the life forms in that universe spontaneously merge consciousnesses with other life forms and separate randomly, so whatever writing we have received is in a style that assumes the reader can effortlessly deduce that the author is actually a froth of many consciousnesses. And in the grand spectrum of how weird things could get, this "exotic" universe I have described is really basically identical to our own, because my imagination is limited.
Learning about a whole exotic universe from just an info dump is the task of GPT-3. For instance, tons of our writing takes for granted that solid objects don't pass through each other. I dropped the book. Where is the book? On the floor. Very few bits of GPT-3's training set includes the statements "a book is a solid object", "the floor is a solid object", "solid objects don't pass through each other", but it can infer this principle and others like it.
From this point of view, its shortcomings make a lot of sense. Some things GPT fails at are obvious to us having grown up in this universe. I imagine we're going to see an explosion of intelligence once researchers figure out how to feed AI systems large swaths of YouTube and such, because then they will have a much higher bandwidth way to learn about the universe and how things interact, connecting language to physical reality.
Paucity of the stimulus is a real problem and may mean our starting point architecture from genetics has a lot of learning built in than just a bunch of uninitialized weights randomly connected. A newborn animal can often get up and walk right away in many species.
https://www.youtube.com/watch?v=oTNA8vFUMEc
Humans have a giant head at birth and muscles too weak, but can swim around like little seals pretty quickly after birth.
And, as you say, there are probably some structural/architectural improvements to be made in the neural network as well. The mammalian brain has had a few hundred million years to evolve such a structure.
It also remains unclear how important learning causal influence is. These networks are essentially "locked in" from inception. They can only take the world in. Whereas animals actively probe and influence their world to learn causality.
Neural plasticity is the algorithm running on top of our neural networks that optimizes their structure as we learn so not only do we get more data, but our brains get better tailored to handle that kind of data. This process continues from birth to death and physical experimentation in youth is a key part of that development, as is social experimentation in social animals.
I think “it remains unclear” only to the ML field, from the perspective of neuroscientists, current neural networks aren’t even superficially at the complexity of axon-dendrite connections with ion channels and threshold potentials, let alone the whole system.
A family member’s doctoral thesis was on the potentiation of signals and based on my understanding if it, every neuron takes part in the process with its own “memory” of sorts and the potentiation she studied was just one tiny piece of the neural plasticity story. We’d need to turn every component in the hidden layers of a neural network into it’s own massive NN with its own memory to even begin to approach that kind of complexity.
I don't doubt that evolution provided us with great priors to help us be fast learners, but there are two more things to consider.
One is scale - the brain is still 10,000x more complex than large language models. We know that smaller models need more training data, thus our brain being many orders of magnitude larger than GPT-3 naturally learns faster.
The second is social embedding - we are not isolated, our environment is made of human beings, similarly an AI would need to be trained as part of human society, or even as part of an AI society, but not alone.
The six year old has the advantage of being immersed in a persistent world where images have continuity and don’t jump around randomly. For example infants learn very quickly that most objects stay put even when they aren’t being observed. In contrast a dataset of images on the internet doesn’t really demonstrate how the world works.
Drawing involves taking a mental image and converting it into a sequence of actions that replicate the image on a physical surface. Imagen does not do that. I think the images it generates are more analogous to the image a person creates in their mind before drawing something.
Effectively, this can broaden the context the network can learn. There are relationships that are readily apparent to something that learned images that might not be apparent to something trained only on text, or vis-versa.
It will be interesting to see where that goes. Will it be possible to make a singular multi-domain encoder, that can take a wide range of inputs and create an embedding (an "mental model" of the input), and have this one model be usable as the input for a wide variety of tasks? Can something trained on multi-domains learn new concepts faster than a network that is single-domain?
text2blueprint or wav2schematic could produce some interesting things.
Later refinements with AlphaZero and MuZero also reduced the driver size.
The most naive form is just Monte Carlo tree search using the neural net as the bias/eval function. A basic MCTS is just 100 lines of code or so.
Edit: Also you are missing the Go engine in that comment, it can't train without a Go engine to train against that evaluates the results of each move. That Go engine is a part of the training algorithm and thus is also a part of the driver code, you would need to produce something similar to train a similar AI for other domains. We don't know how to write similar blueprint engines or text evaluation engines, so we can't expect such AI models to produce similar results.
"The universe" according to most human experience misses SO much information and it will be interesting to see what happens once we have agents that can use all this extra stuff in realtime and "see" things we cannot.
Transformers themselves are a lot more effective compared to n-gram models or non-contextual word vectors. I imagine there is something to Transformers as Transformers are to word2vec.
True vision requires motive and embodied self. I’m ignorant about the state of the art here, but I’m way more terrified of what these things don’t see than interested in what they could show us. It seems to me that the only human motives accessible to machines are extremely superficial and behavioral based.
Knowledge is not some disconnected map of symbols that results in easily measurable behavior, it has a deep and fundamental relation to conscious and unconscious human motivation.
I don’t see any possible way to give a machine that same set of motives without having it go through our same evolutionary and cultural history, and strongly believe most of our true motives are under many protective layers of behavioral feints and tests and require voluntary connection and deep introspection to fractionally expose to our conscious selves, let alone others, let alone a computer.
These models seem to be amazingly good at combining maps of already travelled territory. Trying to use them to create maps for territory that is new to us seems incredibly dangerous.
Am I missing something here, or is it not true that AI models operate purely on bias? What we chose to measure and train the model on seems to predetermine the outcome, it’s not actually empirical because it can’t evaluate whether it’s predictions make sense outside of that model. At some point it’s always dependent on a human saying “success/fail”, and seems more like an incredibly complicated kaleidoscope. Maybe they can cause humans to see patterns we didn’t see before, but I don’t think it’s something that could actually make new discoveries on its own.
> We test a stronger hypothesis: that the conceptual representations learned by text only models are functionally equivalent (up to a linear transformation) to those learned by models trained on vision tasks. Specifically, we show that the image representations from vision models can be transferred as continuous prompts to frozen LMs by training only a single linear projection.
Linearly Mapping from Image to Text Space - https://arxiv.org/abs/2209.15162
People's brains are wired for language, it's natural we have a head-start.
1. hardwired for creating visual models of 3d / 2d space
2. hardwired for fine-grained hand motor movements.
So the learning is just combining those two skills into the ability to make the brush strokes needed to output the model onto paper.
Another issue is what we really care about is scientific reasoning and there, if anything, nature has given an anti-bias, at least at the level of interfacing with facts. People aren't born biased towards learning Metric Tensors and Christoffell Symbols but it takes only a few years at a handful of hours a day using a small number of joules for many humans to get it (I'm counting from all grade school prerequisites vs GPUs watts x time). Much fewer for genius children.
They don't. Human bitrates are quite low all things considered. The eyes which by far produce them most information only have a bitrate equivalent to ~2kbps:
http://www.princeton.edu/~wbialek/our_papers/ruyter+laughlin...
The rest of the input nerves don't bring us over 20kpbs.
The average image recognition system has access to more data and can tell the difference between a cat and a banana. A human has somewhat more capability than that.
If we play "the bus filled with ping-pongs" with that information: it is a 3D structure so if you assume cortex is a perfect cube that feeds to something right behind it, you will get (10^12)^(2 (dimensions)/3 (of 3 dimensions)) channels, e.g. 10^8 channels 2kbps each. E.g. about 25GB/s. Which is less than an order of magnitude off from an estimate you would get from 8000x8000 resolution per eye True Color at 24fps - 9GB/s.
My prediction is we’ll get ‘good enough for prod’ without interactive data, which will let us put interactive systems in the real world at scale, at which point the field’s focus will be able to shift.
One way to look at it is active learning. We all know the game where I think of a number between 0 and 100 and you have to guess it, and I’ll tell you if it’s higher or lower. You’ll start by guessing 50, then maybe 25, and so on, bisecting the intervals. If you want to get within +/-1 of the number I’m thinking of, you need about six data points. On the other hand, if you don’t do this interactively, and just gather a bunch of data before seeing any answers, to get within +/-1, and need 50 data points. The interactivity means you can refine your questions in response to whatever you’ve learned, saving huge amounts of time.
Another way to look at it is like randomized controlled trials. To learn a compact idea (more X means more Y), you can randomize X and gather just enough data on Y to be confident that the relationship isn’t a coincidence. The alternative (observational causal inference) is famously harder. You have to look at a bunch of X’s and Y’s, and also all the Z’s that might affect them, and then get enough data to be confident in this entire structure you’ve put together involving lots of variables.
The way ML has progressed is really a function of what’s easy. If you want a model to learn to speak english, do you want it to be embodied in the real world for two years with humans teaching it full time how the world and language relate? Or is it faster to just show it a terabyte of english?
tl;dr observational learning is much much harder than interactive learning, but we can scale observational learning in ways we can’t scale interactive learning.
While true on the surface, you have to also consider that there is a vast quantity of training data expressed in our DNA. Our 'self' is a conscious thought, sure, but it's also unconscious action and instinct, all of which is indirect lived experience of our forebear organisms. The ones that had a slightly better twitch response to the feel of an insect crawling on their arm were able to survive the incident, etc. Our 'lizard brains' are the result of the largest set of training data we could possibly imagine - the evolutionary history of life on earth.
The upshot is that once you've found an architecture that is already biased towards solving a specific problem, then the training of the weights is faster and results in better performance.
From the abstract, "...In this work, we question to what extent neural network architectures alone, without learning any weight parameters, can encode solutions for a given task.... We demonstrate that our method can find minimal neural network architectures that can perform several reinforcement learning tasks without weight training. On a supervised learning domain, we find network architectures that achieve much higher than chance accuracy on MNIST using random weights."
They are taking many thousands (millions?) of inputs every minute for their surroundings
It may be argued we need more human-like learning mechanisms. Then again, if we need internet-scale data to achieve human-level general intelligence, so what? If it works it works. Of course, the comparison has some value in terms of knowing what can be improved and so on, especially for RL. But I wouldn't call this a 'troubling problem'.
Of course there are differences between the systems, but I'm beginning to be more skeptical that saying that the newer ML system can't learn as much as biological systems given the same input (obviously this is where a lot is hidden).
I'm definitely in agreement that multi-task models represent an ability to learn more than any one specialized model, but I think it's a bit of an open question whether multi-task learning alone can fully close the digital-biological gap. Of course I'd be very happy to be proven wrong on this though by empirical evidence in my lifetime :)
With vision, pointing to Parti and DALLE as scaling is quite dumb. They perform similarly but are DRASTICALLY different in size. Parti has configurations with 350M, 750M, 3B, and 20B parameters. DALLE has 3.5. Imagen uses T5-XXL which alone has 11B parameters, just in the text part.
Not only this, there are major architecture changes. If scaling was all you needed then all these networks would still be using CNNs. But we shifted to transformers. THEN we have shifted to diffusion based models. Not to mention that Parti, DALLE, and Parti have different architectures. It isn't just about scale. Architecture matters here.
And to address concerns, diffusion (invented decades ago) didn't work because we just scaled it up. It worked because of engineering. It was largely ignored previously because no one got it to work better than GANs. I think this lesson should really stand out. That we need to consider the advantages and disadvantages of different architectures and learn how to make ALL of them work effectively. In that manner we can combine them to work in ideal ways. Even Le Cun is coming to this point of view despite previously being on the scaling side.
But maybe you NLP folks disagree. But the experience in vision is far more rich than just scaling.
from wikipedia (https://en.wikipedia.org/wiki/DALL-E):
DALL-E's model is a multimodal implementation of GPT-3 with 12 billion parameters which "swaps text for pixels", trained on text-image pairs from the Internet. DALL-E 2 uses 3.5 billion parameters, a smaller number than its predecessor.
At the other extreme, some recent works [1,2] show why it’s sometime better to scale down instead of up, especially for some humanlike capabilities like generalization:
[1] https://direct.mit.edu/tacl/article/doi/10.1162/tacl_a_00489...
From a value perspective a very high fidelity model with extremely unexpected behavior seems really low value since you need a human there full time to make sure that the model doesn't go haywire that 1-5% of the time
Therefore, the model uses other information than humans in order to classify, and the information it uses is wrong.
Therefore, the model needs to be trained ON THE ADVERSARIAL EXAMPLES in order to gain robustness.
Similarly to GANs using a classifier adversary for augmenting a generative network, one could use a generative adversary for augmenting a classifier network.
You can repeat until the adversarial examples look ambiguous even to a human.
For example, my work in grad school is deploying AI-based control on drones such that they are fault-tolerant. We can have a stronger model at the ground station communicating with the vehicle, which introduces lag. Or the ground station model is more supervisory, and bare metal control is deferred to your classical algorithms. Or we run a smaller, faster network(s) on board which can act and learn fast.
The other thing is model interpretability. For some of my work with drones or smart buildings, the stakeholders wouldn't let a neural network run on a real vehicle or building until it was thoroughly certified through whatever criteria they had. Easier with a smaller network.
However the core also seems to be training data. SEER from meta/Facebook is a billion images, far bigger than imagenet. OpenAI doesn’t release its training dataset for whisper.
It seems we are in a place where AI will most likely increase the digital divide and wealth gap because the big datasets + large model training ability will only be accessible to mega corps.
This is not a path to "AGI", this is just building a search engine with a little better querying power.
"AI" systems today are little more than superpositions of google search results, with their parameters being a compression of billions of images/documents.
This isnt even on the road to intelligence, let alone an instance of it. "General intelligence" does not solve problems by induction over billions of examples of their prior solutions.
And exponential scaling in the amount of such remembering required is a fatal trajectory for AI, and likewise and indication that it doesnt deserve the term.
No intelligence is exponential in an answer-space, indeed, i'd say that's *the whole point* of intelligence!
We already know that if you compress all possible {(Question, Answer)} pairs, you can "solve" any problem trivially.
> Parameters are just the database of the system.
Would any equations parameters be considered just the database then? C in E=MC^2, 2 in a^2+b^2=c^2?
I suppose those numbers are basically a database, but the relationships (connections) they have to the other variables (inputs) represent a demonstrable truth about the universe.
To some degree every parameter in a nn is also representing some truth about the universe. How general and compact that representation is currently is not known (likely less than we'd like of both traits).
Also, the C in E=MC^2 has units which define what it means in physical terms. How can you define a "unit" for a neural network's output?
Now, my thoughts on this are contrary to what I've said so far. Even though neural network outputs aren't easily defined currently, there's some experimental results showing neurons in neural networks demonstrating symbolic-like higher-level behavior:
https://openai.com/blog/multimodal-neurons/
Part of the confusion likely comes from how neural networks represent information -- often by superimposing multiple different representations. A very nice paper from Anthropic and Harvard delved into this recently:
"Connectionists" always want to reduce everything to formulae with no natural semantics and then equivocate this with science. Science isnt mathematics. Mathematics is just a short hand for a description of the world made true by the semantics of that description.
E=mc^2 isnt true because it's a polynomial, and it doesnt mean a polynomial, and it doesnt have "polynomial properties" because it isnt about mathematics. It's about the world.
E stands for energy, m for mass, and c for a geometric constant of spacetime. If they were to stand for other properties of the world, in general, the formulae would be false.
I find this "connectionist supernaturalism" about mathematics deeply irritating, it has all the hubris and numerology of religions but wandering around in a stolen lab coat. Hence the tone.
What can one say or feel in the face of the overtaking of science by pseudoscience? It seems plausible to say now, today, more pseudoscientific papers are written than scientific ones. A generation of researchers are doing little more than analysing ink-blot patterns and calling them "models".
The insistence, without explanation, that this is a reasonable activity pushes one past tolerance on these matters. It's exasperating... from psychometrics to AI, the whole world of intellectual life has been taken over by a pseudoscientific analysis of non-experimental post-hoc datasets.
From [1], with my own editing...
When comparing the difference between now and human performance
> ...[huamns] can achieve closer to 0.7 bits per character . What is in that missing >0.4?
> Well—everything! Everything that the model misses. While just babbling random words was good enough at the beginning, at the end, it needs to be able to reason our way through the most difficult textual scenarios requiring causality or commonsense reasoning... every time that it lacks the theory of mind to compress novel scenes describing the Machiavellian scheming of a dozen individuals at dinner jockeying for power as they talk...
> If we trained a model which reached that loss of <0.7, which could predict text indistinguishable from a human, whether in a dialogue ...how could we say that it doesn’t truly understand everything?
There's an a priori "limit" on the form of any time/space complexity "intelligence" can have: it cannot be exponential.
We know that if you exhaust all infinite number of ways of describing a system by facts (/data), then you don't need intelligence. Whatever intelligence is, it's a strategy animals use for competent action in the face of their bodies not being able to exponentially expend energy to solve every trivial problem.
Exponential complexity is a killer to the project of building intelligence in every respect. The universe, a priori, cannot require an exponential space complexity description of it to be parsable: since that would require exponentially many universes.
Exponential scaling, for me (via ordinary conceptual analysis), seems a hallmark of any system not using intelligence. Exp. scaling is "the shortcut" to solving the problem without intelligence. It's the hallmark of the system not building O(log_{large number}) rich representations of the world.
At the crudest level this immediately provides for more than canned lookup as asserted; analogical reasoning is a much-documented emergent property.
But analogies are merely the simplest, first-order abstraction, which are easy for humans to point at.
Inference and abstraction across multiple levels, means the behaviors of these systems is utterly unlike simple stores. One clear demonstration of this is the effective "compression" of image gen networks. They don't compress images. For lack of any better vocabulary, they understand them well enough produce them.
The hot topic is precisely whether there are boundaries to what sorts of implicit reasoning can occur through scale, and, what other architectures need to be present to effect agency and planning of the kind hacked at in traditional symbolic systems AI.
It might be worthwhile to read contemporary work to get up to speed. Things are already a lot weirder than we have had time to internalize.
The problem is not that really what they are doing, which is more akin to what e.g our own visual processing stack does...
...it's that we don't have language for a lot of the territory these systems operate in.
We are used to a world defined by very clear boundaries between agents (which e.g. have state, and agency, and carry world representations which we can infer and reason about and behave against) and tools (potentially complex but inert, at best narrowly-constrained),
these exist in the in-between where until now we only really had "animals" or maybe "institutions" but that is more of an abstract analogy.
Like say in insect, they understand but poorly—compared to us as the gold standard. And we have silos of deep specialization (language) and world knowledge beyond the purely visual. Eventually it's clear there will be a knitting together of systems which provide those, as we cobble together something that "understands" more than just the visual.
But already when you look closer it's clear they already understand some things as well as we do, or better.
Pareidolia is a fabulous word and frame, I've used the behavior of these models also to explain what hallucinations are like (and are) when on psychedelics to people who haven't experienced them.
Anyway. What does a toddler understand? Or a dog? Or...
On the bright side I am seeing language come into use before our eyes. The most established and interesting one to emerge in common use this year IMO is the use of "latent" to describe that which these things understand... they render that which is latent in their complex encodings. A great adaptation of a domain specific term of art into lay use.
Fascinating times to be sure.
We don't know if your database analogy is also true for the human brain.
If it isn't, then you have a point. If it is, then you probably don't.
The AI stalks the convicted and finds ways to cause suffering, but learning from performing it against thousands of people, many simultaneously and even adversarially, optimizing for the most reflected cruelty it interprets from posts it gauges sentiment of in its environment. It's a plausible reason for why Roko's Basilisk would be made to exist, as we would have to think it was something we were inflicting on the other or another. Criminal punishment via AI as a means to wash our hands of our judgments seems like just the sort of thing we might sadly do. Surely it will only be reserved for the most serious of criminals.
The effects of white torture are arguably worse, but they let us pretend to be less barbaric.