Large language models: Scaling laws and emergent properties
cthiriet.com
cthiriet.com
"... has discovered empirically that the capability for moral self-correction emerges at around 22B parameters. For more scientific details, the paper is here. [https://arxiv.org/pdf/2302.07459.pdf]"
This would suggest that moral philosophy could be transformed - in much the same way that natural philosophy turned into physics.
I’m just a layperson and don’t really know what I’m talking about, but I find this all very exciting.
I don't think they're "dangerous" per se, I think metaphors matter and we should choose the best ones.
So we can actually look at concepts like “morality” and see how that is encoded. And I’m confident that this will give us empirical insight into these concepts in a way no philosophy has been able to do until now.
What AI might enable however is a superior form of democratic process, wherein an AI surveys the entire population through a natural language interface and synthesizes the nuanced and conflicting desires of the entire society. This deliberative democracy process would ameliorate the distorting effects of campaign funding and issues like uninformed or misinformed voters and low voter participation. It could also allow a sort of citizen feedback line-by-line on proposed legislation and government action.
Why would something being subjective imply it is meaningless?
This part seems unsubstantiated. The author implies that next-token-prediction does not lead to reasoning and creativity "of an animal level". But arguably GPT-4 is already at that level (can your dog predict what a novel Python function would output when it is executed, or write a rap song about autoregressive training where every word starts with the letter "a"¹?).
Regardless of whether GPT-4 is there already, it would certainly be a risky bet to say that it is impossible for a GPT-5 to get to "animal level" (whatever that means) reasoning and creativity while still using next-token-prediction.
As the article points out in the next paragraphs, there is no obvious limitation to next-token-prediction, because making the best prediction requires fully understanding the words in the context, as well as their true meaning and the reason they were written.
Related to creativity, here's a Midjourney image resulting from my spouse prompting "a studious corn snake": https://chris.printf.net/a_studious_corn_snake.png
¹: via GPT-4:
==
(Verse 1) Autoregressive, arranging, assessing answers, Adjacent activations, astounding, acting as anchors, Associating abstracts, analyzing antecedents, Aptly approximating, articulating, all adjacent aspects
(Chorus) Astonishing achievements, astounding accolades, Anomalous arrangements, ascending always, Audacious algorithms, autoregressive amaze, Applying ample artistry, as alliteration ablaze
(Verse 2) Accelerating acumen, anonymous aspirations, Anecdotal adventures, allegorical animations, Applied apparatus, assertive applications, Astral alliterations, assembling aggregations
(Chorus)
(Outro) Amplified abilities, augmenting atmosphere Adapting, acclimating, always aiming austere Autoregressive training, ascending, advocates Awe-inspiring artistry, as alliteration activates
There really is no arguably about it. There isn't a verifiable definition/criteria for general intelligence that GPT-4 fails that a significant chunk of the human population doesn't also fail.
GPT-4 performs nearly all tasks it's given at at least the average human level, usually well above the average baseline. In some cases (nlp, analogical reasoning - https://arxiv.org/abs/2212.09196, and some others it's top %)
You basically have to make up your own definition that can't be substantiated (don't worry, plenty seem to do this). It almost always boils down "Well it's not "real" understanding, the difference is huge, i can't show you this supposed huge difference but trust me bro, it's not real"
yet, you will afraid to trust gpt-4 to file your taxes through installed browser plugin.
There is a huge gap between some narrow cherry picked (researchers have strong incentive to report mind boggling results) and diverse and multi modal real life situations
I don't need to cherry pick, I use GPT4 daily. At this point naysayers can only be willfully ignorant or uncreative in their use of it.
It's not possible for me to believe it is anything else.
you should try to work on your limitations..
The systems are very impressive and I think there is a 'there' there, they're clearly intelligent to some significant degree. But it's incorrect to say that they pass every test.
It scored 3/4, incorrectly answering "the pouch" instead of "the baby kangaroo" for Question 2 of Schema 1. And now that I look at the question, one could easily imagine the kangaroo being afraid to jump into a pouch that's too large! In hindsight, if a human complained to me that this question was ambiguous, I'd have a hard time disputing them.
Granted these are just two examples, but they aren't cherry-picked. Moral of the story: before one makes a strong claim such as "there isn't any evidence a model can do X", it's often worth making a minimal effort to investigate one's own claim.
========================================
Winograd Schema 1:
Sentence 1: The baby kangaroo was afraid to jump into the pouch because it was too small. Question 1: What was too small? (Correct answer: the pouch)
Sentence 2: The baby kangaroo was afraid to jump into the pouch because it was too large. Question 2: What was too large? (Correct answer: the baby kangaroo)
Winograd Schema 2:
Sentence 1: The gardener couldn't grow the flowers in the greenhouse because it was too hot. Question 1: What was too hot? (Correct answer: the greenhouse)
Sentence 2: The gardener couldn't grow the flowers in the greenhouse because they were too delicate. Question 2: What were too delicate? (Correct answer: the flowers)
Summing up, out of your four (not an acceptable sample size to form evidence) attempts to subject a SOTA LLM to Winograd schemas, two of them were actually Winograd schemas and only one of those got answered correctly. That isn't evidence that LLMs can pass novel Winograd schemas. If you actually come across such evidence, you should post it publicly on at least a blog page if you can't access the infrastructure to publish the results in a paper, even though the finding would certainly be novel enough to deserve a paper. As I said: there is no evidence that LLMs can pass novel Winograd schemas as of yet. We could get said evidence tomorrow, but it doesn't exist right now.
I would put a condescending moral of the story here, but I don't feel any particular need to belittle you, and instead hope you just learn from the information as presented.
Do you think this was a strong refutation ? I just flipped a coin for fun and also got 3/4 correct.
I understand your point, but I mean, something that's intelligent really should've got the second question in that test correct.
Interesting! So there is published evidence, that it cannot pass novel Winograd schemas? Could you please provide me with the sources? I would like to have a look at the prompts. From my experience you can make GPT-4 pass novel Winograd Tests by giving GPT-4 additional context.
>So there is published evidence, that it cannot pass novel Winograd schemas?
There is no published evidence that they can, which is what I said.
>Could you please provide me with the sources?
I cannot prove a negative to you with a source. You will have to review the literature like I did.
>From my experience you can make GPT-4 pass novel Winograd Tests by giving GPT-4 additional context.
This is why I am pretty sure you're trolling. This is a lot like saying "You can make GPT-4 do prime factorisation of a semiprime by noting the constituent primes." The point of testing the model is to see if the model can give you the answer, not to give it the answer. The point of Winograd schemas as a test is that they test the ability to attribute pronouns using common sense, without additional context. The reason it's a good test is that every competent human being can pass the test, and as of yet no computer can.
> This is why I am pretty sure you're trolling. This is a lot like saying "You can make GPT-4 do prime factorisation of a semiprime by noting the constituent primes." The point of testing the model is to see if the model can give you the answer, not to give it the answer.
It is not "giving the answer to the model" if you adjust the prompt with additional instructions like "this is a test question" or "use common sense". It is about creating the correct frame of context. With these kind of adjustments GPT-4 passes every novel Winograd scheme I have presented it with.
The usual broad indicia of intelligence (which, e.g., IQ is considered as a useful proxy measure for in humans because it predicts/correlates with them) are lifelong success in a wide variety of tasks GPT-4 can’t even begin to approach.
Right. and what are these tasks and why is intelligence the barrier? (meaning don't tell me something like driving, blind people are not unintelligent)
This sounds quite similar to Bing famously explaining to a user that 2022 is in the future while we are in 2023.
It's something you would expect maybe a child to do (it reminds of "Patrick Star's Wallet" meme), but to me that clearly fails the notion of intelligence in a very fundamental way. What would be the counter-argument?
Sentences look something like this to the model, [754 76 4752 67 456]
If I try to reformulate your answer, this is like words are a different domain than what GPT has been trained on (which is tokens), and so it is not unexpected that GPT cannot generalize to this other domain.
Is that a correct interpretation of your statement?
In my view failing to see the similarity in the relationship between class objects is a sign of a lack of 'intelligence', whatever that may be in a very precise sense.
Ok but not all relationship similarities transfer across representations like that.
It's like you give an instruction in English, "Find a word that rhymes with ingenious" to a french man. The french man sees, "Trouver un mot qui rime avec ingénieux". Guess what ? His response to "Trouver un mot qui rime avec ingénieux" would not fulfill the request in english that it might in French.
Ingenious and heterogeneous rhyme. ingénieux and hétérogène have no such property
The french man is not unintelligent to be unable to fulfill your request.
Analog computers also don't have infinite states, because of the noise floor.
A synapse can encode at most some low tens of states[1], synapses are discrete, spiking is discrete, receptor density is discrete, each neurotransmitter release is made up of discrete number of vesicles with a discrete number of molecules, even epigenetic expression is discrete. That it's not all perfectly algorithmically executed is not key to the function being performed – and anyway, at the limits of high-performance compute we also start to deal with esoteric processes, and ways of suppressing them. We didn't design the brain nor do we have perfect ways of observing it in vivo, so we can't always neatly distinguish the wheat and the chaff; but this cannot be seriously taken as a cause to think it's all wheat. Non-human animals show that it's not magic either; even apes, with brains architecturally almost (not wholly) identical to ours but smaller, cannot learn any of the hard symbol-manipulating stuff that we or later GPTs can and that we colloquially refer to as intelligent behavior and reward economically. This, I think, is a very strong prior for that sort of intelligence not being dependent on some low-level expressivity of biological neural computational substrate.
Penrose-Hamerroff's "quantum effects in microtubules" thesis is crank science not supported by any direct evidence.
==
Me: Is it true that, in humans, one neuron has an infinite number of possible states (as far as we know)?
GPT-4: It is not accurate to say that a single neuron in humans has an infinite number of possible states. Neurons are complex cells that can exhibit a wide range of behaviors and responses, but the number of states is not infinite.
Neurons communicate with each other through electrical and chemical signals, and their primary functional unit is the synapse, where communication between neurons occurs. The strength and patterns of these connections can vary, resulting in a large number of potential states for a single neuron. However, this number is still finite, albeit extremely large.
Moreover, the activity of a neuron can be described in terms of its membrane potential, which is a continuous variable ranging from hyperpolarized to depolarized states. While there are many possible values for the membrane potential, it is still a bounded range, not infinite.
So, while the number of possible states for a single neuron is vast and complex, it is not infinite. The complexity and diversity of neuronal states contribute to the rich repertoire of cognitive and behavioral functions that humans possess.
(Would it bother you to be corrected by a lookup table?)
https://phys.org/news/2014-01-discovery-quantum-vibrations-m...
Neurons are not of unbounded size nor energy.
The size of the earth's biosphere
The size of the largest animal.
The size of a human skull cavity.
The size of a human skull cavity, divided by the number of neurons therein ( a few Billion).
Is this not obvious? There's nothing futile about stating these obvious facts.
Do not fool around with "if I move a potassium ion farther away, isn't that a different bound" ? correct, but moving the bound slightly, is irrelevant to the obvious statement that there are still bounds, and they are practical.
Stating that the neuron is not well-defined is similar irrelevance, if you want to precisely measure the bound, you will need to know precisely what is bounded, but I did not, merely stated the obvious truth that every such definition will come with a bound. One can say "definitely smaller than this" in a wide range of ways, without defining the exact edge.
What's the alternate argument: "I won't define it, therefor it might be larger than the solar system". This is not a serious thing to say.
And to think that such bounded systems "have infinite states would thereby be capable of storing infinite information" is pseudo-mystical foolery. Large, or even "vast", is not infinite. And as sibling comment has mentioned, there is a "noise floor" below which analog resolution means nothing.
But you seem to be suggesting that all animal brains are equally powerful, because they have infinite neuron states, quantum superpositioned neurons, etc.
As far as I can tell, you seem to be suggesting that my old cat’s brain has the same complexity as my brain - right?
So, I’m interested in understanding why my cat is demonstrably stupider than I am, despite having infinite complexity, and why whatevrr you come up with doesn’t also apply to LLM size.
But of course it does imply that. It's vast but necessarily finite number of "brain states" however that's defined, which is itself a subset of the vast but also finite number of states of the finite number of component particles.
That's the minimum. It's far above average for most things.
>Maybe, because quite a lot of people aren’t very smart.
Lol okay..and ?
>But even GPT4 struggles with complex logic problems
So does everyone else
>and when it makes logic errors it makes ones a human would spot
people spot mistakes they're capable of making all the time
and so can GPT-4 - https://arxiv.org/abs/2303.11366 https://twitter.com/johnjnay/status/1639362071807549446
>It also has trouble counting letters and words, though that’s largely due to its autoregressive nature.
It sees tokens not words or letters and there aren't many instances of counting occurring in text so it doesn't get learnt.
>and I really think it’s emergent, but at its core it doesn’t seem very “intelligent”
and we're back to the trust me bro argument. It's not intelligent because you say it isn't not because it doesn't perform a great many tasks intelligently. if you say so
(No, I don't know why.)
E.g. here, when I say modular -- I mean that a "fix" can't be implemented in a local, small way. I need to try again with a new prompt, etc -- and it's not clear to me how to "fix" my prompt to get the right thing.
one way to support this sort of modularity might be with a GPT-4 / langagent integration which would query Midjourney until it gets a picture which satisfies my criteria, changing the prompt, etc.
The problem is that, if the integration or langagent has an issue -- how do I modularly fix those systems? And so on.
This sort of question "how do I figure out how to fix this prompt-based thing" makes me worry about a sort of "turtles all the way down" problem. Okay, so maybe I don't need to understand exactly how to use these systems to get what I want -- just let another LLM do that ... but what if that LLM has an issue ... and what if the LLM which is prompting that LLM has an issue ...
It concerns me. I'm not sure how well founded my concerns really are.
inpainting.
It seems as if animals and humans can learn on far less training data and on a much dirtier and noisier data set. The data an animal or a human trains on has not been curated to remove misleading data or noise.
It seems that one can produce models with amazing capabilities by training them on vast curated data sets, but I get the sense real brains still have a lot of tricks that we don’t understand.
The raw volume of the data from the human senses is high but a lot of it is low entropy. The high resolution video from our retinas of reading the pages of a book over many hours, and the ASCII representation of said text are equivalent, modulo information about the physical object of the book.
Another salient quote is from Andrej Karpathy (ex-Tesla AI team lead) who said that a camera is a high-bandwidth input "that puts many constraints on the world". Children learn from multi-modal inputs, and vision especially provides a large number of constraints that they can use to learn how the world works.
I have a two-year old, and something I've noticed is that infants have a very strong instinctive urge to gain agency over the world. They try very hard to control things, to make things move, to make sounds, to be able to affect things in all sorts of ways. This must be a very critical part of learning, because they'll kick and scream if you remove their agency. It's as strong an instinct as wanting to eat or sleep. No current LLM has any kind of feedback loop, and has essentially zero agency.
And yet, you can take a 16 year old who's never driven, and teach them within a week to be decent at it and maybe 50-100 hours of driving training and they're competent. You don't need to show them a billion man-hours of driving footage. Even the first people to buy cars in the late 1800s when they were first invented, were able to pick it up almost right away (there weren't even driving licenses back then).
At any rate, driving is just one example. Despite being one of the oldest futuristic sci-fi examples, I don't see restaurants powered by AI. I don't see housekeeping powered by it.
Ok, those are embodied examples, so you say they're unfair. Fine. What remote-friendly jobs are being swept away by AI? Can we even do customer service with AI right now? No, outside of some "front line" chat bot (which just replaces phone trees and terrible localized search engines), we can't. Even if a GPT is trained on a business's proprietary documentation, it's wrong or unresponsive enough that it would cost you more than it would save you by firing your customer support staff.
A 16 year old has a decade of vision input, knows tens of thousands of words, has a coherent theory-of-mind, and has self-preservation instincts honed over hundreds of millions of years of evolution.
The equivalent scenario would be take a fully general AGI(!) that already has a body and has learned to manipulate the physical world, put that in the car and have it learn to drive.
A lot of people seem confused about these scenarios. The currently popular LLMs are like a child raised in a black box, and are weirdly retarded in the same way you would expect a child raised in a black box to be.
Similarly, driving AIs don't speak English, can't take instructions, and are simultaneously learning physics, theory-of-mind, the rules of the rode, and signage conventions without having agency during most of their training.
No they don't. Where were you driving 23 days ago? Or, if you could categorize your driving data in your head, what did you pass, precisely, in the car while you were driving 96 hours ago? AI training data has all of this. It has perfect timestamps with 25-30 frames per second (or more in some cases, with multiple cameras feeding frames in to the dataset) with full 360 view in many cases, adding up to millions of man-hours of driving data from thousands of drivers, which is longer than any human lifespan. A lot of times this data is supplemented with IR representations or LIDAR representations of the environment as well, something a human can't even see.
> Similarly, driving AIs don't speak English, can't take instructions, and are simultaneously learning physics, theory-of-mind, the rules of the rode, and signage conventions without having agency during most of their training.
They're not learning that on their own. They're being poked and prodded by humans in the loop (and indeed, automated tagging tools) who have tweaked the models by adding numerical weights to "bad outcomes" and "good outcomes". Or "good categorizations" and "bad categorizations", or "aligned responses" and "unaligned responses". And because it's just a dumb statistical model, if you tell it the color blue is bad and that it should resist answering any questions related to the color blue, it will agree, because its entire heuristic model for organizing the world was designed by humans.
Similarly, it doesn't have a theory of mind. If it did, it would also have agency and it would already be AGI. Instead, it has access to examples of data of humans exhibiting theory of mind with other humans. And it's got enough parameters in its statistical model that it can accommodate a weighting for this "theory of mind" impact on what the expected output should be.
LLMs are nothing like an intelligent child raised in a black box. I mean, we already have examples of this, vis-a-vis being both blind and deaf: https://en.wikipedia.org/wiki/Helen_Keller.
Just because I cannot do that, still my brain had this input, and was "trained", and we know the real neurons of the brain work completely different in all regards (connecitvities forming, activations, weighting) than our gross simplification of one activation function with a weight.. so not getting your arguments at all.
I have been trained on all these inputs, and even if I cannot recall them now they manifested in a my superior brain somehow.
Nice that in theory the AI training data has all this available, still the resulting model is much simpler, and also cannot recite all of its inputs seen, too??
This is like how LLMs are difficult to train to the point that they understand English, but then relatively easy to specialise.
Similarly, students don’t start their education at University, they spend nearly two decades getting to that point by learning the prerequisites.
> Similarly, it doesn't have a theory of mind. If it did, it would also have agency
These are unrelated concepts. There is evidence that GPT 4 has a rudimentary model of mind but it isn’t conscious itself and is a static model with zero agency over the world.
If I'm talking to you, especially in person, I won't be able to tell you what the second word of three sentences ago was. In fact, I might not have even been paying attention, because I'm already thinking at a higher level about what you're communicating to me and the word wasn't all that important. But a LLM certainly uses this information to guide its responses (along with its ginormous lookup table).
They don’t build a world model at all. They make inferences from text.
Considering they have been trained on 100,000 books, and all of Wikipedia, they are remarkably unintelligent, and really only able to produce text that is consistent with what they have been trained on.
And just like inventing airplanes caused us to say 'all things that fly are not birds', now we may start saying 'all things that display intelligence are not us'
(where intelligence is => human level intelligence and "us" could be swapped out with alive/sentient/conscious/etc at least for now and perhaps in all cases)
The numbers involved are also an order magnitude greater. In a human brain, there are approximately 60 trillion connections between neurons. Made with a variety neurotransmitters and regulated by a host hormones as well, not single floating point weights.
A few more breakthroughs every few years (not guaranteed) could mean better results with 1/100th the data.
Similarly, many animals also pass the knowledge, like certain birds spread the songs among the population.
Besides, individual human knowledge is really specialized. LLM is expected to know everything.
This clickbait in abstracts will hopefully, eventually, have the same effect as clickbait everywhere: people skip reading it.
1. Adenosine, i.e. the sleep molecule affects all outputs of dopamine including the two heterodimers of dopamine D1 and D2 that happen to be combined with Adenosine receptors A1 and A2 and also impacts the two optical computing components the eye itself as its bathed in dopamine and how we detect colored light and the other neuron component of the eye at the back of the human brain.
Okay, so in my ADHD research I got a bit focused on dopamine.
Hello engineering, my old friend. Nothing inspired more confidence than a set of random numbers cast into a formula and then whimsically called a fundamental "law".
They admit that this law is only useful if you already know where the 'breaks' are: "If an additional break of sufficient sharpness happens at a scale that is sufficiently larger than the maximum (along the x-axis) of the points used for fitting, there does not (currently) exist a way to extrapolate the scaling behavior after that additional break."