The main thesis here seems to be that LLMs behave like almost all other machine learning models, in that they are doing pattern matching on their input data, and short circuiting to a statistically likely result. Chain of thought reasoning is still bound by this basic property of reflexive pattern matching, except the LLM is forced to go through a process of iteratively refining the domain it does matching on.
Chain of thought is interesting, because you can combine it with reinforcement learning to get models to solve (seemingly) arbitrarily hard problems. This comes with the caveat that you need some reward model for all RL. This means you need a clear definition of success, and some way of rewarding being closer to success, to actually solve those problems.
Framing transformer based models as pattern matchers makes all the sense in the world. Pattern matching is obviously vital to human problem solving skills too. Interesting to think about what structures human intelligence has that these models don't. For one, humans can integrate absolutely gargantuan amounts of information extremely efficiently.
LLMs are trained, as others have mentioned, first to just learn the language at all costs. Ingest any and all strings of text generated by humans until you can learn how to generate text in a way that is indistinguishable.
As a happy side effect, this language you've now learned happens to embed quite a few statements of fact and examples of high-quality logical reasoning, but crucially, the language itself isn't a representation of reality or of good reasoning. It isn't meant to be. It's a way to store and communicate arbitrary ideas, which may be wrong or bad or both. Thus, the problem for these researchers now becomes how do we tease out and surface the parts of the model that can produce factually accurate and reasonable statements and dampen everything else?
Animal learning isn't like this. We don't require language at all to represent and reason about reality. We have multimodal sensory experience and direct interaction with the physical world, not just recorded images or writing about the world, from the beginning. Whatever it is humans do, I think we at least innately understand that language isn't truth or reason. It's just a way to encode arbitrary information.
Some way or another, we all grok that there is a hierarchy of evidence or even what evidence is and isn't in the first place. Going into the backyard to find where your dog left the ball or reading a physics textbook is fundamentally a different form of learning than reading the Odyssey or the published manifesto of a mass murderer. We're still "learning" in the sense that our brains now contain more information than they did before, but we know some of these things are representations of reality and some are not. We have access to the world beyond the shadows in the cave.
Humans can carve the world up into domains with a fixed set of rules and then do symbolic reasoning within it. LLMs can't see to do this in a formal way at all -- they just occasionally get it right when the domain happens to be encoded in their language learning.
You can't feed an LLM a formal language grammar (e.g. SQL) then have it only generate results with valid syntax.
It's awfully confusing to me that people think current LLMs (or multi-modal models etc) are "close" to AGI (for whatever various definitions of all those words you want to use) when they can't do real symbolic reasoning.
Though I'm not an expert and happy to be corrected...
Adult humans can do symbolic reasoning, but lower mammals cannot. Even ones that share most of our brain structure are much worse at this, if they can do it at all; children need to learn it, along with a lot of the other things that we consider a natural part of human intelligence.
That all points towards symbolic reasoning being a pretty small algorithmic discovery compared to the general ability to pattern match and do fuzzy lookups, transformations, and retrievals against a memory bank. It's not like our architecture is so special that we burned most of our evolutionary history selecting for these abilities, they're very recent innovations, and thus must be relatively simple, given the existence of the core set of abilities that our close ancestors have.
The thing about transformers is that obviously they're not the end of the line, there are some things they really can't do in their current form (though it's a smaller set than people tend to think, which is why the Gary Marcuses of the world always backpedal like crazy and retcon their previous statements as each new release does things that they previously said were impossible). But they are a proof of concept showing that just about the simplest architecture that you could propose that might be able to generate language in a reasonable way (beyond N-gram sampling) can, in fact, do it really, really well even if all you do is scale it up, and even the simplest next-token prediction as a goal leads to much higher level abilities than you would expect. That was the hard core of the problem, building a flexible pattern mimic that can be easily trained, and it turns out to get us way further along the line to AGI than I suspect anyone working on it ever expected it would without major additions and changes to the design. Now it's probably time to start adding bits and bobs and addressing some of the shortcomings (e.g. static nature of the network, lack of online learning, the fact that chains of thought shouldn't be constrained to token sequences, addressing tokenization itself, etc), but IMO the engine at the heart of the current systems is so impressively capable that the remaining work is going to be less of an Einstein moment and more of an elbow grease and engineering grind.
We may not be close in the "2 years of known work" sense, but we're certainly not far in the "we have no idea how to prove the Riemann Hypothesis" sense anymore, where major unknown breakthroughs are still required which might be 50+ years away, or the problem might even be unsolvable.
Yes, I've always thought that LLMs need the equivalent of a limbic system. This is how we solved this problem in organic computers. There is no static 'reward function'. Instead, we have a dynamic reward function computer. It decides from day to day and hour to hour what our basic objectives are. It also crucially handles emotional 'tagging' of memory. Memories that we store are proportionally more likely to be retrieved under similar emotional conditions. It helps to filter relevant memories, which is something LLMs definitely could use.
I think the equivalent of an LLM limbic system is more or less the missing piece for AGI. Now, how you'd go about making one of those I have no idea. How does one construct an emotional state space?
Companies are bad about doing this on purpose. If they set out to build AGI and accomplish something novel, just call that AI and go on fund raising from people who don't know better (or more likely don't care and just want to gamble with others' money).
Continuous RL in a sense. There maybe an undiscovered additional scaling law around models doing what you describe; continuous LLM-as-self-judge, if you will.
Provided it can be determined why a user ended the chat, which may turn out to be possible in some subset of conversations.
And also sometimes write down the conclusion and work backwards, without considering that the reason most likely for the conclusion isn't necessarily going to have the conclusion as the most likely conclusion — I hope I phrased that broken symmetry correctly.
I did think of sour grapes (only thing that came to my mind) and was hoping for something better. Sour grapes doesn't seem too interesting. I think most people can tell you, if you actually ask them, what the differences between their actual successes and their wildest dreams are. But any improvement is still a success and I think that's valid.
The milliken oil drop experiment, “winning “ the space race, mostly anything C levels will tell the board and shareholders at a shareholder meeting, the American wars in Iraq and Afghanistan, most of what Sam Altman or Elon musk has to say, this list continues.
I think you're approaching it form very high level, when you should think about it from much lower level, i.e. success is being determined by stress/dopamine hormones or similar
This article is kind of vague on that tbf:
To conclude, we observed no credible evidence for a beneficial effect of L-dopa (vs. Haloperidol) on reinforcement learning in a reward context, as well as the proposed mechanistic account of an enhanced striatal prediction error response mediating this effect.
Is that controversial? I would say everything a human does is to feel better, and everything someone does that doesn’t make them feel better immediately is just done in the expectation of even greater pleasure later.
Well mine can, with some tactics and strategy layered on top. If I do something I don’t like, I only do it because the payoff later makes it worth it (or at least I think it will from my current knowledge).
It is important that “profit”, comes in various forms, which exchange rates are problematic to calculate (or maybe there can’t be any): not hungry, not thirsty, tastes good, not cold, feel safe, feel excited, feel righteous, feel powerful, listen to music, watch a movie, get curious, satisfy curiosity, laugh, love, sex, rock n roll.
Most behavior we believe is some kind of rational action when it is really blind actions based on fiction or just completely random with rationalizations for the behavior after the fact.
>Interesting to think about what structures human intelligence has that these models don't.
Kant's Critique of Pure Reason has been a very influential way of examining this kind of epistemology. He put forth the argument that our ability to reason about objects comes through our apprehension of sensory input over time, schematizing these into an understanding of the objects, and finally, through reason (by way of the categories) into synthetic a priori knowledge (conclusions grounded in reason rather than empiricism).
If we look at this question in that sense, LLMs are good at symbolic manipulation that mimics our sensibility, as well as combining different encounters with concepts into an understanding of what those objects are relative to other sensed objects. What it lacks is the transcendental reasoning that can form novel and well grounded conclusions.
Such a system that could do this might consist of an LLM layer for translating sensory input (in LLM's case, language) into a representation that can be used by a logical system (of the kind that was popular in AI's first big boom) and then fed back out.
>Such a system that could do this might consist of an LLM layer for translating sensory input (in LLM's case, language) into a representation that can be used by a logical system (of the kind that was popular in AI's first big boom) and then fed back out.
This just goes back into the problems of that AI winter again though. First Order Logic isn't expressive enough to model the real world, while Second Order Logic dosen't have a complete proof system to truly verify all it'sstatements, and is too complex and unyieldy for practical uses. The number of people I would also imagine that are working on such problems would be very few, this isn't engineering that it is analytic philosophy and mathematics.
Kant predates analytical philosophy and some of its failures (the logical positivism you are referring to). The idea here is that first order logic doesn't need to be expressive enough to model the world. Only that some logic system is capable of modeling the understanding of a representation of the world mediated by way of perception (via the current multimodal generative AI models). And finally, it does not need to be complete or correct, just equivalent or better than how our minds do such.
With DeepSeek-R1-Zero, their usage of RL didn't have reward functions really that indicated progress towards the goal afaik.
It was "correct structure, wrong answer", "correct answer", "wrong answer". This was for Math & Coding, where they could verify answers deterministically.
> Framing transformer based models as pattern matchers makes all the sense in the world. Pattern matching is obviously vital to human problem solving skills too. Interesting to think about what structures human intelligence has that these models don't. For one, humans can integrate absolutely gargantuan amounts of information extremely efficiently.
What is also a benefit for humans, I think, is that people are typically much more selective. LLMs train to predict anything on the internet, so for example for finance that includes clickbait articles which have a lifetime of about 2 hours. Experts would probably reject any information in these articles and instead try to focus on high quality sources only.
Similarly, a math researcher will probably have read a completely set of sources throughout the life than, say, a lawyer.
I’m not sure it’s a fundamental difference, but current models do seem to not specialize from the start unlike humans. And that might be in the way of learning the best representations. I know from ice hockey for example, that you can see within 3 seconds whether someone played ice hockey from young age or not. Same with language. People can usually hear an accent within seconds. Relatedly, I've used OpenAI's text to speech a while back and the Dutch voice had an American accent. What this means is that even if you ask LLMs about Buffett's strategy, maybe they have a "clickbait accent" too. So with the current approach to training, the models might never reach absolute expert performance.
When I was doing some NLP stuff a few years ago, I downloaded a few blobs of Common Crawl data, i.e. the kind of thing GPT was trained on. I was sort of horrified by the subject matter and quality: spam, advertisements, flame wars, porn... and that seems to be the vast majority of internet content. (If you've talked to a model without RLHF like one of the base Llama models, you may notice the personality is... different!)
I also started wondering about the utility of spending most of the network memorizing infinite trivia (even excluding most of the content above, which is trash), when LLMs don't really excel at that anyway, and they need to Google it anyway to give you a source. (Aside: I've heard soke people have good luck with "hallucinate then verify" with RAG / Googling...)
i.e. what if we put those neurons to better use? Then I found the Phi-1 paper, which did exactly that. Instead of training the model on slop, they trained it on textbooks! And instead of starting with PhD level stuff, they started with kid level stuff and gradually increased the difficulty.
You can get rid of the trivia by training one model on the slop, then a second model on the first one - called distillation or teacher-student training. But it's not much of a problem because regularization during training should discourage it from learning random noise.
The reason LLMs work isn't because they learn the whole internet, it's because they try to learn it but then fail to, in a useful way.
If anything current models are overly optimized away from this; I get the feeling they mostly want to tell you things from Wikipedia. You don't get a lot of answers that look like they came from a book.
I don't know, babies hear a lot of widely generic topics from multiple people before learning to speak.
I would rather put it that humans can additionally specialize much more, but we usually have a pretty okay generic understanding/model of a thing we consider as 'known'. I would even wager that being generic enough (ergo, has been sufficiently abstracted) is possibly the most important "feature" human's have? (In the context of learning)
> For one, humans can integrate absolutely gargantuan amounts of information extremely efficiently.
What we can integrate, we seem to integrate efficiently*; but compared to the quantities used to train AI, we humans may as well be literally vegetables.
* though people do argue about exactly how much input we get from vision etc., personally I doubt vision input is important to general human intelligence, because if it was then people born blind would have intellectual development difficulties that I've never heard suggested exist — David Blunket's success says human intelligence isn't just fine-tuning on top of a massive vision-grounded model.
Low level details like that aren’t relevant to this discussion. Most human processing power is at the cellular level. The amount of processing power in a single finger literally dwarfs a modern data center, but we can’t leverage that to think only live.
So it’s not a question of ‘a lot’ it’s a question of orders of magnitude vs “the quantities used to train AI”
Library of congress has what 39 million books, tokenize every single one and you’re talking terabytes of training data for an LLM. We can toss blog posts etc to that pile but every word ever written by a person isn’t 20 orders of magnitude larger or anything.
>Hearing is also well into the terabytes worth of information per year.
If we assume that the human auditory system is equivalent to uncompressed digital recording, sure. Actual neural coding is much more efficient, so the amount of data that is meaningfully processed after multiple stages of filtering and compression is plausibly on the order of tens of gigabytes per year; the amount actually retained is plausibly in the tens of megabytes.
Don't get me wrong, the human brain is hugely impressive, but we're heavily reliant on very lossy sensory mechanisms. A few rounds of Kim's Game will powerfully reveal just how much of what we perceive is instantly discarded, even when we're paying close attention.
The sensory information form individual hairs in the ear start off with a lot more data to process than simple digital encoding of two audio streams.
Neural encoding isn’t particularly efficient from a pure data standpoint just an energy standpoint. A given neuron not firing is information and those nerve bundles contain a lot of neurons.
Is that a positive thing? If anything I would consider that as the reverse - LLMs have the "intelligence of vegetables" because even with literally the whole of human written knowledge they can at most regurgitate that back to us with no novelty whatsoever, even though a 2 years old with a not even matured brain can learn a human language from orderS of magnitude less and lower quality input from a couple of people only.
But any Nobel-price winner has read significantly less than a basic LLM, and we see no LLM doing any tiny scientific achievement, let alone that high impact ones.
It's perfectly legit to call these models "thick" because they *need* to read such a vast quantity of text that a human would literally spend two thousand lifetimes to go through it even if that was all the human did with their days.
It also remains the case that, unlike us, they can go through all of that in a few months.
> with no novelty whatsoever, even though a 2 years old with a not even matured brain can learn a human language from orderS of magnitude less and lower quality input from a couple of people only.
You're either grossly underestimating AI or overestimating 2 year olds, possibly both.
I just about remember being a toddler, somewhere between then and 5 was around the age I had the idea that everyone got an invisible extra brain floating next to them for every year they lived. Took me an embarrassingly long time (teens, IIRC) to realise that the witch-duck-weight-comparison scene in Monty Python and the Holy Grail wasn't a documentary, thanks to the part of the film captioned "Famous Historian". One time my dad fell ill, and he was talking to mum about "the tissue being damaged" while I was present, so I gave him a handkerchief (AKA "a tissue"). And while I don't remember this directly, my mum's anecdotes include me saying "fetrol fump", waving a spoon in a jam pan and calling this act "spelling", and when discovered running around with my pockets inside-out explaining myself as trying to fly because I apparently thought that the lining of a pocket was called a "wing".
When it comes to human novelty, I also quite often find there's a lot of remixing going on that just isn't immediately apparent. As Steve Jobs apparently once said, “Good artists copy; great artists steal.”, except Jobs stole that quote from Picasso.
It's easy to categorise different levels with AI, but which one of these counts as "novelty", and how often do humans ever achieve each of these grades?
0. Memorisation of the training set. Think: bunch of pictures, pick best fit.
1. Linear interpolation between any pair of elements in the training set. Think: simple cross-fade between any two pictures, but no tracking or distorting of features during that fade.
2. Let the training set form a basis vector space, and interpolate freely within the constraints of the examples. Think: if these pictures are faces, it would make any hair colour between the most extreme limits shown, etc.
3. Extrapolate beyond the examples. Think: Even if no black or white hair was visible, so long as several shades of grey were, it could reach the ideas of black or white hair.
4. Invent a new vector. Think: even if it had been trained only on black-and-white images, it could still invent green hair.
> But any Nobel-price winner has read significantly less than a basic LLM, and we see no LLM doing any tiny scientific achievement, let alone that high impact ones.
We do see them doing *tiny* scientific achievements, with extra emphasis on "tiny". Just like with using them in software, even the best "only" act like fresh graduates.
When any AI gets to high-impact… the following (fictional) quote comes to mind: "as soon as we started thinking for you, it really became our civilization."
> that a human would literally spend two thousand lifetimes to go through it even if that was all the human did with their days.
Well, `cp` would go over that data even faster, but depending on what retention/conclusion is reached from that it may or may not be impressive.
Humans are fundamentally limited by our biology, and rotating a tiny sphere and turning pages and serial processing does make certain hard limits on us.
A two years old can definitely say stupid stuff, or have wildly incomplete/incorrect models of their reality, but can most certainly already think and reason, and update their internal models at any point.
> Tiny scientific achievements, only acting as fresh graduates with regards to software
I don't believe they are anywhere close to being as good at software as a fresh graduate. Sure, many people write terrible code, and there are a lot of already solved problems out there (not even just solved, but solved thousands times) - LLMs are definitely a novel tool when it comes to finding information based on some high-ish level patterns (over exact string match, or fuzzy match), and they are very good at transforming between different representations of said data, with minimal (and hard limited) reasoning capabilities, but I have never seen evidence of going any further than that.
I don't think your grades are "correct" - e.g. a random generator can easily create new vectors, but I wouldn't call that intelligence. Meanwhile, that two years old can do a novel discovery from their POV every couple of day, potentially turning around their whole world model each day. To me, that sounds way "cooler" than a statistically likely token given these previous tokens, and LLMs definitely need some further structure/architecture to beat humans.
--
I do like your last quote though, and definitely agree there!
> Well, `cp` would go over that data even faster, but depending on what retention/conclusion is reached from that it may or may not be impressive.
Sure, but it would be a level zero on that list, right?
I'd say even Google would be #0.
> A two years old can definitely say stupid stuff, or have wildly incomplete/incorrect models of their reality, but can most certainly already think and reason, and update their internal models at any point.
I think that this presumes a certain definition of "think" and "reason". Monsters under the bed? To move from concrete examples to the abstract, from four apples to the idea of four?
Imagine a picture of a moon's orbit around the parent planet and the planet's orbit around a star, first at one time of year, then again 60° later, the circular orbits of each drawn clearly, with the two positions of the moon's orbits aligned at the top of the image; exaggerate the scale for clarity, and find it in an astronomy book — my peers at age 6 or 7 thought it was a picture of a mouse.
Imagine teachers and an ambulance crew explaining to the class how blood is donated, showing that they're putting a bag up the teachers sleeves and explaining how they'll demonstrate this by taking "blood" (fake? No idea at this point) from that bag. Everyone's looking, we see it go up the sleeve. We see the red stuff come out. Kid next to me screams "they're killing her!". Rather than say "we literally saw the bag go up the sleeve", 5-year-old-me tried to argue on the basis that killing a teacher in front of us was unlikely — not wrong, per say, but a strange argument and I wondered even at the time why I made it.
Are these examples of "reason"? Could be. But, while I would say that we get to the "children say funny things" *with far fewer examples than the best AI*, it doesn't seem different in kind to what AI does.
> LLMs are definitely a novel tool when it comes to finding information based on some high-ish level patterns (over exact string match, or fuzzy match), and they are very good at transforming between different representations of said data, with minimal (and hard limited) reasoning capabilities, but I have never seen evidence of going any further than that.
Aye. So, where I'm going with #2 and #3: even knowing what the question means well enough to respond by appropriately gluing together a few existing documents correctly, requires the AI to have created a vector space of meaning from the words — the sort of thing which word2vec did. But:
To be able to translate questions into answers when neither the question nor the answer are themselves literally in the training set, requires at least #2. (If it was #1, you might see it transition from "Elizabeth II was Queen of the UK" to "Felipe VI is King of Spain" via a mid-point of "Macron is Monarch of France").
For #3, I've tried the concrete example of getting ChatGPT (free model a few months back now) to take the concept of the difference between a racoon and a wolf and apply this difference again on top of a wolf, and… well, their combination of LLM and image generator gave me what looked like a greyhound, so I'm *not* convinced that OpenAI's models demonstrate this in normal use — but also, I've seen this kind of thing demonstrated with other models (including Anthropic, so it's not a limit of the Transformer architecture) and the models seem to do more interesting things.
Possibly sample bias, I am aware of the risk of being subject to a Clever Hans effect.
For #4, this seems hard to be sure it has happened when it seems to have happened. I don't mean what word2vec does, which I realise now could be described in similar language, as what word2vec does is kinda a precursor to anything at least #1. Rather, what I mean, in a human, would seem like "spots a black swan before it happens". I think the invention of non-Euclidian geometry might count, but even then I'm not sure.
I feel like if you take the underlying transformer and apply to other topics, e.g., eqtransformer, nobody questions this assumption. It’s only when language is in the mix do people suggest they are something more and some kind of “artificial intelligence” akin to the beginnings of Data from Star Trek or C3P0 from Star Wars.
Human processing is very interesting and should likely lead to more improvements (and more understanding of human thought!)
Seems to me humans are very good at pattern matching, as a core requirement for intelligence. Not only that, we are wired to enjoy it innately - see sudoku, find Waldo, etc.
We also massively distill input information into short summaries. This is easy to see by what humans are blind to: the guy in a gorilla suit walking through a bunch of people passing a ball around, or basically any human behavior magicians use to deceive or redirect attention. We are mombarded with information constantly. This is the biggest difference between us and LLMs as we have a lot more input data and also are constantly updating that information - with the added feature/limitation of time decay. It would be hard to navigate life without short term memory or a clear way to distinguish things that happened 10 minutes ago from 10 months ago. We don't fully recall each memory of washing the dishes but junk the vast, vast majority of our memories, which is probably the biggest shortcut our brains have over LLMs.
Then we also, crucially, store these summaries in memory as connected vignettes. And our memory is faulty but also quite rich for how "lossy" it must be. Think of a memory involving a ball from before the age of 10 and most people can drum up several relevant memories without much effort, no matter their age.
> Interesting to think about what structures human intelligence has that these models don't.
Pain receptors. If you want to mimic human psyche you have to make your agent want to gather resources and reproduce. And make it painful to lack those resources.
Now, do we really have to mimic human intelligence to get intelligence? You could make the point the internet is now a living organism but does it have some intellect or is it just some human parasite / symbiote?
>Interesting to think about what structures human intelligence has that these models don't.
If we get to the gritty details of what gradient descent is doing, we've got a "frame", i.e a matrix or some array of weights contains the possible solution for a problem, then with another input of weights we're matching a probability distribution to minimize the loss function with our training data to form our solution in the "frame". That works for something like image recognition, where the "frame" is just the matrix of pixels, or in language models where we're trying to find the next word-vector given a preceding input.
But take something like what Sir William Rowan Hamilton was doing back in 1843. He know that complex numbers could be represented in points in a plane, and arthimetic could be performed on them, and now he wanted to extend a similar way for points in a space. With triples it is easy to define addition, but the problem was multiplication. In the end, he made an intuitive jump, a pattern recognition when he realized that he could easily define multiplications used quadruples instead, and thus was born the Quaternion that's a staple in 3D graphics today.
If we want to generalize this kind of problem solving into a way that gradient descent can solve, where do we even start? First of all, we don't even know if a solution is possible or coherent or what "direction" we are going towards. It's not a systematic solution, it's rather one that pattern in one branch of mathematics was recognized into another. So perhaps you might use something like Category Theory, but then how are we going to represent this in terms of numbers and convex functions, and is Category Theory even practical enough to easily do this?
> Interesting to think about what structures human intelligence has that these models don't
Chiefly?
After having thought long and hard, building further knowledge on the results of the process of having thought long and hard, and creating intellectual keys to further think long and hard better.
You do not need sensorial feedback to do math. And you do not need full sensors to have feeback - one well organized channel can suffice for some applications.
To learn new math, a professional mathematician foremostly just thinks further (it's its job); to discuss with other entities (and acquire new material), textual input and outputs suffice.
Your statement, not mine. And I wrote intelligence, not sentience.
People who become quadriplegic as adults (or older children) have already developed intelligence before.
My theory (which I have not researched in any way) implies that someone born fully quadriplegic would be severely impaired in developing intelligence. Sight and hearing are of course also important sources of feedback, the question is whether they are sufficient.
You might get a kick out of this essay by Robert Epstein from 2016: https://aeon.co/essays/your-brain-does-not-process-informati... (The empty brain - Your brain does not process information, retrieve knowledge or store memories. In short: your brain is not a computer)
Maybe I misunderstood it, but I feel that it's a weird article, because it fails to establish any vocabulary and then seem to uses words in uncertain ways, as if constructing the narrative by specifically crafting (but never truly explaining/define) some model that's not true, but presenting the argument with significantly expanded scope. Drastically reduced (which is not really correct, but may help me to convey my general impression/feelings only) it's kinda sorta like-ish "we aren't doing it the way our computers do, thus the information processing metaphor is wrong".
Like when talking about that experiment and an image of the dollar bill, it never talks about what's an "image", just states that there wasn't one stored in a brain, in "any sense". And then goes on describing the idea that seem to match the description of a "mental image" from cognitive science.
As I [very naively] get it... Information theory is a field of mathematics. Unlike all those previous concepts like humours, mechanical motions or electric activities, math is here to establish terminology and general principles that don't have to fundamentally change if^W when we learn more. And that's why it got stuck.
There is a whole genre of essays like this talking about behaviour in a human specific way. But, I wish they engaged with the notions of the Church-Turing thesis and the Universal Turing Machine which indicates that any behaviour following standard physics principles is in fact computable.
(FWIW, I dont think that humans can be reduced to computing, but the Church-Turing thesis is a powerful counterargument which more biologists and psychologists should engage with).
I stopped reading before reaching 2/3 of it but the start is already strawman after strawman (or misunderstanding to be generous).
I don’t think most people believe the brain is made up of a discrete Processing part that accesses information from a memory part that’s encoded in binary there. But just because the brain doesn’t contain a literal encoding of something in binary doesn’t mean the neurons don’t store the information.
If you download the weights of an LLM, you’re not going to find the text it can output „from memory“ anywhere in the file, but the weights still encode the information and can retrieve it (with some accuracy).
Coming up with a reward model seems to be really easy though.
Every decidable problem can be used as reward model. The only downside to this is that the LLM community has developed a severe disdain for making LLMs perform anything that can be verified by a classical algorithm. Only the most random data from the internet will do!
Your post on Twitter uses slightly more words than the ones preceding it above to make the exact same point. Was there really any reason to link to it? Why not expand on your argument here?
"LLMs are fundamentally matching the patterns they've seen, and their abilities are constrained by mathematical boundaries. Embedding tricks and chain-of-thought prompting simply extends their ability to do more sophisticated pattern matching."
Chain of thought is interesting, because you can combine it with reinforcement learning to get models to solve (seemingly) arbitrarily hard problems. This comes with the caveat that you need some reward model for all RL. This means you need a clear definition of success, and some way of rewarding being closer to success, to actually solve those problems.
Framing transformer based models as pattern matchers makes all the sense in the world. Pattern matching is obviously vital to human problem solving skills too. Interesting to think about what structures human intelligence has that these models don't. For one, humans can integrate absolutely gargantuan amounts of information extremely efficiently.