I believe LLM training is happening at a scale great enough for models to start picking up on those patterns. Whether or not this can ever be equivalent to living through the experience personally, or at least asymptomatically approach it, I don't know. At the limit, this is basically asking about the nature of qualia. What I do believe is that continued development of LLMs and similar general-purpose AI systems will shed a lot of light on this topic, and eventually help answer many of the long-standing questions about the nature of conscious experience.
I dunno. I figure it's more likely we keep emulating behaviors without actually gaining any insight into the relevant philosophical questions. I mean what has learning that a supposed stochastic parrot is capable of interacting at the skill levels presently displayed actually taught us about any of the abstract questions?
IMHO a lot. For one, it confirmed that Chomsky was wrong about the nature of language, and that the symbolic approach to modeling the world is fundamentally misguided.
It confirmed the intuition I developed of the years of thinking deeply about these problems[0], that the meaning of words and concepts is not an intrinsic property, but is derived entirely from relationships between concepts. The way this is confirmed, is because the LLM as a computational artifact is a reification of meaning, a data structure that maps token sequences to points in a stupidly high-dimensional space, encoding semantics through spatial adjacency.
We knew for many years that high-dimensional spaces are weird and surprisingly good at encoding semi-dependent information, but knowing the theory is one thing, seeing an actual implementation casually pass the Turing test and threaten to upend all white-collar work, is another thing.
--
I realize my perspective - particularly my belief that this informs the study of human mind in any way - might look to some as making some unfounded assumptions or leaps in logic, so let me spell out two insights that makes me believe LLMs and human brains share fundamentals:
1) The general optimization function of LLM training is "produce output that makes sense to humans, in fully general meaning of that statement". We're not training these models to be good at specific skills, but to always respond to any arbitrary input - even beyond natural language - in a way we consider reasonable. I.e. we're effectively brute-forcing a bag of floats into emulating the human mind.
Now that alone doesn't guarantee the outcome will be anything like our minds, but consider the second insight:
2) Evolution is a dumb, greedy optimizer. Complex biology, including animal and human brains, evolved incrementally - and most importantly, every step taken had to provide a net fitness advantage[1], or else it would've been selected out[2]. From this follows that the basic principles that make a human mind work - including all intelligence and learning capabilities we have - must be fundamentally simple enough that a dumb, blind, greedy random optimizer can grope its way to them in incremental steps in relatively short time span[3].
2.1) Corollary: our brains are basically the dumbest possible solution evolution could find that can host general intelligence. It didn't have time to iterate on the brain design further, before human technological civilization took off in the blink of an eye.
So, my thinking basically is: 2) implies that the fundamentals behind human cognition are easily reachable in space of possible mind designs, so if process described in 1) is going to lead towards a working general intelligence, there's a good chance it'll stumble on the same architecture evolution did.
--
[0] - I imagine there are multiple branches of philosophy, linguistics and cognitive sciences that studied this perspective in detail, but unfortunately I don't know what they are.
[1] - At the point of being taken. Over time, a particular characteristic can become a fitness drag, but persist indefinitely as long as more recent evolutionary steps provide enough advantage that on the net, the fitness increases. So it's possible for evolution to accumulate building blocks that may become useful again later, but only if they were also useful initially.
[2] - Also on average, law of big numbers, yadda yadda. It's fortunate that life started with lots of tiny things with very short life spans.
[3] - It took evolution some 3 billion years to get from bacteria to first multi-cellular life, some extra 60 million years to develop a nervous system and eventually a kind of proto-brain, and then an extra 500 million years iterating on it to arrive at a human brain.
I agree. But there's a very strong incentive to not to; you can't simply erase hundreds of millennia of religion and culture (that sets humans in a singular place in the cosmic order) in the short few years after discovering something that approaches (maybe only a tiny bit) general intelligence. Hell, even the century from Darwin to now has barely made a dent :-( . Buy yeah, our intelligence is a question of scale and training, not some unreachable miracle.
> so if process described in 1) is going to lead towards a working general intelligence, there's a good chance it'll stumble on the same architecture evolution did.
I wouldn't be so sure of that. Consider that a biased random walk using agents is highly dependent on the environment (including other agents). Perhaps a way to convey my objection here is to suggest that there can be a great many paths through the gradient landscape and a great many local minima. We certainly see examples of convergent evolution in the natural environment, but distinct solutions to the same problem are also common.
For example you can't go fiddling with certain low level foundational stuff like the nature of DNA itself once there's a significant structure sitting on top of it. Yet there are very obviously a great many other possibilities in that space. We can synthesize some amino acids with very interesting properties in the lab but continued evolution of existing lifeforms isn't about to stumble upon them.
> the symbolic approach to modeling the world is fundamentally misguided.
It's likely I'm simply ignorant of your reasoning here, but how did you arrive at this conclusion? Why are you certain that symbolic modeling (of some sort, some subset thereof, etc) isn't what ML models are approximating?
> the meaning of words and concepts is not an intrinsic property, but is derived entirely from relationships between concepts.
Possibly I'm not understanding you here. Supposing that certain meanings were intrinsic properties, would the relationships between those concepts not also carry meaning? Can't intrinsic things also be used as building blocks? And why would we expect an ML model to be incapable of learning both of those things? Why should encoding semantics though spatial adjacency be mutually exclusive with the processing of intrinsic concepts? (Hopefully I'm not betraying some sort of great ignorance here.)
I'm not the poster, but my answer would be because symbolic manipulation is way too expensive. Parallelizing it helps, but long dependency chains are inherent to formal logic. And if a long chain is required, it will always be under attack by a cheaper approximation that only gets 90% of the cases right—so such chains are always going to be brittle.
(Separately, I think that the evidence against humans using symbolic manipulation in everyday life, and the evidence for error-prone but efficient approximations and sloppy methods, is mounting and already overwhelming. But that's probably a controversial take, and the above argument doesn't depend on it.)
There is no evolution == better/more fit, as long as reproduction cascade goes uninterrupted, genes can evolve any which way and still survive whether they're neutral or a negative.
The meaning of words is derived from our experience of reality.
Words is how the experiencing self classifies experienced reality into a lossy shared map for the purposes of communication with other similarly experiencing selves, and without that shared experience words are meaningless, no matter what graph you put them in.
It's true that sometimes something can get taken along for the ride by luck of the draw. In which case what's really being selected for is some subgroup of genes as opposed to an individual one. In those cases there's some reason that losing the "detrimental" gene would actually be more detrimental, even if indirectly.
I didn't say "words". I said "concepts"[0].
> Words is how the experiencing self classifies experienced reality into a lossy shared map for the purposes of communication with other similarly experiencing selves, and without that shared experience words are meaningless, no matter what graph you put them in.
Sure, ultimately everything is grounded in some experiences. But I'm not talking about grounding, I'm talking about the mental structures we build on top of those. The kind of higher-level, more abstract thinking (logical or otherwise) we do, is done in terms of those structures, not underlying experiences.
Also: you can see what I mean by "meaning being defined in terms of relationships" if you pick anything, any concept - "a tree", "blue sky", "a chair", "eigenvector", "love", anything - and try to fully define what it means. You'll find the only way you can do it is by relating it to some other concepts, which themselves can only be defined by relating them to other concepts. It's not an infinite regression, eventually you'll reach some kind of empirical experience that can be used as anchor - but still, most of your effort will be spent drawing boundaries in concept space.
--
[0] - And WRT. LLMs, tokens are not words either; if that wasn't obvious 2 years ago, it should be today, now that multimodal LLMs are commonplace. The fact that this - tokenizing video and audio and other modalities into the same class of tokens as text, and embedding them in the same latent space - worked spectacularly well - is pretty informative to me. For one, it's a much better framework to discuss the paradox of Sapir-Whorf hypotheses than whatever was mentioned on Wikipedia to date
A word is a lexical unit, whereas a concept consists of 1) a number of short designations (terms, usually words, possibly various symbols) that stand for 2) a longer definition (created traditionally through the use of other terms, a.k.a. words).
> I'm talking about the mental structures we build on top of those
Which are always backed by experience of reality, even the most “abstract” things we talk about.
> You'll find the only way you can do it is by relating it to some other concepts
Not really. There is no way to fully communicate anything you experience to another person without direct access to their mind, which we never gain. Defining things is a subset of communication, and just as well it is impossible to fully define anything that involves experience, which is everything.
So you are reiterating the idea of organising concepts into graphs. You can do that, but note that any such graph:
1) is a lossy map/model, possibly useful (e.g., for communicating something to humans or providing instructions to an automated system) but always wrong with infinite maps possible to describe the same reality from different angles;
2) does not acquire meaning just because you made it a graph. Symbols acquire meanings in the mind of an experiencing self, and the meaning they acquire depends on recipient’s prior experience and does not map 1:1 to whatever meaning there was in the mind of the sender.
You can feel that I am using a specific narrow definition of “meaning” but I am doing that to communicate a point.
You're looking at Structuralism. First articulated by Ferdinand de Saussure in his Course in General Linguistics published in 1916.
This became the foundation for most of subsequent french philosophy, psychology and literary theory, particularly the post-structuralists and postmodernists. Lacan, Foucault, Derrida, Barthes, Deleuze, Baudrillard, etc.
These ideas have permeated popular culture deeply enough that (I suspect) your deep thinking was subconsciously informed by them.
I agree very much with your "Chomsky was wrong" hypothesis and strongly recommend the book "Language Machines" by Leif Weatherby, which is on precisely that topic.
Even if we were to obtain evidence clearly demonstrating that LLMs don't reason symbolically, why should we interpret that as an indication of what humans do? Certainly it would be highly suggestive, but "hey we've demonstrated that thing can be done this way" doesn't necessarily mean that thing _is_ being done that way.
They make people willing to seriously consider a wider ranger of possibilities. Without the example of LLMs, people tended to be very attached to a "hoomanz special, you need to have the exact same physical substrate to do anything remotely similar." With LLMs, now you have the equally misguided (IMHO) "LLMs talk like people, so they must be doing the same thing as people".
> Couldn't you have argued the same thing prior to LLMs? That evolution is a greedy optimizer etc etc therefore humans don't perform symbolic reasoning.
Could and did. I've long argued that usually when we think we (or others) are thinking logically, that that's a retconned explanation for a decision or behavior that was really arrived at in a messier and more error-prone but also more powerful mechanism. ("Powerful" as in, with wider applicability and generalizability. Not necessarily more capable of arriving at "correct" solutions.)
> humans don't perform symbolic reasoning [is] merely a hypothesis - there's zero evidence one way or the other
"Zero evidence" is inaccurate. There is lots of evidence for what Kahneman calls system 1 and system 2 thinking. (The reports of the death of this theoretical model are greatly exaggerated -- while lots of the research covered in "Thinking, Fast and Slow" has been debunked, the existence of and distinction between these modes of thought are empirically supported.) There's also tons of evidence for how we graft explanations for our past decisions onto them after the fact (even when we are fooled into thinking we made a decision that we didn't, we'll still generate and believe a logical explanation).
But also, I'm not claiming that neither LLMs nor humans ever reason symbolically. I think both do, occasionally. I claim that the bulk of the behavior of both LLMs and humans is not decided upon via symbolic reasoning. The basic reasons are similar -- it's cheaper and more efficient to use other approximate mechanisms, and both of us learn to do what works rather than what is correct, at least most of the time.
> Even if we were to obtain evidence clearly demonstrating that LLMs don't reason symbolically, why should we interpret that as an indication of what humans do? Certainly it would be highly suggestive, but "hey we've demonstrated that thing can be done this way" doesn't necessarily mean that thing _is_ being done that way.
Agreed. But "highly suggestive" is all I'm going for. (And only highly suggestive that neither of us rely heavily on symbolic reasoning, not highly suggestive that we work the same way as LLMs.)
It's tricky, because LLMs are almost designed to introduce as many confounding factors as possible. For example, it's popular to claim that you can't have "real" intelligence without embodiment. (Though that position seems to be declining in popularity with LLM advances.) You need skin that can feel a breeze, a body that can feel pain, a mind that can suffer. You need neurons that live in a chemical bath whose composition and history are part of the processing mechanism. But LLMs are trained out tons and tons of output that was generated by embodied creatures, and so that input data "carries along" the results of a processing mechanism that relies on embodiment. An LLM that claims to enjoy long walks on the beach and the feeling of sand between its toes, and that gets grumpy during the dark season, isn't lying. It was built to emulate the output of beings that do enjoy sandy toes, and can generate new "thoughts"/outputs that are produced via mechanisms that take that into account.
To the extent that chatbots live, they live vicariously through us.
> Possibly I'm not understanding you here. Supposing that certain meanings were intrinsic properties, would the relationships between those concepts not also carry meaning? Can't intrinsic things also be used as building blocks? And why would we expect an ML model to be incapable of learning both of those things? Why should encoding semantics though spatial adjacency be mutually exclusive with the processing of intrinsic concepts? (Hopefully I'm not betraying some sort of great ignorance here.)
I probably shouldn't respond to this part, because I don't really agree with the original assertion. Or rather, I think this ends up boiling down to a disagreement over semantics, and so isn't a particularly interesting question.
Relationships between concepts covers a lot of what "meaning" is. You can teach a computer to translate from language X to Y purely based on it learning the relationships of words to each other in each language, and then generating a mapping between the weight-graph of X to the weight-graph of Y. (Yeah, citation needed; I remember reading some specific evidence for this, but I don't remember where.) So you can get a long way with just relationships.
At the same time, I don't think that proves that the relationships between concepts are everything. A human getting burned and learning the word "hot" could be described as "hot" having an intrinsic meaning. But you could equally describe it as a relationship between the action taken, the sensation experienced, and the phonemes heard. If all those are "concepts", then the relationships between concepts are everything. If they're not, then you can call something intrinsic. Personally, that strikes me as a pointless philosophical question.
I guess you could argue that if you have an LLM trained on mostly English but also enough Chinese to be able to translate, and it generates text including the word "hot", then if you compare that to the same LLM generating text including the Chinese word for hot, that there's more opportunity for drift in the Chinese output. The first case has the chain of a human feeling pain => writing text containing "hot" => generating text containing "hot", whereas the second has the chain of a human feeling pain => writing text containing "hot" => encoded associations between English and Chinese concepts embedded in weights => writing Chinese text containing Chinese "hot". The English "hot" output is more tightly connected to and more directly derives from the physical sensation of burning. (This is of course assuming majority English training data, and in particular a relative lack of Chinese training data containing the "hot" word/concept.) So in a way, you could claim that the question of whether the word "hot" has an intrinsic meaning is relevant and useful. But it seems to me that's just one way of describing the origins of training data; use it if it's useful.
But this forces one into the position that whatever a LLM is doing is not real language, just an imitation of language.
If you take the fact that LLMs are emitting "real language" at face value, then you need to adopt a more structuralist view of language, in which "meaning" is part of the system of language itself and does not need to be grounded biologically.
I don't think holding a structuralist view of language precludes believing that humans have a biological facility for language, or even that language is shaped and ultimately a result of that biological facility. Its more an argument over what language IS -- a symbolic system, or an extension of the human brain.
If you believe in esoteric stuff such as "qualia", then LLMs also tell you nothing. You can continue to believe that true experience requires a human brain and all that a computer does is imitation. But this has no observable consequences that can be used to falsify these ideas, so is not a scientific concept in the first place.