[HN Gopher] The illusion of state in state-space models
       ___________________________________________________________________
        
       The illusion of state in state-space models
        
       Author : canjobear
       Score  : 55 points
       Date   : 2024-06-07 11:53 UTC (11 hours ago)
        
 (HTM) web link (arxiv.org)
 (TXT) w3m dump (arxiv.org)
        
       | optimalsolver wrote:
       | >Figure 1: We prove that SSMs, like transformers, cannot solve
       | inherently sequential problems like permutation com- position
       | (S5), which lies at the heart of state-tracking prob- lems like
       | tracking chess moves in source-target notation (see Section 3.2),
       | evaluating Python code, or entity tracking. Thus, SSMs cannot, in
       | general, solve these problems either
       | 
       | Do Microsoft & friends who are about to build trillion dollar AI
       | data centers know about these proven limitations of transformer-
       | based architectures?
        
         | mjburgess wrote:
         | Yes. The solution is to hoover up all the queries people are
         | currently putting to ChatGPT, save how people have "prompt
         | engineer"ed the solution (ie., answered the question
         | themselves); and then hope that you can just feed this back to
         | people without them noticing.
         | 
         | The only open question is whether people have common-enough
         | queries for this charade to work out. It seems there's quite a
         | lot at least. But this number will decrease over time for
         | various reasons. So it's a game of building a system that can
         | be retrained on the answers people are giving it fast enough
         | that people don't notice where the answers are coming from.
        
           | FeepingCreature wrote:
           | Counterpoint: the more networks internalize the patterns
           | behind engineered prompts, the closer they get to being
           | general reasoners.
        
             | mjburgess wrote:
             | Statistical patterns in text tokens have nothing to do with
             | reasoning. They lack such a property to learn in the first
             | place. This would be more obvious, I guess, if the tokens
             | were in an alien language. Consider a translation to 1,000
             | different alien languages of any given novel.. there is no
             | distributional property they would share.
             | 
             | Statistical AI is just a way of sampling from a historical
             | dataset with a similarity metric. It only works to answer
             | questions if you're sampling from a (Q, A) database in the
             | same language the user already understands. The question
             | was answered by a reasoner, it is now answered by a system
             | which replays answers.
        
               | FeepingCreature wrote:
               | This is just chinese room. But also, I disagree that the
               | languages would not share properties. A novel is too
               | small. Consider a network trained on GPT-3 scale datasets
               | in 1000 alien languages. The shared structures behind the
               | sentences will be the same, even if the grammar is
               | completely different. Stars will be stars, moons will be
               | moons. I'd bet you that the model would be able to
               | translate shared concepts between those languages even if
               | it had _not_ been trained on the same novels, same as GPT
               | can translate terms that it has not seen in dictionary
               | pairs.
               | 
               | (I have no idea if it can! But I'm confident enough that
               | I'm willing to just say it can, at risk of being proven
               | wrong. If GPT can't do that, I'm fundamentally
               | misunderstanding how it works - that is, I don't have a
               | paper offhand showing that GPT shares concept neurons
               | between languages, but I'm willing to bet I could find
               | one if I went looking.)
               | 
               | In other words, if you co-trained GPT-4 on Earth Internet
               | and Alien Internet, there's a good chance it'd end up
               | able to translate English to Alienese, purely as an
               | emergent ability, if it had the concept of translation at
               | all.
               | 
               | Intelligence is compression. With sufficient abstraction
               | (layers) and sufficient volume (dataset), any description
               | of the same reality will assume the same structure. And
               | no learning algo worth its salt will keep two identical
               | structures around.
        
               | mjburgess wrote:
               | Almost every predicate in natural languages is an
               | arbitrary association of properties in the world. There
               | isn't any property a person has "bald", nor are there
               | "tree"s in the world. What bundles of properties the vast
               | majority of language names are a product of historical
               | and contingent associations we've made for practical
               | reasons.
               | 
               | Likewise, languages do not have the same distributional
               | structure. There is no reason an alien language would use
               | discrete tokens to name properties, nor co-locate tokens
               | by linear position in a 'sentence'. Historical human
               | languages did not; using, eg., the full 2D structure of
               | the clay tablet.
               | 
               | To suppose that it is the glyphs and their colocation
               | which somehow bare meaning is a nonesensical
               | superstition. The world is what our words mean, and it is
               | we, in that world, who provide them their meaning. We can
               | do so with arbitrary linguistic structures.
        
               | FeepingCreature wrote:
               | Right, but neural networks at a certain level of scale
               | begin to compress _the perception described by the text_
               | rather than the text per se. The lexical features are
               | abstracted away and the network begins to understand
               | meaning directly. Of course, you 'd still have
               | differentiation by different perceptions and focus on
               | different parts of reality, but I see no reason why that
               | should not likewise abstract out at even higher levels.
               | Ultimately, the whole point of language is that it's
               | about reality, and there is only one reality no matter
               | how many different cultures you feed in. The network
               | isn't doing magic - it's just doing exactly the same
               | thing that we're doing with language. The reason why
               | language is useful is the same reason it's learnable.
        
               | mjburgess wrote:
               | If you want to understand how statistical AI systems
               | work, then you should understand the basics of formalism.
               | They are curve-fitting algorithms that approximate the
               | probability distribution of a hisotrical dataset. That's
               | all they are.
               | 
               | By analogising to any animal or human mind, you aren't
               | describing anythign that acutally exists. A neutral
               | network algorithm isnt neural and it isnt a network. It's
               | a statistical curve-fitting algorith. One oughtnt study
               | trees to understand a decision tree either.
               | 
               | This language is entirely metaphorical. There are no
               | neurones in an NN, there are just summation entires in a
               | matrix. This matrix comprises weights, which define the
               | orientation and scale of the line pieces which form the
               | curve being fit to the data.
        
               | FeepingCreature wrote:
               | They're, in particular, _multi-level_ curve-fitting
               | algorithms. A formalism which, I 'm sure you are aware,
               | can express any bounded computation. In particular,
               | depending on how you set up the tape, neural networks are
               | comfortably Turing complete.
               | 
               | We're not talking about 2-layer perceptrons anymore here.
               | 
               | In my opinion, you should look less at the formalism and
               | more at the empirically demonstrated performance.
        
               | mjburgess wrote:
               | I'm not aware that it can express any bounded
               | computation. But such is irrelevant anyway: there is no
               | function from the distributional structure of text tokens
               | to reasoning. There is nothing to approximate, and this
               | is a system trained just to approximate some presumed
               | function: our mental capacities arent patterns in glyphs.
               | 
               | Whatever property you might imagine a transformer
               | architecture to have (and it is vastly fewer than the set
               | needed for general computation), the problem here is that
               | it's being applied to approximate the structure of
               | historical text data which isnt being generated from such
               | a function.
               | 
               | Indeed, there is no function which generates text data,
               | there's a very large number of independent generating
               | processes that give rise to the distribution of text. The
               | phrase "the war in ukraine" acquires a different
               | semantics over 2010-2030 in a radically different way
               | than, "I liked that film" does.
               | 
               | The capacities which produce distributions of text tokens
               | are highly varied, complex, invovle a vast array of our
               | mental processes, and so on. There's literally almost
               | nothing in the distribution of text tokens that
               | corresponds to any features of these processes.
               | 
               | The structure of language is conventional, and rests on
               | our familiarity with such conventions. Otherwise, let's
               | end all science, everything to be known about the world
               | derives from how "e" occurs alongside "lectron"
        
               | stevenhuang wrote:
               | Computability is of critical relevance.
               | 
               | Our mental capacities _can_ be patterns.
               | https://en.m.wikipedia.org/wiki/Predictive_coding
               | 
               | > There's literally almost nothing in the distribution of
               | text tokens that corresponds to any features of these
               | processes.
               | 
               | There's nothing in evolutionary fitness that necessitates
               | intelligence or reasoning ability either. Yet here we
               | are.
               | 
               | Sorry but if you don't see the connections then you need
               | to do some reading on theory of mind, cognition,
               | information theory, physics, philosophy. All of the
               | fundamental basis are met to allow reasoning to emerge in
               | LLMs.
               | 
               | It is clear now where your confusion lies, and why you
               | are led to believe so strongly that LLMs cannot reason:
               | it is because you are an ML practitioner you overweight
               | your expertise yet you don't know what you don't know,
               | and have foundational gaps in your knowledge. You lack
               | the context in these other fields. If you had them, your
               | position should be closer to agnostic than this strong
               | belief of yours that LLMs in their current form cannot
               | reason.
        
               | mjburgess wrote:
               | lol, well when I've written my PhD on those areas we can
               | return to whether I'm an expert on them or not.
               | 
               | One does not form "connections" between them as in some
               | wide-eye conspiracy theorist.. "predictive coding" has
               | little to do with "prediction" in the ML sense. The
               | latter concerned with making a quantitive estimate of
               | some variable by summarising historical data.
               | 
               | What we are doing when we revise a "mental model" is done
               | by counter-factual simulation of possible future states.
               | Statistical AI models conditional probability structures,
               | and computes predictions as an expectation over weighted
               | summarised historical data. This is not a means of
               | performing counter-factual simulation.
               | 
               | One trivial, sadly empirical, way to see this is to note
               | that each marginal token generated is of constant time
               | and energy use. Yet trivially, reasoning and a variety of
               | other mental capacities should require abitarily
               | different time to run. Eg., simulating a complex scenario
               | is necessarily more intensive than a simple one, and so
               | on.
               | 
               | Yet I am quite annoyed that we need such dumb
               | observations to make this point. It speaks of a profound
               | ignorance of "theory of mind, cognition, information
               | theory, physics, philosophy" and especially neurology and
               | zoology which are your most significant missing terms.
               | 
               | It is no real mystery what the structure of various
               | mental capacities involves; nor any mystery what
               | s(Ws(Ws(WX+B)+B...)...)...) computes. Even involving
               | anything beyond trivial applied statistics and trivial
               | results in science shouldnt be required here. This stuff
               | is very obvious.
        
               | FeepingCreature wrote:
               | > Indeed, there is no function which generates text data,
               | there's a very large number of independent generating
               | processes that give rise to the distribution of text. The
               | phrase "the war in ukraine" acquires a different
               | semantics over 2010-2030 in a radically different way
               | than, "I liked that film" does.
               | 
               | But those semantics _are_ revealed in the greater context
               | of the phrase! That 's why it's so important that
               | transformers can attend to large context ranges; that's
               | what lets them learn the greater semantic patterns to
               | begin with. And at the limit, at a scale greater than
               | phrases, I simply reject the idea that the same article -
               | the same _book_ - can have totally different meanings
               | depending on context. Language isn 't just shaped by
               | context, it _shapes_ context itself. Because language
               | cannot be considered without context, it reveals
               | information about that context, and in fact any
               | compression of language ultimately requires modelling the
               | person and even society that produced it. That 's what
               | the network learns.
               | 
               | If your words don't have meaning beyond themselves, what
               | are you even talking about?
        
               | mjburgess wrote:
               | > If your words don't have meaning beyond themselves,
               | what are you even talking about?
               | 
               | ...err... of course? That's the whole point.
               | 
               | "Context" here isnt other _words_. It 's the _world_. The
               | meaning of words is the world.
               | 
               | When I say, "I like what you're wearing" i'm not
               | summarising a history of prior texts; it has nothnig to
               | do with any statistical operation over historical
               | documents. It has entirely to do with what you're
               | wearing.
               | 
               | Langauge use is a side-effect of being embedded in
               | reality, directly attentive to it, and so on. Words are
               | mere symptoms of how we are situated in the world.
               | 
               | LLMs merely replay these back to us. They are not in the
               | world. They cannot, in principle, ever _mean_ , "I like
               | what you're wearing"
        
               | FeepingCreature wrote:
               | I'd say exactly the opposite. It's because human speech
               | is about the world, and LLM speech is about human speech,
               | that LLM speech is about the world, by transitivity. "The
               | ball fell to the <floor>", the LLM predicts, ultimately,
               | not because of any feature of the human brain, or any
               | feature of the English language, but because of
               | _gravity._ That the causal arrow passes through a human
               | brain does not make this any less true! Because human
               | speech is inextricable from physical reality, language
               | models likewise learn to model reality. They learn this
               | poorly, incrementally, making missteps on the way -
               | granted! But that they learn it at all shows that there
               | 's more going on than statistical modelling - or that
               | statistical modelling _of language_ is more than it
               | sounds like.
        
               | canjobear wrote:
               | Languages do in fact have similar distributional
               | structure, so that it is possible to learn how to
               | translate words without supervision:
               | https://arxiv.org/abs/2203.04863
               | 
               | In principle languages are arbitrary, but in practice
               | they're describing the same world and the same concepts
               | end up being useful.
        
               | mjburgess wrote:
               | Those are on embedding vectors, not on words. They use
               | embeddings created from translated sources, eg.,
               | wikipedia articles.
               | 
               | Yes, if you construct an embedding vector on texts-A, and
               | another on texts-B where (A, B) are translations of each
               | other, then an "unsupervised" algorithm really will give
               | you the dizzying heights of a little above coin-flip
               | accuracy on highly engineered self-selected benchmarks.
               | 
               | They evaluate by by looking at the most in-use words in
               | each vocab.. so you take the most in-use words on
               | translations of Wikipedia, whose frequency is decided by
               | the need of translation.. and then you use that to
               | evaluate.
               | 
               | It is blindingly obvious that the structure and frequency
               | of heirglphys on tombs, Chinese glyphs in poetry, and
               | latin in medieval liturgical literature are not
               | distributed by Reality.. written in this order by God so
               | that the Langauge of Reality is what places "d" alongside
               | "oor". We already know this to be the case. The
               | assumption of its opposite is rank pseudoscience.
        
               | FeepingCreature wrote:
               | > It is blindingly obvious that the structure and
               | frequency of heirglphys on tombs, Chinese glyphs in
               | poetry, and latin in medieval liturgical literature are
               | not distributed by Reality.. written in this order by God
               | so that the Langauge of Reality is what places "d"
               | alongside "oor". We already know this to be the case. The
               | assumption of its opposite is rank pseudoscience.
               | 
               | I mean, on the first level, of course they're _entirely_
               | determined by reality in the sense that the human brain
               | is a real, physical object. But also on a second level
               | they 're _still_ entirely determined by reality because
               | the shape of the human _mind_ is also entirely determined
               | by reality. What is the mind _for_ except reflecting
               | reality? What is language for except communicating it?
               | Sure that reality is warped, filtered, reduced and
               | biased, but the data is still in there. That 's in large
               | part why LLMs need such ludicrously large training runs.
               | 
               | I don't think letter frequency is objectively determined,
               | but we know for a fact (many studies!) that the features
               | that large language models learn are far, far above the
               | scale of letters. Even arguing about phrases isn't
               | engaging with the current state of the art.
               | 
               | We're not talking about Markov chains here.
        
               | empath75 wrote:
               | Have you ever actually bothered testing your assumption
               | that chatgpt can't reason. And I don't mean that it fails
               | to reason properly about certain questions, that's
               | trivially easy to show, but is less interesting than many
               | people think it is, because humans can't reason properly
               | about many questions (see, for example, the Monty Hall
               | problem). Test your assumption that it can never reason
               | properly about any question not in its data set. It's way
               | easier, IME, to find examples where it does give the
               | correct answer to novel problems than it is to trip it up
               | with a difficult problem.
        
               | mjburgess wrote:
               | It's data set is, approximately, everything ever written.
               | We have no access to it. And this is widely studied. You
               | can find trivial mistakes in apparent reasoning.
               | 
               | There are two hypotheses: H1, the structure of a response
               | from any given prompt is computed using distributional
               | properties of historical data; H2: the response is
               | computed via deduction from premises to conclusions of
               | agent employing the semantics of the terms, their logical
               | connections, and connections of relevance.
               | 
               | In many cases a prompt/reply will confirm both
               | hypotheses, hence confirmation bias and why we dont
               | bother "confirming" any hypothesis. Rather to choose
               | betweeen them, if you wanted to use data, you just find
               | cases where reasoning fails in such a way that H1 is the
               | more plausible answer. Such cases are easy to find, and
               | across the literature.
               | 
               | This whole thing is pointless however, because it's
               | blindingly obvious from what a statistical AI algorithm
               | does, which is empirical function fiting on historical
               | datasets to approximate conditional probability
               | distributions. This process is extremely well-understood,
               | and we know _necessarily_ that it is just approximating a
               | historical distribution.
        
               | stevenhuang wrote:
               | > This process is extremely well-understood, and we know
               | necessarily that it is just approximating a historical
               | distribution.
               | 
               | Prove that this precludes formation of mind. Prove that
               | human minds do not work this way.
               | 
               | Until such premises are established, your argument is not
               | sound. This leaves open the possibility what we're seeing
               | is in fact a degree of reasoning ability in LLMs.
        
               | mjburgess wrote:
               | As I said in my other comment. Statistical models
               | generate output with constant time/energy. So for _any_
               | input y, and _any_ output x, the computation is constant
               | time. eg., generating _any_ 100 tokens of output takes
               | the same time as any other.
               | 
               | If you reflect on that carefully, you'll note the only
               | way this is possible is if the algorithm does not use any
               | semantic features of the input. Indeed, if it has no
               | relevant capacities at all.
               | 
               | Consider e.g., the prompt, "imagine a world where..., and
               | then infer..., and then what would be... ?" for various
               | "..." we can choose, the output ought take arbitrarily
               | longer to compute; but its constant.
               | 
               | For any word: imagine, reason, suppose, simulate,
               | believe, remember... denoting any alleged mental
               | operation, we can trivially construct cases where
               | actually performing this operation would take more or
               | less time.
               | 
               | "Given premises A, B, C, D,...; infer conclusion..." has
               | a prompt reply which is constant in time regardless of
               | how many premises we input. "Recall memories A, B,C,.."
               | likewise.
               | 
               | The only way that answering "yes" to a question, say,
               | _regardless of what that question is_ , taking the same
               | time/energy/etc. to answer.. would be if no semantic
               | feature of that question was being used in its answering.
               | 
               | This is an LLM, it's all statistical AI: the number of
               | operations performed per prediction is constant _for all
               | predictions_. Almost no alleged capacities are consistent
               | with this fact.
        
               | FeepingCreature wrote:
               | The human brain computes precisely the same amount of
               | information every millisecond. That is, the computational
               | power of a single neuron is necessarily capped by its
               | switching speed.
               | 
               | This proves humans cannot think.
        
               | stevenhuang wrote:
               | The guy you're responding to can be seen routinely
               | interjecting with these kinds of unfounded assertions
               | that LLMs can't reason everywhere LLMs are discussed. He
               | can't help himself it seems. You'd think he'd engage with
               | the responses he receives, maybe down-weight his
               | convictions through exposure to reasonable arguments, but
               | nope.
               | 
               | An LLM would have the wherewithal to consider the
               | possibility they're wrong, or to be agnostic in their
               | beliefs until we ourselves understand better what it
               | means to reason.
        
         | empath75 wrote:
         | Not being able to solve a problem "in general" does not mean
         | that it can't solve specific and useful instances of the
         | problem. There are lots of SAT solvers that can't solve SAT
         | problems in general (in any reasonable amount of time), but
         | nevertheless can solve many useful categories of SAT problems.
        
         | canjobear wrote:
         | People at Microsoft Research certainly do. But the bet is that,
         | with scale, these limitations won't matter in practice. For
         | example Transformers can't recognize well-nested brackets (like
         | {[()]}) to infinite depth; the depth that Transformers can
         | recognize is limited by how many self-attention layers they
         | have. But in practice, you rarely need much depth.
        
           | toxik wrote:
           | Notably, humans also cannot track this to infinite depth.
           | Somehow we know this and reach for algorithmic solutions
           | (like pen and paper).
        
             | logicchains wrote:
             | Interestingly LLMs also do better with pen and paper
             | reasoning: https://arxiv.org/abs/2310.07923 .
        
       | gaudat wrote:
       | Another name hijacked by AI/ML... I was hoping to see some
       | control thoery...
        
         | throwaway42668 wrote:
         | This is what they actually mean by AI will be in everything.
         | Just a relentless campaign of appropriating names that used to
         | have meanings.
        
           | nico wrote:
           | Interesting. This is how I sometimes feel about physics
           | 
           | The field completely appropriated and redefined so many terms
           | of common language, that now it's hard to talk in plain
           | language with someone formally trained in physics about
           | physical phenomena
           | 
           | For example, everyone has some sort of intuitive idea about
           | what Energy is. But if you use that word with a physicist,
           | watch out, for them it means something super specific within
           | the context of assumptions and mathematical models and they
           | will assume you don't know what you are talking about because
           | you are not using their definitions from their models
           | 
           | Same thing happens with infinite in math
        
         | aerospace_guy wrote:
         | This has been happening for years now. Just adding "AI" or
         | "LLM" gets the views and $$ these days.
        
         | szvsw wrote:
         | I mean, if you like state space models, then you should read
         | the paper on Mamba if you haven't already! Because it quite
         | literally uses state spaces... and you will probably think it's
         | a really cool application of state spaces!
         | 
         | Apologies if you know the following already, but maybe others
         | reading your comment feeling similarly will not be familiar and
         | might be interested.
         | 
         | At least intuitively, I like to motivate it this way- pick your
         | favorite simple state space problem. Say a coupled spring
         | system of two masses, maybe with some driving forces. Set it
         | up. Perturb it. Make a bunch of observations at various points
         | in time. Now use your observations to figure out the state
         | space matrices.
         | 
         | There's fundamentally not really anything different (in my
         | opinion) with using Mamba (or another state space model) as a
         | function approximation of whatever phenomenon you are
         | interested in. Okay Mamba has more moving parts, but the core
         | idea is the same: you are saying that on some level, a state
         | space is an appropriate prior for approximation of the dynamics
         | of the quantities of interest. It turns out being pretty
         | remarkable the number of things this can work out quite well
         | for. For instance, I use it to model the 15-min interval data
         | for heating, cooling, and electricity usage of a whole building
         | given 15 min weather data, occupancy schedules, and
         | descriptions of the building characteristics (eg building
         | envelope construction, equipment types, number of occupants,
         | etc).
        
         | duped wrote:
         | It's not hijacked, the formulation is the same. For any layer
         | there is a state-space formulation                  h = Ah + Bx
         | y = Ch + Dx
         | 
         | where x is the input, y is the output, and h is the state. They
         | use "h" instead of "s" for the state variables because they're
         | called "hidden states" in the literature. edit: it is obnoxious
         | they've flipped the convention for A/B/C/D which is the one
         | thing controls people agree on (we can't even agree on the
         | signs and naming of transfer function coefficients!).
         | 
         | Where this diverges from dynamical systems/controls is that
         | they're proving that when x/h/y are represented with finite
         | precision numbers, the model is limited in the problems it can
         | represent (no surprise from controls perspective), and they
         | prove this by using an equivalence to the state-space formulae
         | that's consistent with evaluating it on massively parallel
         | hardware.
         | 
         | The classical controls theory is not super applicable here,
         | because what controls people care about (is the system stable,
         | is its rise/fall time in bounds, what about overshoot, etc) is
         | _not_ what ML researchers care about (what classes of AI
         | problems can be modeled and evaluated using this computational
         | architecture).
        
         | 3abiton wrote:
         | To be fair state-space models come originally from physics
        
       | briandw wrote:
       | The early research in neural networks was hampered by proofs that
       | perceptrons could never solve certain functions, like XOR. DNNs
       | could have been developed much sooner otherwise. I view these
       | proof papers with some skepticism, since they can be
       | unnecessarily dismissive of good ideas.
        
         | nickpsecurity wrote:
         | I've only read the abstract. It says that they have experiments
         | to back their claims. So, that's a proof claim with
         | experimental data. It's in a field where most learning happens
         | by experimental exploration, too.
         | 
         | I don't think it will hold us back. If anything, it's very
         | exciting to see how many people in the ML field are challenging
         | the status quo from many, different angles.
        
         | logicchains wrote:
         | >The early research in neural networks was hampered by proofs
         | that perceptrons could never solve certain functions, like XOR.
         | DNNs could have been developed much sooner otherwise.
         | 
         | These proofs still hold; pure MLPs (without a modern activation
         | function) aren't very useful, both in theory and practice. What
         | made them useful was the realisation that combining them with a
         | proper activation function makes them much more useful, both
         | theoretically and practically; this discovery took time.
        
         | nyrikki wrote:
         | How did papers that demonstrated single layer neural networks
         | can't do XOR, but multi layer neural networks can, hamper
         | development?
         | 
         | XOR is simply not linearly serperatable, requiring an MLP or
         | kernel trick still holds.
         | 
         | It is a similar reason that attention works for majority gates
         | but not parity gates in the general case.
         | 
         | Acknowledging that reality resulted in new developments, but is
         | still a limitation.
         | 
         | Perceptrons are binary classifiers.
        
       | empath75 wrote:
       | There's nothing wrong with this paper and it's important to
       | understand exactly what LLMs (and related technologies) are
       | capable of and what they aren't, but I do have a problem with the
       | general argument that <ai technology> can't do <some task> and so
       | that means it's useless and won't lead to AGI. I feel like a lot
       | of people think that if a single neural network architecture
       | can't achieve something on it's own that building a system that
       | supplements the AI with other tools is somehow _cheating_.
       | General AI if it ever happens is going to very likely be a system
       | of specialized components working together, and there's not going
       | to be a single piece of it that you can point to and say "that is
       | where the intelligence lives" -- it's sort of liking digging
       | through the brain looking for the homunculus that's piloting the
       | ship. If an LLM can't track state, hook it up to something that
       | can. If it doesn't have long term memory and retrieval, hook it
       | up to something that does.
        
         | andoando wrote:
         | I think a way to connect these different systems as a
         | homogenous system is the difficult part.
        
         | szvsw wrote:
         | Very well said. It absolutely drives me crazy - the focus on
         | the model architecture for autoregressive text generation as
         | the "key" (or blocker) to AGI seems so misguided as a debate.
         | It seems clear to me (and maybe I'm wrong!) that a much larger
         | system of modules interacting is essential (actuators, sensors,
         | short term and long term memory, working memory, etc), and I
         | completely fail to understand why it feels like the debate
         | revolves around neural network architectures as opposed to the
         | larger system architecture for agency. The language and
         | decision making component is obviously extremely important and
         | necessary, but it feels to me like it will always just be a
         | necessary but not sufficient component for AGI.
         | 
         | The fact that we have already obviously developed a technology
         | which has an excellent grasp of language, can select between
         | tools, perform knowledge retrieval from external databases,
         | perform some form of simplistic reasoning and planning... it
         | really feels (again, only intuitively!) like everything outside
         | the perimeter of LLMs is where the actual frontier of AGI will
         | be.
         | 
         | It feels like with enough sensorimotor affordances and
         | mechanisms for saving and accessing memory, all that is left is
         | the (obviously complex) framework for connecting all of these
         | together, with potentially many instances of each module
         | working together... especially when you consider that one such
         | system can potentially use human actors as tools (eg "hello
         | fellow human, please complete this captcha for me, as I am low-
         | vision").
         | 
         | I guess more succinctly - I do not believe that GPT-4 et al are
         | remotely close to AGI on their own, _but_ I do feel (again,
         | this is just "vibes," and obviously people smarter than me
         | disagree!) that GPT-4 could be a component combined with lots
         | of other tech we already have to essentially achieve AGI
         | already. Perhaps what I am describing would just be a pale
         | imitation of what others mean by AGI and be reductively called
         | another mechanical Turk. I am probably the misguided one, but
         | nonetheless, I can't escape the feeling that "the one model to
         | rule them all" is a red herring vis-a-vis a complex network
         | (ha) of modules interacting together to achieve meaningful
         | agency. Maybe it's just that I am more interested in
         | "meaningful agency" as a guiding principle over "AGI."
        
           | genevra wrote:
           | Exactly, we're already adding basic memory functions to GPT
           | models
           | 
           | Text generation can only get so much better but the other
           | modules required to simulate intelligence have tons of room
           | for improvement
        
         | logicchains wrote:
         | >the general argument that <ai technology> can't do <some task>
         | and so that means it's useless and won't lead to AGI
         | 
         | The paper isn't making that argument. The point it makes is
         | that parallelisable SSMs are theoretically no more powerful
         | than transformers, contrary to some people's assumption that
         | they'd be theoretically equivalent to RNNs and hence more
         | efficient at certain kinds of problems.
        
       | Lerc wrote:
       | Reading that abstract, I'm not sure what is being claimed here.
       | They say "the "state" in common SSMs is an illusion:" Is this a
       | claim that the activations within a state space model contain no
       | state or that it cannot contain an arbitrary state. The former
       | seems trivially false and the latter seems completely
       | uncontroversial.
        
         | logicchains wrote:
         | If you read the rest of the paper it's clear what they mean.
         | SSMs are theoretically no more powerful than transformers,
         | meaning without chain of thought they can't efficiently solve
         | state-tracking problems, unlike RNNs:
         | https://arxiv.org/abs/2207.00729 .
        
           | opprobium wrote:
           | Not just efficiently, can't solve.
        
             | logicchains wrote:
             | They can solve it if you keep adding layers to the
             | transformer, it's just not efficient; you'd need
             | exponentially more layers than a similarly sized RNN.
        
           | yobbo wrote:
           | The difference to RNNs is that the state in SSMs is a linear
           | combination of previous inputs, which makes it possible to
           | parallelise training in various ways.
           | 
           | Gates and non-linearities in RNNs allow the state to be "any"
           | function of previous inputs.
        
       | duped wrote:
       | I disagree with the editorialized abstract. State is not an
       | illusion. There is a limit to what stateful LTI systems are
       | capable of modeling. It is very interesting that they've been
       | able to prove what that limit is in the context of computation.
       | 
       | From my reading, "non gated SSMs" are LTI systems. They're linear
       | (the paper ignores the activation function) and the "does not
       | depend on input" criteria is the controls way of saying "time
       | invariant" (replace time with the domain of the input vector). It
       | is not surprising that a system that cannot adapt to changing
       | input conditions (by construction) is fundamentally limited to
       | the problems it can model. This is why adaptive filters are
       | studied - LTI systems are useful computationally, but limited
       | because they cannot adapt. Relaxing the "TI" constraint greatly
       | expands the domains of problems that can be solved.
       | 
       | When the feedback matrix is diagonal you have an FIR system.
       | While "finite impulse response" has as specific mathematical
       | definition, conceptually it means the state at time n depends
       | solely on past inputs and states n - N where N is the size of the
       | state vector. So of course, if the feedback matrix is diagonal,
       | the system is limited in the kinds of stateful problems it can
       | handle.
       | 
       | What this paper is missing is that connection to the limits of
       | LTI and connection to TC0, which is very interesting indeed.
       | 
       | In other words, the state of the network is still infinitely long
       | and definitely not an illusion. However if the problem _requires_
       | the network to adapt to its input then a non-gated SSM is not
       | going to be sufficient. That 's an interesting research space.
        
       ___________________________________________________________________
       (page generated 2024-06-07 23:01 UTC)