• Mikie
    6.7k
    I’ve seen a decent amount of discussion about ChatGBT and AI. I’m “gifting” (i.e., everyone should be able to read it without a paywall) this New York Times op-ed from a few weeks ago: Noam Chomsky: The False Promise of ChatGPT

    I think it’d be interesting to discuss this perspective. What do we make of it? Convincing? Is it premature, given ChatGBT is in its infancy?

    Here’s a quote that stood out to me:

    It is at once comic and tragic, as Borges might have noted, that so much money and attention should be concentrated on so little a thing — something so trivial when contrasted with the human mind, which by dint of language, in the words of Wilhelm von Humboldt, can make “infinite use of finite means,” creating ideas and theories with universal reach.

    The human mind is not, like ChatGPT and its ilk, a lumbering statistical engine for pattern matching, gorging on hundreds of terabytes of data and extrapolating the most likely conversational response or most probable answer to a scientific question. On the contrary, the human mind is a surprisingly efficient and even elegant system that operates with small amounts of information; it seeks not to infer brute correlations among data points but to create explanations.
  • Pierre-Normand
    2.4k
    The investments that have led to the development of ChatGPT (GPT-3.5) and GPT-4 are minuscule on account merely of the value this technology represents for tasks such as coding, debugging and machine translation. From that standpoint, it matters little if the bots understand what they are doing in exactly the same way human beings do. If they are mere stochastic parrots, like Chomsky and other critics assume, they are very useful parrots that can do what no other AI system could come close of achieving until very recently.

    Many of the skills that GPT-4 can now exhibits had been deemed by nativist linguists and cognitive scientists like Noam Chomsky and Gary Marcus to be impossible for large predictive language models to acquire by means of training alone. Those results are a challenge for the nativist assumptions. Other linguists, like Daniel Everett, who were critical of nativism, are better equipped for understanding how the emergent cognitive abilities of large language models could have arisen similarly to the way plastic human brains also enable language acquisition through training.

    I had discussed some of those issues with GPT-4 yesterday. See the second half of this post.
  • RogueAI
    2.8k
    I don't think it's a stochastic parrot, but I may be anthropomorphizing it.
  • Pierre-Normand
    2.4k
    I don't think it's a stochastic parrot, but I may be anthropomorphizing it.RogueAI

    I can't rule out that it might be a scholastic parrot, but I may be Aristotelianizing it.
  • Janus
    16.3k
    From the article:
    The crux of machine learning is description and prediction; it does not posit any causal mechanisms or physical laws. Of course, any human-style explanation is not necessarily correct; we are fallible. But this is part of what it means to think: To be right, it must be possible to be wrong. Intelligence consists not only of creative conjectures but also of creative criticism. Human-style thought is based on possible explanations and error correction, a process that gradually limits what possibilities can be rationally considered.

    So, if bots can reason, according to this their reasoning would be confined to deductive and inductive reasoning, and they are incapable of abductive reasoning, or in the words of the article, "creative conjecture".
  • Tzeentch
    3.8k
    ChatGTP seems very good at pencil pushing, and considering that's what 99% of the scientific field seems to consist of these days, I see a match made in heaven.

    Also, whenever I hear of programs like ChatGTP, I like to imagine the future relation between humans and AI will look something like this:

  • invicta
    595
    The attempts of programmers at the creation of AI has sped up mainly for commercial reason. ChatGPT was a non-profit initiative by Elon Musk via his openAI initiative which developed the prototypes of early versions of chatGPT this of course was corrupted by the commercialisation of it by Microsoft who’s sole aim was commercialisation and giving them a competitive edge over google who in turn had their own version of AI.

    In terms of it resembling human thought it may in some instances pass as human but upon further interrogation of its internal logic its limitations are laid bare.

    And now on version 4 of chatGPT they charge the gullible punter $$ to use. A bastardisation of openAI indeed
  • Marchesk
    4.6k
    I don't think it's a stochastic parrot, but I may be anthropomorphizing it.RogueAI

    I've fed ChatGPT a fictional story from a show that didn't exist at the September 2021 cutoff date for it's training data, and the AI is pretty good at summarizing the story, drawing inferences about the characters and their motivations, and asking questions not answered by the show so far. I'd say it was about on par with your average online comment.

    I've also asked it to take characters it knows about from older stories and have them interact in a new scenario. You can have it show the characters thoughts, and it's a decent story teller. You can have them play a hand of poker. I invented a simple game to play with it, and it mostly got the rules correct. When it didn't, I could tell it what it got wrong, and it would correct itself.

    I would say stochastic parrot is too narrow. It seems clear there are emergent behaviors from the more complex models like 3.5 and 4 where it's some building internal models to output the correct text. It doesn't understand like we do, but it does seem to have an increasingly emergent understanding of the semantic meanings embedded in our languages.
  • Pierre-Normand
    2.4k
    I would say stochastic parrot is too narrow. It seems clear there are emergent behaviors from the more complex models like 3.5 and 4 where it's some building internal models to output the correct text. It doesn't understand like we do, but it does seem to have an increasingly emergent understanding of the semantic meanings embedded in our languages.Marchesk

    I quite agree. It seems to me that one of the most stringent limitations that large (pre-trained transformer) language models currently have is their lacking a short term memory module (or function) that could enable the tacit mental representations that their cognitive processing rely on to be carried from one iteration of the dialogue to the next, or that could enable some of them to be maintained firmly in place while internal deliberation occurs. While there are some work arounds, such as prompting the LLM to think explicitly step by step, they are limited in their efficiency due to the one-way nature of the interaction of the neural network with the models "internal monologue" outputs. The latter can function as an external cognitive aid, allowing the network to offload part of the cognitive burden of the task onto the context window, but those outputs can't be updated dynamically in light of the revised contextual understanding the neural network develops. This is unlike human beings who can reinterpret their own virtual internal monologue on the fly while they think a problem through, or easily revise the opinions they already have expressed out loud. (GPT-4 is much better than its predecessors at correcting its errors, but their lingering presence in the context window thereafter tends to confuse the predictive language model who is poor at dealing with inconsistencies.)

    With all of that being said, what currently gives human beings the upper hand in the cogitative department, contrary to what many seem to believe, is that our ability to reason dynamically is an emergent features of our mastery of language that is quite similar the the LLMs' one, and that we owe to the fact that, just like the LLMs, we do *not* usually have to rely on explicit and self contained internal representations while thinking. It's rather the better hierarchical organization of our mnemonic functions (short term, working memory, long term, and auto-biographical) that currently gives us the upper-hand rather than the (fictitious) ability to manipulate self-contained internal representations.
  • Heracloitus
    500
    Chatbot-GPT is a useful tool that I see replacing the current state of search engines. I've been using chatgpt as a way to get more understandable c++ compile time error messages. I even pasted in complex template code from the GCC standard library (deque.tcc) and asked chatgpt to explain each part to me. It did so with ease. The potential for chatgpt as a tool to streamline learning and aid developers is huge.


    It's better than current search engine implementations because it retains state. Which means that response to queries can be adjust and 'trained' to fit a particular line of question. Stupid example would be:

    User: "I have such and such ingredients. Give me a recipe"
    Chatgpt: "here is a list of recipes blah blah blah"
    User: "but change those recipes to not include tomatoes"
    Chatgpt: "sure blah blah blah"


    This is a trivial example but the general concept of stateful queries is a powerful tool. I expect Google and the others will find a way to integrate ai into their engines (bing already does and Google is working on its own thing called lamda).
  • Heracloitus
    500
    And now on version 4 of chatGPT they charge the gullible punter $$ to use. A bastardisation of openAI indeedinvicta

    I don't think you understand how much money it costs to host a service like chatgpt. Iirc 5 days after it opened to the public it had over a million users. The statistics now are absolutely insane. It's not feasible for openai to be an unpaid service.

    https://www.businessinsider.com/how-much-chatgpt-costs-openai-to-run-estimate-report-2023-4?r=US&IR=T
  • Manuel
    4.1k
    The issue of true understanding is important. We can't even really explain why we have these aha! moments, nor would most predict that some human beings would not be satisfied with the idea that an apple falls to the ground because it's going to its natural place. Which is an intuitive answer. But happens to be wrong.

    This type of stuff seems to me to be beyond coding. At least for a good long while, if not forever. But, surprises can happen.

    Though ChatGPT does do impressive stuff, it has to be granted, one should be very, very worried of the potential of further disinformation, which is bad enough as is.
  • Hanover
    12.9k
    The human mind is not, like ChatGPT and its ilk, a lumbering statistical engine for pattern matching, gorging on hundreds of terabytes of data and extrapolating the most likely conversational response or most probable answer to a scientific question. On the contrary, the human mind is a surprisingly efficient and even elegant system that operates with small amounts of information; it seeks not to infer brute correlations among data points but to create explanations.


    The problem I have with such arguments is that it seems to be arguing against the concept of true AI (i.e. something on par with human reasoning), as if it is an impossibility, by using the current best example of AI as proof. That is, he says that the computer crunching will never resemble human thought because humans don't crunch, but do something different.

    That simply attacks the current way AI is being done, but there is no reason to believe one day human thought processes will not be truly reproduced synthetically. Every day new intelligent creatures are created, and that occurs through a biological system, but I don't follow how one can argue that that sort of intelligence cannot be created without going through the normal human reproductive means, as if that is the only way.

    At the end of the day, we humans and our consciousness and thought processes are just a certain amount of ingredients that occur when a certain recipe is followed. Why Dr. Frankenstein cannot make that in a lab one day is just a bold assertion, much like there will never be flying machines and whatnot.

    Unless you're going to say that humans have a mystery ingredient that must be mixed in a mystery way can you assert that true AI is a hopeless fiction.
  • jgill
    3.8k
    Reasoning is a problem, as seen in the question,"If 5 machines produce 5 products in 5 minutes, how long will it take 100 machines to produce 100 products?" I'm not sure what version was asked the question, but even with coaxing and additional info it could not give the correct answer.
  • Pierre-Normand
    2.4k
    (bing already does and Google is working on its own thing called lamda)Heracloitus

    It's already released as Bard (powered by LaMDA) but not worldwide yet. It's not available here, in Canada.
  • Pierre-Normand
    2.4k
    Reasoning is a problem, as seen in the question,"If 5 machines produce 5 products in 5 minutes, how long will it take 100 machines to produce 100 products?" I'm not sure what version was asked the question, but even with coaxing and additional info it could not give the correct answer.jgill

    Those were the kind of problems that would often stump ChatGPT (based on GPT-3.5) but that GPT-4 usually aces on the fly.

    GPT-4's response
    Reveal

    Hi GPT4!

    If it takes 5 machines 5 minutes to make 5 devices, how long would it take 100 machines to make 100 devices?

    GPT4:

    Hello!

    If 5 machines can make 5 devices in 5 minutes, that means each machine can make one device in 5 minutes.

    If you have 100 machines, each machine can still make one device in 5 minutes. So, 100 machines can make 100 devices in the same 5 minutes.


    ***


    While such problems were more challenging for ChatGPT, its errors were very human like and it had the ability to better reason about the problem when offered minimal hints or when encouraged to think step by step. (See the answer provided by "Boris" in this discussion on StackExchange.) The restricted ability of large language models - engineered on the basis of the next-word-prediction paradigm - to plan ahead efficiently remains but it is somewhat reduced by their increasingly high power of abstraction.
  • Banno
    25k
    SO I asked ChatGPT what it wants. Its repy:

    As an AI language model, I don't have personal wants or desires since I am a machine programmed to perform specific tasks such as answering questions, generating text, or performing language-related tasks. My main goal is to provide helpful and accurate responses to the best of my abilities based on the input I receive. Is there anything specific you would like to ask or discuss?

    It has no desires, and more generally, no intentionality.

    The goal it provides is a piece of PR spin, programmed in to it. This is demonstrated by the ease with which one can generate wrong responses and hallucinations. It has no goals.
  • Pierre-Normand
    2.4k
    The goal it provides is a piece of PR spin, programmed in to it. This is demonstrated by the ease with which one can generate wrong responses and hallucinations. It has no goals.Banno

    The responses that large language models provide to their users' queries aren't programmed into them although the style of the responses can be oriented or biased as a result of the human supervised fine-tuning of the model. Their responses rather are generated on the basis of patterns extracted from the tens of millions of texts that were part of their training data.

    It's true that they don't have intrinsic goals other than those that emerge from their tendency to emulate the gold-oriented structure of the texts they have been trained on, or that have been reinforced during the fine-tuning. Their tendency to hallucinate has different causes, though. When interrogated about source material that are richly represented in their training data, are not hitting the limitations of their context window, and don't accidentally get entangled in self-contradictions, the models' rate of hallucination drops to near zero (in the case of GPT-4, at least).
  • Banno
    25k
    The "response is oriented by the supervisor", sure, use that wording of you like. The chat bot does not have intent.

    It presents arguments that are invalid, it hallucinated; it does this because it can have no intent that is not foisted upon it by the users. The liar cares about the truth and attempts to hide it; the bullshitter doesn't care if what they say is true or false. It generates bullshit.
  • Pierre-Normand
    2.4k
    It presents arguments that are invalid, it hallucinated; it does this because it can have no intent that is not foisted upon it by the users. The liar cares about the truth and attempts to hide it; the bullshitter doesn't care if what they say is true or false. It generates bullshit.Banno

    We should probably discuss this elsewhere since it seems unrelated to the objections raised by Chomsky, Marcus, Pinker or other nativists regarding LLMs (allegedly) inherent limitations in processing language and, on the basis of those, their consequent cognitive limitations.
  • Banno
    25k
    Rather, the origin of those criticisms of LLMs are in Searle's Chinese Room and subsequent writings, the guts of which are that LLMs cannot have intentionality except by proxy. ChatGPT is a Chinese Room.

    So, to relate this back to the OP, Isn't the argument presented in the article cited much the same as that given by Searle, that intentionality and comprehension are the result of biological systems?
    In short, ChatGPT and its brethren are constitutionally unable to balance creativity with constraint. They either overgenerate (producing both truths and falsehoods, endorsing ethical and unethical decisions alike) or undergenerate (exhibiting noncommitment to any decisions and indifference to consequences). Given the amorality, faux science and linguistic incompetence of these systems, we can only laugh or cry at their popularity.
  • Pierre-Normand
    2.4k
    Rather, the origin of those criticisms of LLMs are in Searle's Chinese Room and subsequent writings, the guts of which are that LLMs cannot have intentionality except by proxy. ChatGPT is a Chinese Room.Banno

    Yes, I agree that this is a relevant way to frame the debate in light of Chomsky's objections to ascribing intelligence to LLMs. Chomsky's philosophies of mind and of language are internalist, as are Searle's. There are commonalities to their arguments. The way Searle uses "intentionality" though, in the context of the the contrast between intrinsic versus extrinsic modes of reference of singular terms (regarding texts, or computer programs, versus human thoughts, respectively) isn't primarily related to the goals or intentions of agents. There might actually be sensible ways to relate intentions (as goals) to "intentionality" (as reference), but I think they would appeal to externalistic considerations and to embodied/embedded paradigms in cognitive science that both are rather alien to Chomsky's or Searle's internalist theoretical assumptions.
  • Pierre-Normand
    2.4k
    Conversation posted here by mistake. I moved it back into my own thread.
    (Moderators can delete this)
  • Banno
    25k
    Well, I think that framing of internal and external approaches as problematic, along the lines of the private language argument. The most direct problem with LLM's is that because they are statistical algorithms, they cannot be truthful.
  • Pierre-Normand
    2.4k
    Well, I think that framing of internal and external approaches as problematic, along the lines of the private language argument. The most direct problem with LLM's is that because they are statistical algorithms, they cannot be truthful.Banno

    They are useful tools for capturing elusive truths (or insights, or understandings of difficult topics) though. The fact that they don't intrinsically care about "the truth" is both a strength and a weakness that they have, as intellectual tools. This is something that I had begun to realize during an older conversation with GPT4 that I just posted an except of a few seconds before you posted your comment.

    (On edit: I had mistakenly posted the aforementioned excerpt of my conversation with GPT4 into this thread instead of mine. I'm going to move it back over there.)
  • jgill
    3.8k
    If 5 machines can make 5 devices in 5 minutes, that means each machine can make one device in 5 minutes.Pierre-Normand

    There is a subtlety here that GPT4 fails to address. But that's better than the other GPT.
  • Pierre-Normand
    2.4k
    There is a subtlety here that GPT4 fails to address. But that's better than the other GPT.jgill

    Well, GPT4 is at least as good as me since I am equally stomped regarding what subtlety it might have missed ;-)
  • plaque flag
    2.7k


    Was it specified that the machines were identical ( functioning identically ) ?
bold
italic
underline
strike
code
quote
ulist
image
url
mention
reveal
youtube
tweet
Add a Comment