4 ms·

I was thinking last night about LLMs with respect to Wittgenstein after watching this interesting discussion of his philosophy by John Searle [1]. I think Witt

I was thinking last night about LLMs with respect to Wittgenstein after watching this interesting discussion of his philosophy by John Searle [1]. I think Wittgenstein's ideas are pertinent to the discussion of the relation of language to intelligence (or reasoning in general). I don't meant this in a technical sense (I recall Chomsky mentioning that almost no ideas from Wittgenstein actually have a place in modern linguistics) but from a metaphysical sense (Chomsky also noted that Wittgenstein was one of his formative influences). The video I linked is a worthy introduction and not too long so I recommend it to anyone interested in how language might be the key to intelligence. My personal take, when I see skeptics of LLMs approaching AGI, is that they implicitly reject a Wittgenstein view of metaphysics without actually engaging with it. There is an implicit Cartesian aspect to their world view, where there is either some mental aspect not yet captured by machines (a primitive soul) or some physical process missing (some kind of non-language system). Whenever I read skeptical arguments against LLMs they are not credibly evidence based, nor are they credibly theoretical. They almost always come down to the assumption that language alone isn't sufficient. Wittgenstein was arguing long before LLMs were even a possibility that language wasn't just sufficient, it was inextricably linked to reason. What excites me about scaling LLMs, is we may actually build evidence that supports (or refutes) his metaphysical ideas. 1. https://www.youtube.com/watch?v=v_hQpvQYhOI&ab_channel=PhilosophyOverdose https://www.youtube.com/watch?v=v_hQpvQYhOI&ab_channel=Philo...

throwup238 3y ago
Wittgenstein is the perfect lens through which to be skeptical about LLMs and AGI and I think you may be the one not fully engaging with his work. He saw that languages are inseparable from the context in which they are used and that context is much bigger than language itself. Part of learning those language games is experimenting in the real world - interacting with other people, playing language games, and seeing how they react is how we build up our internal dictionaries and innate knowledge. The fidelity of text is simply too low to communicate the amount of information humans use to build up general intelligence. Without the ability to interact with the physical world, LLMs will never be able to reach AGI. It can kind of simulate it to some extent, but it'll never get there LLMs can't even form their own memories, the context has to be explicitly fed back to them.
- zoogeny 3y ago
  > He saw that languages are inseparable from the context in which they are used That is one of the things that stood out to me in Searle's summary of his later work because I consider how the transformer architecture works and the way in which the surrounding context plays into the meaning of the words. > Part of learning those language games is experimenting in the real world It is interesting that the article we are responding to talks about how we have only just begun to experiment with RL on top of transformers. In the same way that Alpha Go engaged in adversarial play we can envision LLMs being augmented to play language games amongst themselves. That may result in their own language, distinct from human language. But it also may result in the formation of intelligence surpassing human intelligence. > The fidelity of text is simply too low to communicate the amount of information humans use to build up general intelligence. This does not at all follow from anything I've encountered in Wittgenstein. It is an empirical claim that we (as in humanity) are going to test and not something that I would argue either one of us can know simply reasoning from first principles. What does follow for me is closer to what Steven Pinker has been proposing in his own critiques of LLMs and AGI, which is that there is no necessary correlation between goal seeking (or morality) and intelligence. I also feel this is concordant with Wittgenstein's own work. > Without the ability to interact with the physical world, LLMs will never be able to reach AGI Again, a confident claim that is based on nothing other than your own belief. As I stated in my last comment, I am excited to see if that is empirically true or false. We are definitely going to scale up LLMs in the coming decade and so we are likely to find out. My suspicion is that people don't want this scaling up to work because it would force them to let go of metaphysical commitments they have on both the nature of intelligence as well as the nature of reality. And for this reason they are adamantly disbelieving in even the possibility before the evidence has been gathered. I'm happy to stay agnostic until the evidence is in. Thankfully, it shouldn't take too long so I may be lucky enough to find out in my own lifetime.
  - throwup238 3y ago
    It's not a belief or logically derived claim, it's as close to empirical fact as we're going to get. We have zero evidence that intelligence without physical experimentation is possible because we have no other examples of intelligence except humans (and nonhuman animals), all of whom learned experimentally with a physical feedback loop. Even the most extreme cases like Helen Keller depended on it - her story is perhaps far more useful to grounding theories about AGI than any philosophical text as Wittgenstein himself would likely argue (Water!). His contempt, for lack of a better word, for philosophy on those terms is clear. I'm excited to see how LLMs scale but it won't reach AGI without a much richer architecture that is capable of experimentation, capable of playing "language games" with other humans and remembering what it learned. (I'm fairly certain of my views given my experience in neuroscience but it's fun to talk Wittgenstein in the context of LLMs, something that's been conspicuously missing. Sadly I don't believe discussions of AGI are fruitful, just what LLMs can teach us about the nature of language)
daepyonim 3y ago
> Wittgenstein view of metaphysics lmao. I'd like to see you elaborate on what you think this means! The fact that you quote Searle, though, tells me all I need to know.
- zoogeny 3y ago
  Wittgenstein seemed most interested in the question of how to categorize logical propositions into two categories: those that were sensible and those that were nonsense. I don't know how he explicitly described that process of categorization but I think it is fair to say that an agent that can categorize logical propositions into those two categories could be described as intelligent. In his early work he appeared to believe that the way to do this categorization was to determine if the logical propositions corresponded in their structure to the real world (his picture theory of language). In his later work he appears to have believed the way to do this was to reference some "language game" among a set of communicating agents and then understand the words based on their context within that construct (use theory of language). Both of these concepts do not reference any kind of subjective experience, aesthetics, morality, etc. They are simply a description of how to judge whether or not a proposition is "meaningful". In later Wittgenstein, meaning in language is entirely divorced from any necessity that the statements correspond to an independent objective reality. It is simply necessary that they are consistent amongst the participants in a particular language game. I don't believe that means the entirety of our "consciousness" is related solely to our ability to categorize logical propositions. However, it may suggest that intelligence specifically is related to this ability. In as much as we can say that an LLM is capable of participating in some particular human language game and can successfully categorize logical propositions within that language game - I would say that LLM is demonstrating "intelligence" within that language game. And if we can create LLMs that can participate in arbitrary (or general) language games across a wide variety of domains, we might call that LLM generally intelligent. I believe that current LLMs have achieved the first (demonstrating some intelligence in particular domains) but we have yet to achieve the second (demonstrating consistent intelligence in a wide range of general domains). As for metaphysics, I would argue that Wittgenstein saw this general ability (to categorize logical propositions) as a subset of all possible experience. I believe he saw this categorization activity as the primary aim of philosophy. However, the kinds of experience that were outside of this categorization activity could not be spoken about at all.

throwup238 3y ago

Wittgenstein is the perfect lens through which to be skeptical about LLMs and AGI and I think you may be the one not fully engaging with his work. He saw that languages are inseparable from the context in which they are used and that context is much bigger than language itself. Part of learning those language games is experimenting in the real world - interacting with other people, playing language games, and seeing how they react is how we build up our internal dictionaries and innate knowledge. The fidelity of text is simply too low to communicate the amount of information humans use to build up general intelligence. Without the ability to interact with the physical world, LLMs will never be able to reach AGI. It can kind of simulate it to some extent, but it'll never get there LLMs can't even form their own memories, the context has to be explicitly fed back to them.

zoogeny 3y ago
> He saw that languages are inseparable from the context in which they are used That is one of the things that stood out to me in Searle's summary of his later work because I consider how the transformer architecture works and the way in which the surrounding context plays into the meaning of the words. > Part of learning those language games is experimenting in the real world It is interesting that the article we are responding to talks about how we have only just begun to experiment with RL on top of transformers. In the same way that Alpha Go engaged in adversarial play we can envision LLMs being augmented to play language games amongst themselves. That may result in their own language, distinct from human language. But it also may result in the formation of intelligence surpassing human intelligence. > The fidelity of text is simply too low to communicate the amount of information humans use to build up general intelligence. This does not at all follow from anything I've encountered in Wittgenstein. It is an empirical claim that we (as in humanity) are going to test and not something that I would argue either one of us can know simply reasoning from first principles. What does follow for me is closer to what Steven Pinker has been proposing in his own critiques of LLMs and AGI, which is that there is no necessary correlation between goal seeking (or morality) and intelligence. I also feel this is concordant with Wittgenstein's own work. > Without the ability to interact with the physical world, LLMs will never be able to reach AGI Again, a confident claim that is based on nothing other than your own belief. As I stated in my last comment, I am excited to see if that is empirically true or false. We are definitely going to scale up LLMs in the coming decade and so we are likely to find out. My suspicion is that people don't want this scaling up to work because it would force them to let go of metaphysical commitments they have on both the nature of intelligence as well as the nature of reality. And for this reason they are adamantly disbelieving in even the possibility before the evidence has been gathered. I'm happy to stay agnostic until the evidence is in. Thankfully, it shouldn't take too long so I may be lucky enough to find out in my own lifetime.
- throwup238 3y ago
  It's not a belief or logically derived claim, it's as close to empirical fact as we're going to get. We have zero evidence that intelligence without physical experimentation is possible because we have no other examples of intelligence except humans (and nonhuman animals), all of whom learned experimentally with a physical feedback loop. Even the most extreme cases like Helen Keller depended on it - her story is perhaps far more useful to grounding theories about AGI than any philosophical text as Wittgenstein himself would likely argue (Water!). His contempt, for lack of a better word, for philosophy on those terms is clear. I'm excited to see how LLMs scale but it won't reach AGI without a much richer architecture that is capable of experimentation, capable of playing "language games" with other humans and remembering what it learned. (I'm fairly certain of my views given my experience in neuroscience but it's fun to talk Wittgenstein in the context of LLMs, something that's been conspicuously missing. Sadly I don't believe discussions of AGI are fruitful, just what LLMs can teach us about the nature of language)

daepyonim 3y ago

> Wittgenstein view of metaphysics lmao. I'd like to see you elaborate on what you think this means! The fact that you quote Searle, though, tells me all I need to know.

zoogeny 3y ago
Wittgenstein seemed most interested in the question of how to categorize logical propositions into two categories: those that were sensible and those that were nonsense. I don't know how he explicitly described that process of categorization but I think it is fair to say that an agent that can categorize logical propositions into those two categories could be described as intelligent. In his early work he appeared to believe that the way to do this categorization was to determine if the logical propositions corresponded in their structure to the real world (his picture theory of language). In his later work he appears to have believed the way to do this was to reference some "language game" among a set of communicating agents and then understand the words based on their context within that construct (use theory of language). Both of these concepts do not reference any kind of subjective experience, aesthetics, morality, etc. They are simply a description of how to judge whether or not a proposition is "meaningful". In later Wittgenstein, meaning in language is entirely divorced from any necessity that the statements correspond to an independent objective reality. It is simply necessary that they are consistent amongst the participants in a particular language game. I don't believe that means the entirety of our "consciousness" is related solely to our ability to categorize logical propositions. However, it may suggest that intelligence specifically is related to this ability. In as much as we can say that an LLM is capable of participating in some particular human language game and can successfully categorize logical propositions within that language game - I would say that LLM is demonstrating "intelligence" within that language game. And if we can create LLMs that can participate in arbitrary (or general) language games across a wide variety of domains, we might call that LLM generally intelligent. I believe that current LLMs have achieved the first (demonstrating some intelligence in particular domains) but we have yet to achieve the second (demonstrating consistent intelligence in a wide range of general domains). As for metaphysics, I would argue that Wittgenstein saw this general ability (to categorize logical propositions) as a subset of all possible experience. I believe he saw this categorization activity as the primary aim of philosophy. However, the kinds of experience that were outside of this categorization activity could not be spoken about at all.