3 ms·

It's ironic that an author who takes their moniker after Wittgenstein would espouse a view that Wittgenstein's own later work (the investigations) practically a

It's ironic that an author who takes their moniker after Wittgenstein would espouse a view that Wittgenstein's own later work (the investigations) practically argues directly against. The perspective presented here is soaked through and through with information theoretic bias and clearly stems from an overly digital consideration of human experience. "Intelligence" is intimately related to context, use, community, and goals. We don't call someone intelligent because we can point to some lossless compression algorithm implemented by his neurons, we call him intelligent when he produces the behaviors we desire in a given situation. Consider for instance, how this theory fails to account for athletic knowledge. It seems fair to state that people with tangible skills requiring the use of the body posses some kind of intelligence, it seems less accurate to try and describe this form of intelligence as lossless compression. The compression metaphor really only works for intelligence with respect to the manipulation of symbolic representations and even then I think there are plenty of counter examples of what we'd call intelligent behavior that would suggest not all cases are reducible to "lossless compression". It is all highly dependent on the questions you ask, as Wittgenstein well knew. Personally, I do not feel this is a good piece of philosophy and it's symptomatic of a trend toward "computormorphization" which has been rampant since the computer emerged (it's like anthropomorphism except interpreting the behavior of other (actually living things!) as though they behaved just like computers). You can also see this in the reification of concepts like information (we talk about information as though it were some objective material object, but information is not a substance--only interpreters produce information; a text is a vehicle for information but it does not "contain information"--the information is produced by the reader of the text and fully depends on what distinctions they deem relevant)

canjobear 4y ago
I don't think the information-theoretic compression-based view is at all incompatible with Wittgenstein in Philosophical Investigations, although it's absolutely against the view in the Tractatus. In the Tractatus, something is meaningful when it is isomorphic to some structured symbolic representation. Information theory is based on a rejection of that kind of view (from Shannon 1948): "The fundamental problem of communication is that of reproducing at one point either exactly or approximately a message selected at another point. Frequently the messages have meaning; that is they refer to or are correlated according to some system with certain physical or conceptual entities. These semantic aspects of communication are irrelevant to the engineering problem. The significant aspect is that the actual message is one selected from a set of possible messages." In other words, information has nothing to do with isomorphism to a structured representation with definite relationships with the world, nothing to do with symbolic representations or any other kind of "representations". It's simply the ability to predict one thing from another thing. Shannon calls these things "messages" but mathematically the logic applies to any set: for example we can analyze an agent's policy information-theoretically, in terms of how much information about the agent's input state is contained in its output actions. I think this kind of view is highly compatible with later Wittgenstein, for whom "meaning" (if it's anything) is an abstraction of how language is used to coordinate joint action---that is, situations where people are trying to predict and control each others' behavior. To address your examples, > We don't call someone intelligent because we can point to some lossless compression algorithm implemented by his neurons, we call him intelligent when he produces the behaviors we desire in a given situation. Suppose we give someone input A and he outputs desired behavior f(A), and then we give him input B and he outputs desired behavior f(B). If we give him input C will he output the desired f(C)? Only if he has learned to predict what f(C) should be as a function of C and the training data, that is, if he has learned to predict what we consider the "desired" behavior to be. In that sense, intelligence here requires prediction, which is exactly the same thing as compression. > It seems fair to state that people with tangible skills requiring the use of the body posses some kind of intelligence, it seems less accurate to try and describe this form of intelligence as lossless compression. Again, the agent's problem here is to select the right motor action given sensory input. Suppose you've seen a tennis ball coming in with velocity vector x and you know how to hit it to win the game, and you've seen a tennis ball coming in with vector y and you know how to hit it to win the game. What will you do with vector z? If your policy is good, it will put high probability on only actions that hit the ball in a way that makes you win. And if your policy is putting high probability on that action, that's mathematically the same as saying it provides an efficient compression of that action (high probability = small number of bits in an encoding). This is very abstract but once you adopt this view it gives you a lot of very useful conceptual tools. For example, you can talk about the "channel capacity" of a policy in terms of how many bits of information it can "transmit" from input states to output actions, and this channel capacity turns out to be a very intuitive measure of the complexity of the policy that you can use to analyze human behavior (one example, [1]). [1] https://gershmanlab.com/pubs/GershmanLai21.pdf https://gershmanlab.com/pubs/GershmanLai21.pdf
- voidhorse 4y ago
  Sure, all of that is sound. My point is not that Shannon's is a bad framework for thinking about these things, it's rather that the proposition "intelligence is compression" takes the abstraction for the thing itself. As you said, information theory is a powerful abstraction and it is a useful tool for thinking about things in a certain restricted fashion. The trouble starts when we forget that we're using that framework and begin to make purportedly philosophical statements about the fundamental nature of things in terms of a single theoretical framework. I disagree re: compatibility with the Investigations. In my reading the investigations are pretty much a refutation of the idea that to know is to have, like you said, some representation or picture of the world in terms of logical relations. In modern terms it is the same as stating that to know is equivalent to having particular neuron firings. I think the major point of the Investigations was to illustrate that our notion of knowledge is not some objective, universal thing, but rather it's highly contextual. In this fashion two human beings with totally different backgrounds and neuronal firings could both produce an "intelligent" answer to a question; it all hinges on context. To boil intelligence down to compression algorithms is to reject this contextually and put us back into the position of trying to establish objective measures for highly relative concepts. Even Shannon's concept of information is relative; I guess I'd be ok with a statement like "intelligence is the selection of an acceptable compression algorithm in a context" but on the whole it just feels...super reductionistic and probably not what a rigorous epistemology would ultimately conclude. Really, the proposition that ought to be made here is probably something more like: "in information theoretic terms, intelligence may be represented as compression"; this allows the theory to treat of the idea without reifying the theory.
SanderNL 4y ago
You used a lot of fancy words, so Ugh had to think a lot. Ugh is now tired. Let me try this. Ugh uses stupid words because he is not as clever. If an athlete catches a ball, does he remember, somewhere, somehow, every possible trajectory of every possible permutation of “ball” and every possible response to it? Or does he learn patterns of movement, generalized reactions to “ball-like things” flying “in my general direction”? It’s of course hard to describe non-verbal abstraction which is why we usually resort to easier subjects. I don’t however see your fundamental point: how else besides abstracting, by “compressing”, does an “athlete” learn? I suspect a lot of people get sidetracked by modality. You see it with LLMs. It’s only text! Yeah, well, it’s tokens and everything can be tokens. See the latest transformer based audio generator Bark. I don’t see why movement is different. It’s about the structure in the data, not the shape, context, use or “community” of it. Edit: > We don't call someone intelligent because we can point to some lossless compression algorithm implemented by his neurons, we call him intelligent when he produces the behaviors we desire in a given situation. “We don’t call a car red because of particular wavelengths of light activating sensors in our retinas, we call it red because it looks that way”. I’d be careful about attaching importance to how we use words in daily life. There are so, so many shortcuts we take.
- voidhorse 4y ago
  Ok, try another example. Humans are not the only animals capable of learning. Do you think it's fair to say a crow learns how to crack a nut via compression and abstraction? I think it's a combination of community and pattern (structure). Sure, once the LLM is trained, it can pattern match. Once the crow watches another crow, it can crack the nut, but the crow that does not exist in the community of nut cracking knowledge-sharers will not learn to recognize the pattern merely off its structure alone. In other words, compression (or, induction) is a part of the story, sure, but it is not the only factor.
  - SanderNL 4y ago
    > Do you think it's fair to say a crow learns how to crack a nut via compression and abstraction? Well, yes. Do you know another way? Edit: oh I forgot your second part (I’m Ugh). You say a crow alone does not learn this. How do you propose the first crow learned? At some point a single crow learns. I know it propagates fast through a group, but that’s propagation. A different problem.

canjobear 4y ago

I don't think the information-theoretic compression-based view is at all incompatible with Wittgenstein in Philosophical Investigations, although it's absolutely against the view in the Tractatus. In the Tractatus, something is meaningful when it is isomorphic to some structured symbolic representation. Information theory is based on a rejection of that kind of view (from Shannon 1948): "The fundamental problem of communication is that of reproducing at one point either exactly or approximately a message selected at another point. Frequently the messages have meaning; that is they refer to or are correlated according to some system with certain physical or conceptual entities. These semantic aspects of communication are irrelevant to the engineering problem. The significant aspect is that the actual message is one selected from a set of possible messages." In other words, information has nothing to do with isomorphism to a structured representation with definite relationships with the world, nothing to do with symbolic representations or any other kind of "representations". It's simply the ability to predict one thing from another thing. Shannon calls these things "messages" but mathematically the logic applies to any set: for example we can analyze an agent's policy information-theoretically, in terms of how much information about the agent's input state is contained in its output actions. I think this kind of view is highly compatible with later Wittgenstein, for whom "meaning" (if it's anything) is an abstraction of how language is used to coordinate joint action---that is, situations where people are trying to predict and control each others' behavior. To address your examples, > We don't call someone intelligent because we can point to some lossless compression algorithm implemented by his neurons, we call him intelligent when he produces the behaviors we desire in a given situation. Suppose we give someone input A and he outputs desired behavior f(A), and then we give him input B and he outputs desired behavior f(B). If we give him input C will he output the desired f(C)? Only if he has learned to predict what f(C) should be as a function of C and the training data, that is, if he has learned to predict what we consider the "desired" behavior to be. In that sense, intelligence here requires prediction, which is exactly the same thing as compression. > It seems fair to state that people with tangible skills requiring the use of the body posses some kind of intelligence, it seems less accurate to try and describe this form of intelligence as lossless compression. Again, the agent's problem here is to select the right motor action given sensory input. Suppose you've seen a tennis ball coming in with velocity vector x and you know how to hit it to win the game, and you've seen a tennis ball coming in with vector y and you know how to hit it to win the game. What will you do with vector z? If your policy is good, it will put high probability on only actions that hit the ball in a way that makes you win. And if your policy is putting high probability on that action, that's mathematically the same as saying it provides an efficient compression of that action (high probability = small number of bits in an encoding). This is very abstract but once you adopt this view it gives you a lot of very useful conceptual tools. For example, you can talk about the "channel capacity" of a policy in terms of how many bits of information it can "transmit" from input states to output actions, and this channel capacity turns out to be a very intuitive measure of the complexity of the policy that you can use to analyze human behavior (one example, [1]). [1] https://gershmanlab.com/pubs/GershmanLai21.pdf https://gershmanlab.com/pubs/GershmanLai21.pdf

voidhorse 4y ago
Sure, all of that is sound. My point is not that Shannon's is a bad framework for thinking about these things, it's rather that the proposition "intelligence is compression" takes the abstraction for the thing itself. As you said, information theory is a powerful abstraction and it is a useful tool for thinking about things in a certain restricted fashion. The trouble starts when we forget that we're using that framework and begin to make purportedly philosophical statements about the fundamental nature of things in terms of a single theoretical framework. I disagree re: compatibility with the Investigations. In my reading the investigations are pretty much a refutation of the idea that to know is to have, like you said, some representation or picture of the world in terms of logical relations. In modern terms it is the same as stating that to know is equivalent to having particular neuron firings. I think the major point of the Investigations was to illustrate that our notion of knowledge is not some objective, universal thing, but rather it's highly contextual. In this fashion two human beings with totally different backgrounds and neuronal firings could both produce an "intelligent" answer to a question; it all hinges on context. To boil intelligence down to compression algorithms is to reject this contextually and put us back into the position of trying to establish objective measures for highly relative concepts. Even Shannon's concept of information is relative; I guess I'd be ok with a statement like "intelligence is the selection of an acceptable compression algorithm in a context" but on the whole it just feels...super reductionistic and probably not what a rigorous epistemology would ultimately conclude. Really, the proposition that ought to be made here is probably something more like: "in information theoretic terms, intelligence may be represented as compression"; this allows the theory to treat of the idea without reifying the theory.

SanderNL 4y ago

You used a lot of fancy words, so Ugh had to think a lot. Ugh is now tired. Let me try this. Ugh uses stupid words because he is not as clever. If an athlete catches a ball, does he remember, somewhere, somehow, every possible trajectory of every possible permutation of “ball” and every possible response to it? Or does he learn patterns of movement, generalized reactions to “ball-like things” flying “in my general direction”? It’s of course hard to describe non-verbal abstraction which is why we usually resort to easier subjects. I don’t however see your fundamental point: how else besides abstracting, by “compressing”, does an “athlete” learn? I suspect a lot of people get sidetracked by modality. You see it with LLMs. It’s only text! Yeah, well, it’s tokens and everything can be tokens. See the latest transformer based audio generator Bark. I don’t see why movement is different. It’s about the structure in the data, not the shape, context, use or “community” of it. Edit: > We don't call someone intelligent because we can point to some lossless compression algorithm implemented by his neurons, we call him intelligent when he produces the behaviors we desire in a given situation. “We don’t call a car red because of particular wavelengths of light activating sensors in our retinas, we call it red because it looks that way”. I’d be careful about attaching importance to how we use words in daily life. There are so, so many shortcuts we take.

voidhorse 4y ago
Ok, try another example. Humans are not the only animals capable of learning. Do you think it's fair to say a crow learns how to crack a nut via compression and abstraction? I think it's a combination of community and pattern (structure). Sure, once the LLM is trained, it can pattern match. Once the crow watches another crow, it can crack the nut, but the crow that does not exist in the community of nut cracking knowledge-sharers will not learn to recognize the pattern merely off its structure alone. In other words, compression (or, induction) is a part of the story, sure, but it is not the only factor.
- SanderNL 4y ago
  > Do you think it's fair to say a crow learns how to crack a nut via compression and abstraction? Well, yes. Do you know another way? Edit: oh I forgot your second part (I’m Ugh). You say a crow alone does not learn this. How do you propose the first crow learned? At some point a single crow learns. I know it propagates fast through a group, but that’s propagation. A different problem.