4 ms·

thanks for the info! great stuff! gpt3's generalization-by-description never ceases to amuse me; but the difficult thing here is to get the right abstraction l

thanks for the info! great stuff! gpt3's generalization-by-description never ceases to amuse me; but the difficult thing here is to get the right abstraction layers layered nicely in the conceptual lasagna. This is where category theory becomes extremely powerful. It has occured to me that codex-davinci has an intuitive "understanding" of constructs like monads, or something along that line.

tectonic 5y ago
debuild.co looks cool. Using Codex yet?
- archibaldJ 5y ago
  yes; here is our latest demo https://twitter.com/sharifshameem/status/1425185575645024256 https://twitter.com/sharifshameem/status/1425185575645024256
Y_Y 5y ago
Can you expand on the utility of categories here? There's a lot of space between knowing what defines a monad, when something might be a monad, what you can do with monadic structure etc. Of course if an AI truly understood monads I it would be a bright line marking where the machines have finally surpassed the human mind. Cool.
- archibaldJ 5y ago
  I think it's closely linked to the notions of semantics as approached in CS (ie. PLT) v.s. in linguistics where we are mostly concerned with the "micro-structures" and "meso-structures" that gave rise to qualia we humans experience (e.g. therefore languages with different structural systems such as English vs Chinese encode concepts (as well as intentions) very differently; (for an illustration, see Interality as a Key to Deciphering Guiguzi: A Challenge to Critics [1])), and not so much about how evaluation and execution came about (e.g. as studied from a compiler's persective in denotational semantics, or a more functional perspective in operational semantics, where things like natural transformations are ubiquitous) [1]: https://cjc-online.ca/index.php/journal/article/view/3187/3258 https://cjc-online.ca/index.php/journal/article/view/3187/32... And so categories naturally come in as a way to bridge and compose these two worlds, and that's just the beginning. There're so much more we don't understand yet, such as what understanding really is given a certain set of contexts and constraints as well as in regards to their relaxations. What is understanding if there are no doings? And what is doing with no understandings? How do things compose? These are great mysteries.
  - Y_Y 5y ago
    > How do things compose? I think you pretty much summed it up right there.
antonvs 5y ago
That's a bit like claiming that dogs have an intuitive understanding of calculus because they can catch a ball. There's a rather weak sense in which that claim might be argued to have some validity, but it doesn't have the implications you seem to think. Monads are not a feature of the universe, independent of human minds. They're a concept we impose on things in order to make dealing with them more tractable in some way. It's very unlikely that a machine learning algorithm is relying on anything remotely like "an intuitive understanding of constructs like monads".

tectonic 5y ago

debuild.co looks cool. Using Codex yet?

archibaldJ 5y ago
yes; here is our latest demo https://twitter.com/sharifshameem/status/1425185575645024256 https://twitter.com/sharifshameem/status/1425185575645024256

Y_Y 5y ago

Can you expand on the utility of categories here? There's a lot of space between knowing what defines a monad, when something might be a monad, what you can do with monadic structure etc. Of course if an AI truly understood monads I it would be a bright line marking where the machines have finally surpassed the human mind. Cool.

archibaldJ 5y ago
I think it's closely linked to the notions of semantics as approached in CS (ie. PLT) v.s. in linguistics where we are mostly concerned with the "micro-structures" and "meso-structures" that gave rise to qualia we humans experience (e.g. therefore languages with different structural systems such as English vs Chinese encode concepts (as well as intentions) very differently; (for an illustration, see Interality as a Key to Deciphering Guiguzi: A Challenge to Critics [1])), and not so much about how evaluation and execution came about (e.g. as studied from a compiler's persective in denotational semantics, or a more functional perspective in operational semantics, where things like natural transformations are ubiquitous) [1]: https://cjc-online.ca/index.php/journal/article/view/3187/3258 https://cjc-online.ca/index.php/journal/article/view/3187/32... And so categories naturally come in as a way to bridge and compose these two worlds, and that's just the beginning. There're so much more we don't understand yet, such as what understanding really is given a certain set of contexts and constraints as well as in regards to their relaxations. What is understanding if there are no doings? And what is doing with no understandings? How do things compose? These are great mysteries.
- Y_Y 5y ago
  > How do things compose? I think you pretty much summed it up right there.

antonvs 5y ago

That's a bit like claiming that dogs have an intuitive understanding of calculus because they can catch a ball. There's a rather weak sense in which that claim might be argued to have some validity, but it doesn't have the implications you seem to think. Monads are not a feature of the universe, independent of human minds. They're a concept we impose on things in order to make dealing with them more tractable in some way. It's very unlikely that a machine learning algorithm is relying on anything remotely like "an intuitive understanding of constructs like monads".