3 ms·

Why does she do that? I would guess for performance. So you would be side-stepping some of the compilation?

pjc50 7y ago
I suspect in the end it's not much faster, but it is smaller. Really the fun in the essay is that that's not what any normal person would do as a solution to that problem, it's a stunt to show off deep mastery.
- mikorym 7y ago
  But in this case, is it mastery? To me, mastery would be if you can use mathematics to trivialise an implimentation in a counter intuitive way. I understand that this is a different kind of mastery. But does the essay write code that is a sign of mastery, or does it just use the byte code as a talisman? I don't follow the code so I wouldn't know.
  - gknoy 7y ago
    I think it counts as mastery, precisely because it does choose a thing so counterintuitive that it's past what most of us would even think of as a possible solution. I spent many years writing Java, but never took the time to learn how to write the bytecode (that's why we have a compiler ...). I know that Clojure lets you call Java classes, but it's a completely crazy, magical, frightening, and awesome thing to see someone do it by writing their own class loader, and then writing their own class -- in bytecode! -- to load, when the _simplest_ and clearest solution would be far from that. It's a bit like solving FizzBuzz with Tensorflow: terrifying on at least some level, as it makes it so clear that there's so much more depth I could be learning, yet exciting for almost the same reasons.
    - mikorym 7y ago
      > It's a bit like solving FizzBuzz with Tensorflow I had to wiki that, but it's a good point. The first thing about NN that interested me, rather than images or audio or anything like that, is that a one level NN (IIRC) can approximate any function (by adding more and more nodes to that level). I guess the point is also that any really deep rabbit hole is in some sense better than a selection of random 1000 rabbit holes.

I suspect in the end it's not much faster, but it is smaller. Really the fun in the essay is that that's not what any normal person would do as a solution to that problem, it's a stunt to show off deep mastery.

mikorym 7y ago
But in this case, is it mastery? To me, mastery would be if you can use mathematics to trivialise an implimentation in a counter intuitive way. I understand that this is a different kind of mastery. But does the essay write code that is a sign of mastery, or does it just use the byte code as a talisman? I don't follow the code so I wouldn't know.
- gknoy 7y ago
  I think it counts as mastery, precisely because it does choose a thing so counterintuitive that it's past what most of us would even think of as a possible solution. I spent many years writing Java, but never took the time to learn how to write the bytecode (that's why we have a compiler ...). I know that Clojure lets you call Java classes, but it's a completely crazy, magical, frightening, and awesome thing to see someone do it by writing their own class loader, and then writing their own class -- in bytecode! -- to load, when the _simplest_ and clearest solution would be far from that. It's a bit like solving FizzBuzz with Tensorflow: terrifying on at least some level, as it makes it so clear that there's so much more depth I could be learning, yet exciting for almost the same reasons.
  - mikorym 7y ago
    > It's a bit like solving FizzBuzz with Tensorflow I had to wiki that, but it's a good point. The first thing about NN that interested me, rather than images or audio or anything like that, is that a one level NN (IIRC) can approximate any function (by adding more and more nodes to that level). I guess the point is also that any really deep rabbit hole is in some sense better than a selection of random 1000 rabbit holes.