3 ms·
I thought the same. Perhaps we can think of a way to do it? The only thing coming to mind is either using a markov chain or a char-RNN to generate code files. T
by iraphael 11y ago
I thought the same. Perhaps we can think of a way to do it? The only thing coming to mind is either using a markov chain or a char-RNN to generate code files. The latter has been done here: http://karpathy.github.io/2015/05/21/rnn-effectiveness/ http://karpathy.github.io/2015/05/21/rnn-effectiveness/
- mbwolff 11y agoIf you find a way to do this, please let me know. For bonus points you should follow Roubaud's principle whereby the code "speaks" about its own algorithms.
- iraphael 11y agoOhh, in that case, we could train the LSTM only with code that was used to generate it. Unfortunately, I think that's just not enough data to get interesting results. Maybe we could train it on multiple different kinds of implementations of LSTMs, but that may or may not be cheating...
- JadeNB 11y agoYou're thinking of it much more high-poweredly than I (praise, not criticism!). I meant something simple, like (by analogy with La Disparition) trying to write code without using the letter 'e'. In that vein, perhaps one can view code golf as a sort of programming analogue of Oulipo—or maybe even functional programming (try to write code without using mutation!).
- iraphael 11y agoOh yes, those experiments would have been fun. But let me explain what I had in mind. Potential Literature is about finding meaning in things. Meaning is human-made anyways, so anyone can find meaning in the most random of things. Having a randomly-generated, but still structured, text can allow for new meaning to come through. Think SubredditSimulator - generating random stuff, and letting the users find/create meaning for it.