Hacker News Reader

> Copilot has the ability to produce Quake's source code nearly by itself.

Was it fed the Quake source code while training? Then it's not producing that code, it's just reproducing it, like a fancy (but imperfect) copy machine.

I'm not sure it's accurate to say that the training source code is "compressed" in the parameters of the model, but certainly some approximation of the training source code is stored in the parameters.

View on HN

It is probably a stretch, but I think less of a stretch than saying "it just a machine that learned to code and randomly reproduced these 10+ lines of code". That has IMHO a probability of 0.

So if I rule that out, where does it end up? What if we put this as the grown up ML brother of the chain of LZW, PPM, dictionary assisted compression (e.g. zstd) and various attempts at using neural networks for compression?

I would not want to judge this - that's why I put up the AGPL idea. Or even unlicensed code. It would be a very interesting case to watch.

Reply on news.ycombinator.com