Hacker News Reader

Comment by ghurtado | Hacker News Reader

I didn't buy into the "copying unauthorized zeros and ones is theft" narrative (because it's not even an argument) when old music industry executives were screaming about Napster, and I'm certainly not gonna start now.

There's absolutely a moral debate to be had in this topic, but if your starting position is redefining common language and twist the meaning of words beyond reason, then you've already shut it down before it's even started.

View on HN

I don't think it's twisting common language, but let me clarify: Tensors are a form of compression. If I take a song that a musician put out for free, convert it to a different file format and sell it (that's the key here, the commercial use), that's theft. It was theft when it was Napster, too, but even then there's a big difference between freely distributing pirated stuff and selling it.

The LLMs hoovered up everything, compressed it, and put it up for sale as bootleg IP. That's especially offensive since so much of what they took was already freely available with attribution.

> Tensors are a form of compression

This is an incorrect understanding of machine learning data-structrues.

A data structure which encodes a bunch of square vectors as input, stores them in a multi-gigabyte file, and uses weights across layers of them to generate an output is no different from any simple compression algorithm or a bunch of convolution kernels.

Just saying those words doesn't make it so.

If you read 300 lines of code, that anyone could write in 1995 and train successfully with minimally $1b of GPU power in 2022, it would make sense to you too.

It's the world's most expensive all-inclusive compression and theft algorithm, perfect for cannibalizing and re-selling all human knowledge if you own enough firepower to train it and sell tokens.

Sez me? Yeah, it doesn't matter if me saying it makes it true; an objective look at the thing confirms how it works; each model is a self-contained artifact holding compressed stolen goods.

Reply on news.ycombinator.com