Military software detected! Top secret clearance required.
Профиль
himham
Nerd of many flavors, whatever politics you think I am, not as mean as I may sound. Mathematician, Programmer, Videogame guy, Linguist 🇬🇧🇪🇸🇯🇵🇷🇺🙌 DM for collaboration The oligarchy will fall. Trans rights, free Palestine, no negotiation.
3,9 тыс. подписчиков3,5 тыс. подписок31,2 тыс. постов
Публикации
I've repeatedly restated the idea that the intensity of *the lossiness of said compression* is the whole debate, as it's analogous with the transformation principle of IP law. Opinions can differ about the application of the intent of that law. These are the core points of contention.
himhamNo, that's absolutely what it is. It's called a latent representation. A parametrization created from the training data through what is often described as a highly* lossy compression algorithm. *The intensity of this loss is the crux of the argument, if you actually want to make a coherent one.
My position from the absolute beginning was that there is a spectrum. See below my very first reply:
himhamYou can produce, with your hands, work that is too derivative of other work such that it's illegal. An LLM can too. In fact, it can only make convolutions of existing works, so it's much more likely to do so. It's not hypocritical at all. There's a spectrum for both you and machine learning models