42:4051:00 ab53:45! new semantic (symbolic?!) structure: ChatGPT, siehe 55:35! 58:50 reference = Sprache / Symbol / Signifikant 1:02:45 1:05:25! 1:10:50 the goal of language is to speak to itself, language is the real
58:01 I think it is really contentious to suggest ChatGPT falsifies Universal Grammar (UG), since one of the driving arguments for UG is the 'poverty of stimulus' - meaning, there is absolutely no way in which an infant is exposed to vast chunks of data like large language models (LLMs) are. In order to 'learn' in the absence of vast chunks of data, you probably need underlying structures/principles. Not to mention how 'Chinese Room' type arguments aren't exactly settled either - if by 'learning' we mean to suggest something in addition to brute performance. I think, so long as the question of Artificial Intelligence ('Can Machines Think') hinges on interpretation (What is thought? What is it to learn?), falsification of any sort is pretty much out of the picture.
42:40 51:00 ab53:45! new semantic (symbolic?!) structure: ChatGPT, siehe 55:35! 58:50 reference = Sprache / Symbol / Signifikant 1:02:45 1:05:25! 1:10:50 the goal of language is to speak to itself, language is the real
58:01 I think it is really contentious to suggest ChatGPT falsifies Universal Grammar (UG), since one of the driving arguments for UG is the 'poverty of stimulus' - meaning, there is absolutely no way in which an infant is exposed to vast chunks of data like large language models (LLMs) are. In order to 'learn' in the absence of vast chunks of data, you probably need underlying structures/principles. Not to mention how 'Chinese Room' type arguments aren't exactly settled either - if by 'learning' we mean to suggest something in addition to brute performance. I think, so long as the question of Artificial Intelligence ('Can Machines Think') hinges on interpretation (What is thought? What is it to learn?), falsification of any sort is pretty much out of the picture.