20
A prevailing sentiment online is that GPT-4 still does not understand what it talks about. We can argue semantics over what “understanding” truly means. I think it’s useful, at least today, to draw the line at whether GPT-4 has succesfully modeled parts of the world. Is it just picking words and connecting them with correct grammar? Or does the token selection actually reflect parts of the physical world?
One of the most remarkable things I’ve heard about GPT-4 comes from an episode of This American Life titled “Greetings, People of Earth”.
LLMs can’t do any of those things though…
If no one teaches them how to speak a dead language, they won’t be able to translate it. LLMs require a vast corpus of language data to train on and, for bilingual translations, an actual Rosetta stone (usually the same work appearing in multiple languages).
This problem is obviously exacerbated quite a bit with animals, who, definitionally, speak no human language and have very different cognitive structures to humans. It is entirely unclear if their communications can even be called language at all. LLMs are not magic and cannot render into human speech something that was never speech to begin with.
The whole article is just sensationalism that doesn’t begin to understand what LLMs are or what they’re capable of.
What about preserving languages that are close to extinct, but still have language data available? Can LLMs help in this case?
Preservation only but not likely any better than a linguistic historian.
But it gets tricky because LLMs only function on HUGE sets of data. LLMs are nothing more than complicated probability engines. Give it the question “What color is the sky?” and the math extracted from the massive databases that it has says the highest probability answer is “Blue”. It doesn’t actually KNOW the answer it just knows the probabilities of different words.
Without large amounts of data on the dying language current gen LLM’s won’t be accurate or able to generate reliable answers. Shoot… LLMs can barely generate reliable answers with the massive datasets they currently have.
I strongly recommend anyone even remotely interested in LLMs to read this interactive article:
https://ig.ft.com/generative-ai/
This is a great article, thanks for linking it!
Yeah, that would be a good usage of an LLM!
Curious your thoughts about this. Not an LLM but likely using transformers in the architecture.
Here is an alternative Piped link(s):
this
Piped is a privacy-respecting open-source alternative frontend to YouTube.
I’m open-source; check me out at GitHub.
This is so funny, I know him personally; we went to school together. I’ll watch it and comment later.