from this article, which is well worth the read, if only for the fun of seeing zuck get dunked on
I've been saying this for years. I got on LLM (large learning model) bandwagon years ago for research purposes. There is so much scientific data out there that if a program could help us get through it faster, we could really get somewhere. Anything that needs 4 or more variables assessed in tandem is so much faster with an LLM. Plus doing it as a human is super tedious and prone to errors. No one wants to do it or grad students are forced to do it alongside other responsibilities. So it would be great if an LLM could help.
LLMs learn one thing (and one thing only!) and can repeat the learned data in any way you require. Such as summaries and graphs. They can predict in the sense that after looking at 1000 items in a sequence they can guess what comes next. But the only accurate thing about them is them repeating data back at you in a condensed format.
Then ChatGPT comes out and I'm of course highly dubious of the term "AI". Lo and behold, any further reading on it reveals its just an LLM, same as any other supposed "AI" out there.
Again, LLMs do not produce something new. They repeat back what was given in the way you want. So "AI" scrapes the internet and just spews the main Google page at you but with a twist. The twist is the bad part because every extra instruction you give an LLM on how to repeat back to you, every bit more complicated it gets, the more likely it is to make mistakes. Programmers working with LLMs often have the program spit out the results of every individual step when this happens so they can fix it and move on. "AI" does not do this and people don't think to ask. People don't know to ask. They've been sold the idea that this machine thinks for itself.
It doesn't. It just echoes information from others in any way you like.
The more complicated the order (such as condensing a summary into a few paragraphs, or using a sassy tone, or explaining via pop culture analogies, etc), the more likely for things to go wrong. Because then it's trying to combine multiple learning models in one. An LLM trained on history can probably give good answers. An LLM trained on vocal tone might approach accurate sounds. An LLM trained on analogies will always make mistakes (because it would need a lot of data on each item its trying to compare). But a single LLM trying to do all three at once is going to encounter redundancies and contradictions, each request guiding it to a different outcome.
LLMs, "AI", are an echo. Echoes are just our own words bounced back. Echoes can be fun to play with. Echoes can be somewhat helpful in navigating the world when we analyze the way they bounce back.
Echoes are not another intelligent being. Echoes are not thinking. Echoes are not providing anything else that hasn't already been said before.
Please use LLMs to help make individualized cancer treatments and optimize power grids and accurately predict the weather. Please do not use "AI" for art or writing or war. Thats just the companies trying to get more use and money than the smaller user base they would naturally have.
