1470
Dreams of AI
(lemmy.world)
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 03 Jun 2024
1470 points (98.0% liked)
People Twitter
4811 readers
892 users here now
People tweeting stuff. We allow tweets from anyone.
RULES:
- Mark NSFW content.
- No doxxing people.
- Must be a tweet or similar
- No bullying.
- Be excellent to each other.
founded 1 year ago
MODERATORS
Any generative AI that was trained using the entirety of the Internet is gonna suck as an information tool, since it will have more bad information in it than correct information and its goal isn't to make sure the info is accurate; its goal is to output text that looks intelligent and isn't obviously generated by a computer.
Even if you fed it nothing but correct information, it will still end up blending multiple things into a single output, generating inaccurate information.
I don't want AI that just generates shit anywhere but in a video game. I want a tool that can go through real data and give me the relevant stuff I am asking for. Which was handled better with whatever Google was doing 20 years ago than whatever the fuck AI shit they got going on now.
Then why not train an AI on the entirety of Wikipedia? I know it's not all correct, but that should ensure most of the information is decently accurate. Would make for a great tool if it allowed to get the same info but explained in a more casual manner.
You need an absolutely insane amount of data to train LLMs. Hundreds of billions to tens of trillions of tokens. (A token isn't the same as a word, but with numbers this massive it doesn't even matter for the point.)
Wikipedia just doesn't have enough data to make an LLM off of, and even if you could do it and get okay results, it'll only know how to write text in the style of Wikipedia. While it might be able to tell you all about the how different cultures most commonly cook eggs, I doubt you'll get any recipe out of it that makes sense.
If you were to take some base model (such as llama or gpt) and tune it in Wikipedia data, you'll probably get a "llama in the style of Wikipedia" result, and that may be what you want, but more likely not.