this post was submitted on 29 Apr 2024
195 points (94.9% liked)
Technology
59135 readers
3816 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Uh, I understand the sentiment, but the model doesn't know anything. And it's legit really hard to differentiate between factual things and random bullshit it made up.
Was gonna say, the AI doesn't make up or admit bullshit, its just a very advanced a prediction algorithm. It responds with what the combination of words that is most likely the expected answer.
Wether that is accurate or not is part of training it but you'll never get 100% accuracy to any query
If it can name what the most likely combination is, couldn't it also know how likely that combination of words is?
It's not actually deciding anything, the AI thinking is marketing fluff really. But yes that's called confidence rating and it does. But at the scale of something like chatgpt that uses a snapshot of the entire internet and is non mutable there's no way to train it for every possible question. If you ask about a topic 99% of the internet gets wrong it'll respond the wrong thing with 99% confidence
No, because that requires it to understand the words. It doesn't.
If it has been trained using questionable sources, or if it's training data includes sarcastic responses (without understanding that context), it isn't hard to imagine how confidently wrong some of the responses could be.
Yeah, no one can make it say "I don't know" because it is not really AI. Business bros decided to call it that and everyone smiled and nodded. LLMs are 1 small component (maybe) of AI. Maybe 1/80th of a true AI or AGI.
Honestly the most impressive part of LLMs is the tokenizer that breaks down the request, not the predictive text button masher that comes up with the response.
Yes, exactly! It's ability to parse the input is incredible. It's the thing that has that "wow" factor, and it feels downright magical.
Unfortunately, that also makes people intuitively trust its output.
It "knows" as in it has access to the information and the ability to provide the right info for the right context.
Any part of that process the AI can just "bullshit" and fills in the gaps with random stuff.
Which is what you want when it's "learning". You want it to try so it's attempt can be rated, and the relevant info added to its "knowledge".
But when consumers are using it, you want it to say "I can't answer that". But consumers are usually stupid and will buy/use the one that says "I can't answer that" the least.
Which is why AI should tell end users "I don't know" more often.
If you feel this is a simple solution, I strongly suggest you write up exactly how you do this and make yourself a billion dollars.
It doesn't, though, any more than you have access to the information in a pile of 10 million shredded documents.
Right, in this case that we're talking about...
Do you not understand how "answer unavailable" is a better answer than taking a small percent of strips of paper at random and filling in the rest with words that sound relevant?
It's like a mad libs
Right. They're text generators. That's the technology. It can't do what you're demanding because that's not how it works. LLMs aren't magic answer machines. They don't know when to say "answer not available". They don't know what they're being asked. They don't know anything.
That is what LLMs do in EVERY conversation. Most of the time you don't notice it, because it fits your expectations.
You know that answer unavailable is better because you have real intelligence, an LLM is just some mathematical functions so it can't do that. If it could it would be getting much closer to actually being AI.