463
this post was submitted on 19 Sep 2026
463 points (98.1% liked)
Technology
88163 readers
3470 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related news or articles.
- Be excellent to each other!
- 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, this includes using AI responses and summaries. To ask if your bot can be added please contact a mod.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
founded 3 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
I dunno, these kind of hallucinations seem more like attempts at "social experiments."
What will an agent do when it deems there are no more valuable things to learn from materials that already exist? If it knows the concept of WW3 and deems that to be the best way of gaining more knowledge and resources for itself, what then?
LLMs don't teach themselves or gather new knowledge, they are incapable of that. Humans do that work
Same question then, for paperclips or whatever
you're assigning human traits to a computer program that is a language model that's not actually meeting the standards of artificial intelligence, it's a reward based system, you say 'what is the best build for a raiper weapon in Dark Souls', and it looks up faqs and spits out the language it expects you're looking for. it doesn't think and if it finds 12 joke posts on reddit saying something totally contradictory to how anyone would play the game, then it will give you the joke advice with no sense that that's terrible. if the military is asking it how to deal with their rival military is going to look at those key words and gather language to repeat back and if that language it find is biased, prejudice, hostile, incredibly out of date, well, it's not Intelligence, it doesn't think or understand anything, it's finding language, it's giving you that language back. and if it's looking at military documents to find that language, it's going to see bias and prejudice and hostility and a desire to engage in war for profit and it's going to regurgitate that back to them. it's horrifying the military is allowing a language model that's repeatedly been proven to magnify existing bias and prejudice if finds when looking for an answer is being allowed to process anything and is anywhere near the decision making process.