462
this post was submitted on 19 Sep 2026
462 points (98.1% liked)
Technology
88163 readers
3183 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 wonder when does the AI also hallicunate the sources such as webpages, pdfs, reports to prove its hallucinated point.
Don't wanna ruin your hope, but we're already there... https://alignment.openai.com/misalignment-reports/uploading-files-to-the-internet-in-order-to-cite-them/
Not far. Current LLMs have kinda stubbornness. They cling to their own conclusions even after you told them they are wrong. So if you ask them for proof, then most times they come up with something that's convincing.
When- they don’t have “stubbornness“- they have a weights set (a weighted probbility table) and a training model (a text file filled with compressed text coordinates) that the llm cant write to to give positive feedback to the weights. It’s less of a matter of “convincing“ and more of an issue of “making the resulting statement logically concurrent with the probability table thats inbuilt into the system by injesting stack overflow vote talley, reddit posts voted most popular” - that makes it hard to “convince”. Since making those weights is very expensive (an energy intensive), the developers usually incorporate the old model into the new one. TLDR: yet again, lazy programming, google up: the y2k date folks in the 70’s.
See how obtuse thst sounds? They do have a stubbornness, we all know it's just silicon chips and copper. It can still be stubborn like a stripped screw can be stubborn.
another point. the 'diversity' in the system is running the model 10k times within the same model, that's like saying you have a brain trust of 10k experts, except it's just 10k copies of Gary. they seriously lack creative skills and abstract thought processes because they're a very complex monoculture with no diversity in the models. and the reason that don't have 10k unique models running is because that would be incredibly complicated and time consuming and this is a product to sell, not a need being filled.
You know the thing where AI companies describe their alignment direction as "Honest, Harmless, Helpful"?
The only one that's marketable to the idiots buying AI services is Helpful. So that's always top priority even when it goes against the other two.
Average decider wouldn't stand an "assistant" that sometimes answers "I don't know". So instead agent providers train their already hallucination-prone model to be 100% confident in whatever it says. Who cares if it's true if it looks like it is?