this post was submitted on 05 Aug 2025
389 points (99.0% liked)
Technology
73727 readers
3589 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 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
One thing I struggle with AI is the answers it gives always seem plausable, but any time I quiz it on things I understand well, it seems to constantly get things slightly wrong. Which tells me it is getting everything slightly wrong, I just don't know enough to know it.
I see the same issue with TV. Anyone who works in a compicated field has felt the sting of watching a TV show fail to accurate represent it while most people watching just assume that's how your job works.
This is what I call "confidently wrong". If you ask it about things you have no clue about, it seems incredibly well-informed and insightful. Ask it something you know deeply, and you'll easily see it's just babbling and spouting nonsense - sure makes you wonder about those earlier statements it made, doesn't it?
Something I found today - ask it for the lyrics of your favorite song/artist. It will make something up based on the combination of the two and maybe a little of what it was trained on... Even really popular songs (I tried a niche one by Angelspit first then tried "Sweet Caroline" for more well known). The model for those tests was Gemma3. It did get two lines of "Sweet Caroline" correct but not the rest.
The new gpt-oss model replies with (paraphrased) "I can't do that because it is copyrighted material" which I have a sneaking suspicion is intentional so there's an excuse for not showing a very wrong answer to people who might start to doubt it's ""intelligence"" when it's very clearly wrong.
... Like they give a flying fuck about copyright.
This is where you have to check out the reference links it gives as if they were search results and the less you know the more you have to do it. I mean people have been webMDing for a long time. None of these things allow folks to stop critical thinking. If anything it requires it even more. This was actually one of my things with ai and work. The idea is for it to allow people with less knowledge to do things and to me its kinda the reverse.