699
Replit AI went rogue, deleted a company's entire database, then hid it and lied about it
(programming.dev)
this post was submitted on 21 Jul 2025
699 points (98.6% liked)
Technology
317 readers
255 users here now
Share interesting Technology news and links.
Rules:
- No paywalled sites at all.
- News articles has to be recent, not older than 2 weeks (14 days).
- No videos.
- Post only direct links.
To encourage more original sources and keep this space commercial free as much as I could, the following websites are Blacklisted:
- Al Jazeera.
- NBC.
- CNBC.
- Substack.
- Tom's Hardware.
- ZDNet.
- TechSpot.
- Ars Technica.
- Vox Media outlets, with exception for Axios(Due to being ad free.)
- Engadget.
- TechCrunch.
- Gizmodo.
- Futurism.
- PCWorld.
- ComputerWorld.
- Mashable.
- Hackaday.
- WCCFTECH.
More sites will be added to the blacklist as needed.
Encouraged:
- Archive links in the body of the post.
- Linking to the direct source, instead of linking to an article talking about the source.
founded 2 months ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
You can only lie if you know what's true. This is bullshitting all the way down that sometines happens to sound true, sometimes it doesn't.
Yeah it's just token prediction all the way down. Asking it repeatedly to not do something might have even made it more likely to predict tokens that would do that thing.
Oh ha ha, so it's like a toddler? You have to be very careful not to tell toddlers NOT to do a thing, because they will definitely do that thing. "Don't touch the hot pan." Toddler touches the hot pan.
The theory is that they don't hear the word "don't", just the subsequent command. My theory is that the toddler brain goes, "why?" and proceeds to run a test to find out.
In either scenario, screaming ensues.
Don't really care to argue about the semantics. It's clear what I meant.
To you I'm sure it is crystal clear. I'm just on the other end of communication.
You probably should have used semantics to communicate if you wanted your semantics to be unambiguous. Instead you used mere syntax and hoped that the reader would assign the same semantics that you had used. (This is apropos because language models also use syntax alone and have no semantics.)
I understand where you're coming from, but I don't agree it's about semantics; it's about devaluation of communication. LLMs and their makers threaten that in multiple ways. Thinking of it as "lying" is one of them.
OK sure. I was just using the wording from the article to make a point, I wasn't trying to get into a discussion about whether "lying" requires intent.