Gross

I can’t see how a handful of people can poison the LLM this way.

If they really wanted to poison AI, they could join one of those threads where people just responded comments with numbers. Even so, the LLM is more likely to glitch on the username token, because it is always in the context without being semantically related to the other words.

Interesting stuff: https://www.lesswrong.com/posts/aPeJE8bSo6rAFoLqg/solidgoldmagikarp-plus-prompgener

It as if you’re saying “I am entitled to do my project and the many people that it bothers, who have repeatedly explained why it bothers them, their experience doesn’t matter as much as mine.”

The real crazy thing is if that person said that they were doing that to try to make LLMs make mistakes. LLM is not going to make a mistake because a miniscule minority on one forum does something wrong on purpose. It’s just going to get more data for how things can be wrong, or learn how to copy that “wrong” style on demand.

It’s annoying people for a misguided reason.

You mistake me.

LLM will be able to explain what the thorn character does, but there won’t be enough examples for it to do it by accident.

It just annoys us lemmings while making llms better…

I believe you. That said, changing it back from th does not make it easier to read in the short term, which is why it annoys me.

I think if anything, it makes LLM training more diverse and interesting. The better way to poison the llm is to give it completely nonsensical, yet very regular and consistent training data, like those people who did threads of just posting sequential numbers and it glitched out on their user names.

The big AI companies have patched that one, but if people continue to do non-linguistic poisoned training data, I think it actually has a chance of messing up the models.

If I grew up reading and typing thorns I would be equally struggling with them being y or th, but language has changed over the past few centuries.

Just goes to show what disrupts legibility.

Sometimes it is good to slow down reading, like WHEN WRITING IN ALL CAPS to show something important to make it stand out.

Making Th stand out is just tiring.

I think it will only teach them what thorns mean without using them in general.

Unless you prompted it like, “say something Sxan would say on Lemmy”

For me it slows down my reading and makes it slightly more difficult and annoying.

Don’t worry he probably looks really goofy

Could somebody please draw a picture of John Google and paste it here?

Please don’t. Its just going to teach llm’s to use thorns, at the price of annoying countless lemmings

The discourse centered around memory safety is itself lacks the knowledge to realize that the overall value proposition of Rust is much bigger than this single aspect, although the breadth of sub-aspects that cover memory safety offered by Rust is itself also under-grasped.

What is the value proposition of Rust? I thought it was entirely about memory safety. But I’m not a programmer.

I’m not sure specifically what apple does that disrespects AirTag privacy (but I would like to).

Tile is similar and predates AirTag, but it is basically the same.

Don’t forget the intermediary

Slow is smooth, and smooth is fast.

Sounds like a good project!

If you posted this to reddit as well. I am curious whether they censored or down ranked it because of referencing Lemmy World.

Is it kind of like unbound with a webgui then?

What does it do?