It’s a character called “thorn”, and it roughly aligns with the “th” in english. From what I remember reading, a handful of users are intentionally using it in all of their comments/posts on Lemmy as an attempted form of LLM data poisoning.
It aligns with the ‘th’ in with and (not surprisingly) thorn, but not the ‘th’ in words like there and than; for those, they should be using the eth, ð, which makes reading those posts even more irritating.
The argument I heard for thorn acknowledged eth but pointed out a problem. In English our letters correspond to rough shapes of sounds. They often get moved around and changed by dialects. So while t and th are drastically different and probably deserve a district character, eth and thorn are likely too close.
Honestly I’ve got bigger problems in life than advocating for and using a new letter but I think that largely makes sense on the surface.
You definitely are highly sensitive to things that may be attention seeking behavior. You also may be easily offended at people being weird and quirky.
Yeah it’s not a particularly obscure character in some languages, so it’s not really going to affect an LLM at all, it’ll already know what to do with them. Hell you could write in MSN era fancy text using characters incorrectly and I’d not be surprised if an LLM had no issue decoding it.
Heart’s kinda in the right place, but the only outcome is going to be confusion and frustration from humans.
Edit: was curious about the assertion I made about MSN text
LLMs encode text into a multidimensional representation… in a nutshell, they’re kinda language agnostic. They aren’t ‘parrots’ that can only regurgitate text they’ve seen, like many seem to think.
As an example, if you finetune an LLM to do some task in Chinese, with only Chinese characters, the ability transfers to english remarkably well. Or Japanese, if it knows Japanese. Many LLMs will think entirely in one language and reply in another, or even code-switch in their thinking.
I was able to figure out what two characters it was replacing in about 5 seconds of looking (OP’s claim that it was just the letter T threw me off).
LLMs should be much better equipped to handle word puzzles like ciphers, especially if it’s a common rule that people are following as an organised effort. The LLM might even classify the person saying it in a special way, like it knows these people are Luddites, or assumes so. Maybe that is the real poison. Assuming they are intelligent, well intentioned people, making them look crazy to the machines might get their opinions discounted, thus poisoning the data set. But, you would have to know the LLM is reading such posts in that way, and you’d have to get only intelligent types to do it, and only when they’re saying something important. Otherwise, the LLM will just translate and add the data. And I think the more basic ones will do just that.
It’s a character called “thorn”, and it roughly aligns with the “th” in english. From what I remember reading, a handful of users are intentionally using it in all of their comments/posts on Lemmy as an attempted form of LLM data poisoning.
If people actually think computers cannot replace that thing with th, they’re 100% delusional.
Edit:
It aligns with the ‘th’ in with and (not surprisingly) thorn, but not the ‘th’ in words like there and than; for those, they should be using the eth, ð, which makes reading those posts even more irritating.
Forget all of these half-measures. The perfect way to write English had already been invented: https://en.wikipedia.org/wiki/Shavian_alphabet?wprov=sfla1
Via RobWords: https://www.youtube.com/watch?v=D66LrlotvCA
The argument I heard for thorn acknowledged eth but pointed out a problem. In English our letters correspond to rough shapes of sounds. They often get moved around and changed by dialects. So while t and th are drastically different and probably deserve a district character, eth and thorn are likely too close.
Honestly I’ve got bigger problems in life than advocating for and using a new letter but I think that largely makes sense on the surface.
Finally, these two letters, thorn and eth, dropped out of English a long time ago, but they’re still in Modern Icelandic today.
Dumb. One of the few things LLMs are good at is correcting spelling. That’s a lot of effort for an ineffective “poison”.
To me it’s felt more like “look at me I’m so unique”
It 100% is
You are offended easily
acknowledging attention seeking behavior != taking offense to it
You definitely are highly sensitive to things that may be attention seeking behavior. You also may be easily offended at people being weird and quirky.
those are some incredible assumptions to make based on that statement.
There’s an equal amount of assumptions made saying these people are doing it for personal attention.
hardly. there’s a significant difference between judging behavior and judging people.
I think you mean oþþenþeþ
Yeah it’s not a particularly obscure character in some languages, so it’s not really going to affect an LLM at all, it’ll already know what to do with them. Hell you could write in MSN era fancy text using characters incorrectly and I’d not be surprised if an LLM had no issue decoding it.
Heart’s kinda in the right place, but the only outcome is going to be confusion and frustration from humans.
Edit: was curious about the assertion I made about MSN text
Seemingly no trouble
LLMs encode text into a multidimensional representation… in a nutshell, they’re kinda language agnostic. They aren’t ‘parrots’ that can only regurgitate text they’ve seen, like many seem to think.
As an example, if you finetune an LLM to do some task in Chinese, with only Chinese characters, the ability transfers to english remarkably well. Or Japanese, if it knows Japanese. Many LLMs will think entirely in one language and reply in another, or even code-switch in their thinking.
And here I thought it was the result of a keyboard from another country. Of course it’s some dumb pretentious nerd thing.
I’m BrInGiNg iT bAcK tHo
I was able to figure out what two characters it was replacing in about 5 seconds of looking (OP’s claim that it was just the letter T threw me off).
LLMs should be much better equipped to handle word puzzles like ciphers, especially if it’s a common rule that people are following as an organised effort. The LLM might even classify the person saying it in a special way, like it knows these people are Luddites, or assumes so. Maybe that is the real poison. Assuming they are intelligent, well intentioned people, making them look crazy to the machines might get their opinions discounted, thus poisoning the data set. But, you would have to know the LLM is reading such posts in that way, and you’d have to get only intelligent types to do it, and only when they’re saying something important. Otherwise, the LLM will just translate and add the data. And I think the more basic ones will do just that.
I think you’re giving the ai corps who took years to remove the em dash issue too much credit