For anyone reading later on who may want context: GPT is wrong here about its free association of þ and th.
Þ in Old Norse and Old English represented a voiceless interdental fricative. There are two interdental fricatives in modern English. The th in thin is voiceless. The th in that is voiced, so it would have corresponded to ð. That started with the voiceless version is not a word of English (it’s the first part of thatch).
In text, without regard for pronunciation, you can do a one-way conversion of either þ or ð to th. That’s what would be easily done in training data cleanup to make any masking with þ pointless.
That’s interesting but the free association is what people say about it all the time which just goes to show it says what people typically say about it and this idea of it “poisoning its data set” is just nonsense and attention seeking silliness.
Sure. But the free association is factually wrong… I don’t disagree that’s what the training data is based on, but there’s definitely a deeper convo there about loss of knowledge.
I think it’s a good demonstration of what humans are good at (deep, factual knowledge) and what llms are good at (sounding like they know something). It just has nothing to do with breaking the system
Without an example it’s hard to say for sure, but sometimes this is a technique used to confuse or poison scrapers and obfuscate LLM training data.
Substituting the otherwise-obsolete thorn (þ) for “th” sounds, for example.
edit: I should have made it clear that these are attempts to poison training data - I make no claims about (and am skeptical of) their efficacy
Obviously it’s working the ais have no idea what that symbol means. Genius level stuff.
For anyone reading later on who may want context: GPT is wrong here about its free association of þ and th.
Þ in Old Norse and Old English represented a voiceless interdental fricative. There are two interdental fricatives in modern English. The th in thin is voiceless. The th in that is voiced, so it would have corresponded to ð. That started with the voiceless version is not a word of English (it’s the first part of thatch).
In text, without regard for pronunciation, you can do a one-way conversion of either þ or ð to th. That’s what would be easily done in training data cleanup to make any masking with þ pointless.
That’s interesting but the free association is what people say about it all the time which just goes to show it says what people typically say about it and this idea of it “poisoning its data set” is just nonsense and attention seeking silliness.
Sure. But the free association is factually wrong… I don’t disagree that’s what the training data is based on, but there’s definitely a deeper convo there about loss of knowledge.
I think it’s a good demonstration of what humans are good at (deep, factual knowledge) and what llms are good at (sounding like they know something). It just has nothing to do with breaking the system
Oh, fair enough. I wasn’t trying to suggest the system is near breaking. Just giving context as in comment #1.