I love how this waxes poetically how "The exact words we choose when writing matter.
-
RE: https://mastodon.social/@daringfireball/117106850045916140
I love how this waxes poetically how "The exact words we choose when writing matter. I want any LLM I use to choose the very best, most precise words at every single decision point."
My brother in christ: Then you shouldn't be using a stochastic text extruder.
@tante "it serves to show just how little regard the people behind these generated-text fingerprinting schemes have for the actual craft of writing" – impressive mental acrobatics there…
-
@tante "oh no, now people would see I used AI to write that"
@tymwol @tante I am an AI sceptic, but I came across an argument that kinda makes sense: someone asked whether this watermark would also be there if he used AI for spellcheck, because if it is, there is no way to distinguish a human-written AI-spellchecked text from an AI-written text.
If that's true, it'll be interesting to watch the effects this has on AI usage.
-
@tymwol @tante I am an AI sceptic, but I came across an argument that kinda makes sense: someone asked whether this watermark would also be there if he used AI for spellcheck, because if it is, there is no way to distinguish a human-written AI-spellchecked text from an AI-written text.
If that's true, it'll be interesting to watch the effects this has on AI usage.
@danielaKay @tante you cannot watermark spelling, so it you used it only for spelling, it's a non issue...
To be honest, I'd rather be worried it'd be the other way around: they watermark it by using specific words, etc, so it is a bit like people were treating em-dashes as a sign of AI usage, just now it is done on purpose. So same as before you could get accused of using AI because you used an em-dash, now it could be something else. So yeah, it would likely lead to some new problems we never needed.
-
RE: https://mastodon.social/@daringfireball/117106850045916140
I love how this waxes poetically how "The exact words we choose when writing matter. I want any LLM I use to choose the very best, most precise words at every single decision point."
My brother in christ: Then you shouldn't be using a stochastic text extruder.
The fruit example is especially egregious because a person should know what their favorite fruits are whereas no amount of statistical text generation can answer that question.
-
@tymwol @tante I am an AI sceptic, but I came across an argument that kinda makes sense: someone asked whether this watermark would also be there if he used AI for spellcheck, because if it is, there is no way to distinguish a human-written AI-spellchecked text from an AI-written text.
If that's true, it'll be interesting to watch the effects this has on AI usage.
It doesn't make sense though. A tool swapping the phrase "my favorite fruit is a bonana" with "my favorite fruit is a banana" is extremely different from one that decides for me what I should think my favorite fruit is.
-
@danielaKay @tante you cannot watermark spelling, so it you used it only for spelling, it's a non issue...
To be honest, I'd rather be worried it'd be the other way around: they watermark it by using specific words, etc, so it is a bit like people were treating em-dashes as a sign of AI usage, just now it is done on purpose. So same as before you could get accused of using AI because you used an em-dash, now it could be something else. So yeah, it would likely lead to some new problems we never needed.
-
@Catriona @danielaKay @tante because AI?
-
-
@danielaKay @tante you cannot watermark spelling, so it you used it only for spelling, it's a non issue...
To be honest, I'd rather be worried it'd be the other way around: they watermark it by using specific words, etc, so it is a bit like people were treating em-dashes as a sign of AI usage, just now it is done on purpose. So same as before you could get accused of using AI because you used an em-dash, now it could be something else. So yeah, it would likely lead to some new problems we never needed.
@tymwol No, that's at least one upside: 1. It's done in token space, which is not visible to humans. Tokens don't map cleanly to any human concept like punctuation, words or even syllables. That's not to say that human perception couldn't eventually learn sub-perceptible statistical differences, but … 2. The cryptographic construction with using a key and deciding off of earlier tokens means that the signal is truly 50/50 embedded. A computer without the key can't recover it, let alone a human.
-
@tymwol No, that's at least one upside: 1. It's done in token space, which is not visible to humans. Tokens don't map cleanly to any human concept like punctuation, words or even syllables. That's not to say that human perception couldn't eventually learn sub-perceptible statistical differences, but … 2. The cryptographic construction with using a key and deciding off of earlier tokens means that the signal is truly 50/50 embedded. A computer without the key can't recover it, let alone a human.
@tymwol Basically, if it were constructed to always chose "significant" where it would normally have chosen "important", that's something that would easily be detectable, even by humans.
Instead what it's doing, is choosing "significant" about *half* the time when "important" would be the right word, and choosing "important" about half the time when "significant" would be right. So in the end, the word statistics balance out.
-
RE: https://mastodon.social/@daringfireball/117106850045916140
I love how this waxes poetically how "The exact words we choose when writing matter. I want any LLM I use to choose the very best, most precise words at every single decision point."
My brother in christ: Then you shouldn't be using a stochastic text extruder.
@tante this whole thing is pretty much "tell me you don’t understand how your favourite text extruder works without telling me you don’t understand how your favourite text extruder works".
-
RE: https://mastodon.social/@daringfireball/117106850045916140
I love how this waxes poetically how "The exact words we choose when writing matter. I want any LLM I use to choose the very best, most precise words at every single decision point."
My brother in christ: Then you shouldn't be using a stochastic text extruder.
@tante he repeatedly comes *so close* to getting it here...
-
RE: https://mastodon.social/@daringfireball/117106850045916140
I love how this waxes poetically how "The exact words we choose when writing matter. I want any LLM I use to choose the very best, most precise words at every single decision point."
My brother in christ: Then you shouldn't be using a stochastic text extruder.
@tante
To me it's also like, not everything has to be perfectly optimised. How you write *doesn't* have to be perfect. How you live life doesn't have to be optimised. -
@tante he repeatedly comes *so close* to getting it here...
-
Not the way *he* uses it though, he knows what he's doing
-
RE: https://mastodon.social/@daringfireball/117106850045916140
I love how this waxes poetically how "The exact words we choose when writing matter. I want any LLM I use to choose the very best, most precise words at every single decision point."
My brother in christ: Then you shouldn't be using a stochastic text extruder.
@tante funny how they almost get the point with: "The exact words we choose when writing matter" and then immediately veer off a cliff
-
RE: https://mastodon.social/@daringfireball/117106850045916140
I love how this waxes poetically how "The exact words we choose when writing matter. I want any LLM I use to choose the very best, most precise words at every single decision point."
My brother in christ: Then you shouldn't be using a stochastic text extruder.
@tante I see Gruber's hot takes remain unmoored to reality, he used to be good at this. If you are using Claude then you ain't writing, fucker.
-
@danielaKay @tymwol @tante (AIG is not less crapy and unethical when it's used for spellchecking - and also it's unnecessary, it already works without it)
-
@danielaKay @tante you cannot watermark spelling, so it you used it only for spelling, it's a non issue...
To be honest, I'd rather be worried it'd be the other way around: they watermark it by using specific words, etc, so it is a bit like people were treating em-dashes as a sign of AI usage, just now it is done on purpose. So same as before you could get accused of using AI because you used an em-dash, now it could be something else. So yeah, it would likely lead to some new problems we never needed.
@tymwol @danielaKay @tante Was going to say this also — patterns are likely to transfer from llm to natural writing as people get exposed to the former.
-
RE: https://mastodon.social/@daringfireball/117106850045916140
I love how this waxes poetically how "The exact words we choose when writing matter. I want any LLM I use to choose the very best, most precise words at every single decision point."
My brother in christ: Then you shouldn't be using a stochastic text extruder.
His brain is truly cooked
"LLMs are, in their popular incarnations, non-deterministic. Ask the same question of the same model and you often get at least slightly different answers. Maybe the same meaning, but different phrasing"
"One of my fundamental problems with [watermarking] is that no two synonyms carry the exact same meaning. “He leaped at the chance” and “He jumped at the opportunity” are very similar sentences expressing the same general sentiment, but they are not the same
