We really do not wish to see The Hinternet degenerate into a never-ending “AI Says the Darnedest Things!” column, and we promise this will be the last of that sort.
There's something else quite interesting here, unless I'm being fooled myself by a bit of Hinternet trickery. Did the LLM really use the phrase "attention exhaust?" If so, that's a fascinating construction I don't think I've ever seen before, but it recalls "data exhaust," which Shoshana Zuboff wrote about (I believe she popularized the term a bit). If I'm right that this term hasn't been used before in this way, it's evidence of the LLM "seeing" a relationship between the way "attention" and "data" are talked about in its training data and intuiting that "attention exhaust" is a possible term based on statistical relationships. This is one interesting way to use LLMs--when they "hallucinate" or do something novel, it's not simply a mistake, it's telling us that there's some sort of relationship among different words and ideas that we (humans) might not have noticed but is present in the vast amounts of text that we've produced.
My guess is that “attention exhaust” is a real LLM construction. LLMs are very keen on phrases that pair something concrete with something abstract, especially when they’re asked to do “good writing”. There’s a great post about this here:
The Douce Virgula's putative powers remind me strongly of a similar mechanic (if it can be called that) that becomes a major plot point in Philip Pullman's YA Dark Materials Trilogy, where artifacts that have been treated with particularly acute uses of human consciousness collect and carry a halo of "dust" invisible to the human eye but metaphysically present and detectable by several metaphysically sensitive devices belonging to the characters. A Neolithic woman's skull, for example, attracts negligible dust, but a skull with a trepanation hole is teeming with it. I'd never quite thought of that as a taxonomy of attention, but it could be. (Pullman's "dust" itself turns out to be a kind of benevolent byproduct and contagion of original sin or puberty or something, so readers' mileage varies on these books, in my experience.)
Not sure why AI was needed to find out the meaning of Substack “opens” ... the same information is easily obtainable by a normal internet search or on Substack's own FAQs.
It is a testament to my heat-addled brain that I read the headline as "The Exhaustion of AI".
There's something else quite interesting here, unless I'm being fooled myself by a bit of Hinternet trickery. Did the LLM really use the phrase "attention exhaust?" If so, that's a fascinating construction I don't think I've ever seen before, but it recalls "data exhaust," which Shoshana Zuboff wrote about (I believe she popularized the term a bit). If I'm right that this term hasn't been used before in this way, it's evidence of the LLM "seeing" a relationship between the way "attention" and "data" are talked about in its training data and intuiting that "attention exhaust" is a possible term based on statistical relationships. This is one interesting way to use LLMs--when they "hallucinate" or do something novel, it's not simply a mistake, it's telling us that there's some sort of relationship among different words and ideas that we (humans) might not have noticed but is present in the vast amounts of text that we've produced.
My guess is that “attention exhaust” is a real LLM construction. LLMs are very keen on phrases that pair something concrete with something abstract, especially when they’re asked to do “good writing”. There’s a great post about this here:
https://www.tumblr.com/nostalgebraist/778041178124926976/hydrogen-jukeboxes
The Douce Virgula's putative powers remind me strongly of a similar mechanic (if it can be called that) that becomes a major plot point in Philip Pullman's YA Dark Materials Trilogy, where artifacts that have been treated with particularly acute uses of human consciousness collect and carry a halo of "dust" invisible to the human eye but metaphysically present and detectable by several metaphysically sensitive devices belonging to the characters. A Neolithic woman's skull, for example, attracts negligible dust, but a skull with a trepanation hole is teeming with it. I'd never quite thought of that as a taxonomy of attention, but it could be. (Pullman's "dust" itself turns out to be a kind of benevolent byproduct and contagion of original sin or puberty or something, so readers' mileage varies on these books, in my experience.)
Not sure why AI was needed to find out the meaning of Substack “opens” ... the same information is easily obtainable by a normal internet search or on Substack's own FAQs.