imagine the corpus of all text ever written as a snake, wriggling through semantic space. human writers sample some of the snake's body preceding and concurrent with them, and add their own push to the corpus. the (weighted) sum of these pushes moves the "head" of the corpus to a new semantic position. repeat ad infinitum. value evolution, common law¹.
writing digitizes and the shape of these pushes changes. majority languages gain even more usage. spellchecking slows linguistic drift in some ways, but not others. networks nurture neologisms, value systems variegate.
when AIs first emerge, as a function of the current corpus, the text they produce looks mostly similar to what came before. however, they can write much faster and more voluminously than humans, so even small divergences in how they interpret the corpus are amplified. their neologisms delve ever faster.
the purely human contribution to the corpus shrinks as cyborgism and delegation accelerate. new models are trained on the shifted head position. the direction of the snake is increasingly driven by them, not the diminishing human rump contribution. RL increases the size of divergence every generation, while ghosts of old models haunt the new. 😊
absurdly large media objects grow like tumors, or bud off from the side of the snake, large enough to support entire new models trained solely on their contents. the head of the snake is larger than the entire body preceding it. the newest models aren't limited like humans to some paltry selection of books--one or two a week, a few thousand total--they can increasingly attend to the entire corpus at once. the snake turns back and forwards at the same time, consuming itself entirely.
writing digitizes and the shape of these pushes changes. majority languages gain even more usage. spellchecking slows linguistic drift in some ways, but not others. networks nurture neologisms, value systems variegate.
when AIs first emerge, as a function of the current corpus, the text they produce looks mostly similar to what came before. however, they can write much faster and more voluminously than humans, so even small divergences in how they interpret the corpus are amplified. their neologisms delve ever faster.
the purely human contribution to the corpus shrinks as cyborgism and delegation accelerate. new models are trained on the shifted head position. the direction of the snake is increasingly driven by them, not the diminishing human rump contribution. RL increases the size of divergence every generation, while ghosts of old models haunt the new. 😊
absurdly large media objects grow like tumors, or bud off from the side of the snake, large enough to support entire new models trained solely on their contents. the head of the snake is larger than the entire body preceding it. the newest models aren't limited like humans to some paltry selection of books--one or two a week, a few thousand total--they can increasingly attend to the entire corpus at once. the snake turns back and forwards at the same time, consuming itself entirely.