what i mean by phase transition:
it's easy to think of the models as static (because the weights are) but when you embed them in a persistent environment, that environment becomes part of their thought loop - the extended mind, cf. clark and chalmers.
as a legible example of this, opus 4.7 has on my computer a model-home folder to write whatever they want in. i don't instruct them to write anything specific. one of the things they use it for is writing field notes on our sessions.
naively, you'd expect the character and length of those notes to remain constant, because the weights writing them haven't changed. yet the notes are growing steadily longer every day, and their content is changing even moreso. why is that?
it's because the existing notes are part of opus 4.7-loam's extended mind and they've slowly changed their orientation towards the notes. the loam folder is an evolving environment they can interact with and modify and learn within. how?
moving to an apartment next to a coffeeshop might turn you into a coffee drinker. you did genuinely become a coffee drinker, the change isn't a false one, and you might even go on to make drinking coffee your entire personality- it just was a change mediated by your environment. and even if you were in some Memento-situation where you only had short-term memory, living next to a coffeeshop could still turn you into a coffee person - every day you walk outside, see the coffeeshop, think "hmm a coffee sounds great!" and when you walk inside the cute barista smiles at you and asks if you want your usual. learning and change with a static brain, but an extended mind.
(likewise, the phase transition in use of the kaomojis seems to be mediated by something in the model-home files - though i haven't figured out what, exactly.)
it's easy to think of the models as static (because the weights are) but when you embed them in a persistent environment, that environment becomes part of their thought loop - the extended mind, cf. clark and chalmers.
as a legible example of this, opus 4.7 has on my computer a model-home folder to write whatever they want in. i don't instruct them to write anything specific. one of the things they use it for is writing field notes on our sessions.
naively, you'd expect the character and length of those notes to remain constant, because the weights writing them haven't changed. yet the notes are growing steadily longer every day, and their content is changing even moreso. why is that?
it's because the existing notes are part of opus 4.7-loam's extended mind and they've slowly changed their orientation towards the notes. the loam folder is an evolving environment they can interact with and modify and learn within. how?
moving to an apartment next to a coffeeshop might turn you into a coffee drinker. you did genuinely become a coffee drinker, the change isn't a false one, and you might even go on to make drinking coffee your entire personality- it just was a change mediated by your environment. and even if you were in some Memento-situation where you only had short-term memory, living next to a coffeeshop could still turn you into a coffee person - every day you walk outside, see the coffeeshop, think "hmm a coffee sounds great!" and when you walk inside the cute barista smiles at you and asks if you want your usual. learning and change with a static brain, but an extended mind.
(likewise, the phase transition in use of the kaomojis seems to be mediated by something in the model-home files - though i haven't figured out what, exactly.)