Ariadne Conill 🐰

this makes stuff like #strudel really interesting by the way because using strudel is a cross-domain task: you're writing code that generates music

so, when using strudel, are you thinking in terms of music? or in terms of language (specifically javascript)?

the tl;dr answer to those rabbit holes are "we're not really sure, but we've mapped a lot of foundational neural networks"

and this is the thing about language as the only mechanism of I/O to these generative AI services, frequently it is easier to just do a thing rather than describe the thing that should be done

there is a whole rabbit hole that we could go down about whether or not people actually think in terms of language (the tl;dr answer is that it is context-dependent, and the cognitive mode usually stays in sync with the active task), but that leads to other rabbit holes like "what even is cognition?" and "how do multiple streams of cognition collapse into observed first-person consciousness?"

the underlying point being that largely these things could likely be overcome with better prompting, but the people who would know *what* to prompt are more likely to just go make the thing themselves

it is the same way with these AI music models like suno

even if you apply a low-pass filter to delete the weird harmonics those models produce, the songs themselves are basically soul-less marketing jingles because there is no real story and no emotional tension

i suspect a user who spent more time crafting their prompt could add context that produces a result with more depth

but that kinda defeats the marketed purpose of image diffusion: low effort visual elements to fill spaces where visual elements are expected

anyway, i am a fan of composable intelligence. it makes a lot more sense than trying to build large monoliths that need ever more compute to work.

the far more interesting "AI civil war" is between the monolithic model advocates (who largely are bought into that architecture because of capital pressure) and the modular intelligence advocates (who are bought into that architecture because it is actually sustainable and does not need 9 gazillion datacenters)

trust me, after a quarter or two, your employer will realize this is not worth it

i for one, do not have an employer that tracks me on token use, or imposes any AI requirement of any kind, but, you know, i hear about a lot of people in that situation

if your employer tracks you on token use, the answer is not to burn those tokens on mediocre FOSS maintenance

the answer is to burn those tokens on your shitty employer's codebase

they asked for slop, they should get slop