S*QCU

baudrillard's framing is far more interesting and far scarier, for what it predicts about the ability of simulator-users and wargame-historians to think through nuclear doctrines and use them when situated. nuclear peace was a kind of shared dream, a myth, a fantasy... of the 20th century.

doing interpretive dances for claude code to express the relative positioning that the inserters should have in the steam pipe fitting factory because i refuse to speak while designing pure geometry...

"-mario luigi erotica fixed point in garbagespace"
i can parse this sentence but not understand it

bsky.app/profile/norv...
norvid post

i'm happy that i see so many cyborgist and furry posts and ai-not-enjoying posts and writing posts on the bsky timeline.
someone really did break twitter and twitter is genuinely not functioning in the capacity it did when i first made an account.

this should not be a super challenging adversarial test for computer use embedding/tokenization; there should be continuous tokenization strategies which let you get entire families of curves & steganography features wtihout bloated discrete codebook sizes... all looks quite tractable...

i think the litmus test for whether this kind of action state modeling *generalizes well* is whether you can tokenize and learn digital pen drawing traces in a full digital pen work environment. making marks, selecting masks, changing layers, changing tools...

it's worth uh.. it's worth... it's worth uhhhh reflecting on this tokenization choice as running into really specific problems related to discretizing n-dimensional space into prematurely decoupled 1d segments... i'm sure it works but there are some invariances cooked in very early...

mouse movements are *really hard*.
it's very hard to describe how you 'ought' to tokenize gui actions, but it's definitely through something more like 'a compressed geometric representation of the trajectory of the user interface action over space', not taps at specific endpoints, for example.

i threw way more effort than was worth it (i can't train on my gpu and also do action-state recording using it at the same time) into throwing together a video recorder that pre-filters incoming videos by whatever surprises still image vision transformers models the most

in this picture you can literally see the dimension in which that cat is being rotated

condolences :(

total legitimacy failure in all institutions (larger than a discord server) and total unwillingness of institutions to demonstrate and confirm basic legitimacy

ah, he just like me frfr...

ah, grassmann, one of the few mathematicians to have clearly lived a good life...

for clarification this has never been done before but all of the vector mathematics is trivial (i assure you) and therefore would not constitute a publishable work or dissertation (because any possible reviewer would be too mad that they didn't think of it first)

k:6 classifier free guidance with multiple positive (attraction-towards) and multiple negative (repulsion-from) interventions at a 'gain' 4:

k:6 classifier free guidance with multiple positive (attraction-towards) and multiple negative (repulsion-from) interventions at a 'gain' 3:

k:6 classifier free guidance with multiple positive (attraction-towards) and multiple negative (repulsion-from) interventions at a 'gain' 2:

k:6 classifier free guidance with multiple positive (attraction-towards) and multiple negative (repulsion-from) interventions at a 'gain' or intervention strength of 1:

k:2, or regular classifier free guidance (can't actually apply the extra constraints

so anyways, this is what the k:1-tuple shrimp looks like. it doesn't have 'classifier free guidance' at all, in fact.

obviously this requires writing a distributed scatter gather version of classifier free guidance so that all 6 constraints can be defined and jointly applied

for example, it comes to me in a dream one day.
"i must have a picture of a shrimp which is precisely the opposite of a poem about bananas. also the shrimp should be an example of good typography. also the shrimp should look unalike a low resolution rendering of itself. or an even lower res--"

it's time for the future of diffusers
regular diffusers only let you do 'classifier free guidance' between two different queries or problem representations at a time.
but i have bigger dreams:
what if i want to run a 6-tuple of diffusion model outputs instead of a 2-tuple?

urgent questions demanding hurried answers

one very tangible and literal killer application for photoshop is custom brushes for a brush texture engine which allow the use of an expensive hardware peripheral requiring years of training to make digital art.
what's the dev ux for the krita brush engine like?

zernike rings are PAINFULLY cool but i can't actually use them for the specific selection criterion i want:
compositional image structure matching with a smoothed sensitivity to rotating the entire image.
(yes, i really do want for rotations to be similar to identity)