S*QCU

ouching at that one

ahhh now this is a videogame

it like, well, you see, i can't explain it.
there's a lot more code needed to improve really basic algorithmic features before there's time to show off what's working better than other image synthesizer tools.
these pictures should explain more than words can.

oh yeah the slider thingy

doing proper numbers

mythopoietic, some kind of actual man-vs-nature & man-vs-fallibility-of-oaths system of conflict

bsky.app/profile/norv...
"by scialabba in his shifting perception of hitchen's eloquence..."
dark souls item description prose

there are actually some ways you can use an advanced ai image generation model which has a really obvious motif (BROWN lol its SEPIA in here haha) to inform dataset design, filtering, annotation, and maybe some kinds of augmentation? but it gets trickier and more labor heavy.

unfortunately the statistical error signature to their model (dude BROWN) is so exquisitely varied in its presentation and manifest forms that they are really gonna struggle to use their current tech for synthetic data generation

learning to relax the tension/attachment/binding which yearns (both) to game (and to suppress the yearning for gaming) will release a drain on your attention which necessarily must be present both while gaming and not gaming

wu wei
relax expectations for accomplishment through deliberate effort
balance seemingly purposeful activities with seemingly purposeless ones (going on walks, attentively preparing simple foods, flickshotting heads) to better understand the cost and value of either

type of guy who does a wavelet transform on his kids and says 'look its still peaking second' to the stats poster

i was mortified for a moment that someone was writing arithmetic units from first principles and physically sighed with relief when it was factorio belts on the right hand side

sparse coding of emoji so that storing a large maximum number of emoji per post doesn't crank db storage costs to product: dictionary_size, user_interactions

"even though we know we see phenomena, not noumena (easily revealed by closing your eyes and drinking from the obscured cup) we still reach for phenomena and feel that we're adressing noumena" is, despite it all, something to be said in every generation

some of the most incredible necroposts in the history of net forums are appearing in that lesswrong thread

scribble backlog

i have a hunch that adding architectural details to the hypernetwork's strict task will somehow make learning the effects of architectures explicit and therefore easier (with scale i suppose) than trying to avoid the topic. might make synthesizing training samples easier too!

we can then train our hypernetwork synthesizer inside of this variable upscale ratio autoencoder's 'latents', and get very very very big networks 'S' by using our hypernetwork to *specify* networks within the compressed autoencoder format, then choose very large upscale ratio conditioning inputs

for bonus points, if we want our S to be even bigger, we can train a weird neural-network-autoencoder to compress large output networks and small output networks into similarly-'sized' latents, using a conditional input to dictate upscaling factor from our latents.

now with this suggestion already here, we have a kind of silly way to make a really big output network S!
all we need to do is let the hypernetworkization model write the output network serially, one relatively detailed tokenized chunk at a time!

this means that a valid input to this network can go:
[QPROJ _ _ _ _ ] [K PROJ _ _ _ _ ] [V PROJ _ _ _ _ ] [activation] [swiglu] [FFN [hyper1 hyper2 hyper3]] [UNEMBED MATRIX _ _ _ _ _ _ _ _ _ _ ].
underscores are here to emphasize sequence-length of inputs.

and make sure that the description of an input weight collection tokenizes boring operations like normalizers, feedforwards, activation functions, and hyperparameters for boring functions so that they have an explicit and consistent representation within its learned embeddings of datatypes.

pass the hypernetwork a collection of trainable matrices (multi-'token'-spanning datatypes) with delimiters to separate matrices of variable size (see llava and wan21's autoencoder notes for variable size tiled input encodings) so it can ingest itself for recurrence...

i'll do my best to try to find a way to make it substantially smaller than S.
basic tricks: compose the hypernet from factored matrices wherever possible (so its outputs are bigger than itself)

you could say i'm a bit of a TESCREAList