mfw the retro 90s style headphones sound better than they should and they are remarkably comfortable. why did we stop doing this

i am inside your walls :blobfoxastronout:

:blobfoxsleep:

mewing about relay/portal/community rig thats like open registration but posts don't actually leave the relay until approved later. basically facing the problem that walled garden models like nostr moved to are hostile (there's nowhere to go for someone who "washes ashore")

things almost start to feel vaguely okay and then they don't all at once they finally cared for the first time in 3 months about the bedroom because they had to sleep in my chair for a guest. and its like weekly job nag. (except its not even that, now its just "you need to be a disability slave" they've given up any pretext of caring.)

if i did have hardware though i might see how much fun you can get LFMs to do things. (it might still be possible to unsloth them, on just cpu, but idk.) cactus needle is. eh. a bit moaty. its quantization-aware trained and cut to 2-bits, but the version they give you is 4-bit. have been wondering about synthetic datasets for some of that, but eh. probably need to just make the bench model for the weird shit i was doing first. then when that doesn't work we can all feel bad about it :02_ded:

probably toss the game project through one final rewrite/pass over the design doc. then off to start doing mockups. people are job nagging again.

introducing goonmart, a gnu taler exchange for gooners :blobfoxgoogly: mastercard hates this one weird trick.

amused that Odin actually just has vector types. well, arrays with syntax sugar. but close enough.

step 1: custodial bunker signing servers step 2: normies start using them because basic step 3: data breach all nsecs leaked :02_flushed:

> These results suggest that, in the on-device regime, most of the benefits attributed to recent hybrid SSM/linear-attention blocks can be captured by their short convolutional submodules plus a small number of global attention layers. We therefore carry forward designs that minimize global blocks while prioritizing inexpensive, gated short convolution blocks elsewhere. The released dense backbones (Section 2) and the sparse-FFN placement in LFM2-8B-A1B (Section 2) instantiate this recipe. Detailed speed and memory measurements appear in Section 2.4. looks like the findings from liquid (and to a lesser extent cactus) AI companies is: transformer stacks don't win when cost is a metric. the stack instead convolves the soup and only spends attention on the last step.

:blobfoxsleep:

soundcraft seems to be suffering. its perhaps the least usable of the artcraft suite, at least for anything midi related. then again its a wonder that project has lived past a week as it is.

chaotic evil: supporting gnu taler wallets instead of zap/buttcoin :blobfoxlaughsweat:

emoji broked :02_popcorn: :blobfox0_0:

tale of the slop - someone makes a new llm product - it gets popular because it does something - it turns out to be completely and utterly unmaintainable - big drama occurs :02_popcorn:

i often hope i will feel better after sleep. i do not feel better after sleep.

wonder how much of beads is beads-specific magic and how much is just "any task list bolted on with MCP." i barely use the things let alone MCP servers.

found the fopses :blobfoxbongo:

> global timeline in amethyst auto updates and scrolls so you can't actually read anything 😰