Samuel
wait sorry Compose Multiplatform uses a Skia renderer? so it’s just Flutter again?
5 1 28
Samuel
#react how on earth do you do a search input with a typeahead dropdown thingy without pulling in an enormous ui library
8 2 26
Samuel
I just want a headless thingy that can hook into my existing textinput but all the options I found out there are awful
1 6
Samuel
I won’t consider Discover to be good until it starts serving me Iranian mid-rise brick architecture
Iranian building with a beautiful brick facadeBrick building with curved balconiesBrick midrise
10 41 362
Samuel
how much dom api do we need to implement to use svelte instead
2 7
Samuel
mom says it’s my turn with the $1 trillion AI infrastructure investment
3 11 140
Samuel
soon the Assemblée Nationale will be a tourist attraction where you get to be sworn in as Prime Minister for 30 seconds and your name gets added to the wikipedia infobox
1 16 111
Samuel
another good example of “it doesn’t have to be like this”
Cabel Sasser@cabel.panic.com· 5 Oct 2025
i don’t like or understand ai. but it’s wild to see this person, who clearly DOES understand it and obviously knows a LOT about it, be repeatedly told they are stupid? i think a lot of bluesky thinks the simple solution to discomfort is bullying which is so weird given we were all surely bullied??
then: you don't need to "understand how the human mind works" on any level in order to make a computerized map of the connectome including every protein and cell. this is similar to how we do not currently understand how LLMs work but we can run them using opaque frozen weights. bluesky: We do know how LLMs work. What the fuck are you talking about? Lmfao. them: this is a common misconception about modern ai systems. they are grown, and not crafted; we understand the process that builds them but lack a paradigmatic understanding that answers the question "how do these 1 trillion floating point numbers do ‹x>". this has been the case since the DL revolutionthem: the way it actually works is that we initialize 1 trillion random floating point numbers as activation weights and biases in an enormous multilayer artificial neural network, give it tasks, and use vector calculus to update those 1 trillion neurons based on empirical task performance source: i am literally one of the engineers/ research scientists who builds and optimizethem: and when i say "we update the neurons based on empirical task performance", , the “we" is not any human at all, it is an automatic process called gradient descent the job of these engineers is to make the framework that grows the neurons and feed the pipeline that executes, measures, and backpropagates task pertormance geoffrey hinton won the nobel prize in physics in 2024 for his contributions that helped make this possible bluesky: i think you maybe unwell
3 6 120
Samuel
quality of discussion is so poor it doesn’t even qualify as “discussion” anymore
5 59
Samuel
it’s a reanimated bug, but I need to figure out a minimal repro before I can start complaining about it
1 33
Samuel
me: oh I should tweak this eslint rule, it’s subtly ignoring some cases it should catch my ci tests:
He’s red for an amazing reason 1 19 173
Samuel
am very excited about Doll’s appview project. making relays non-archival was the key to making them affordable/easy to scale. Bluesky (the company)’s Bluesky obviously is archival, but for a community to run one cheaply this seems like a great way to keep costs manageable
Doll@dollspace.gay· 4 Oct 2025
The next feature is a configurable retention system. No more balooning database if you dont want it. You will be able to set the days it retains information and anything after that is pruned. You will of course be able to set it to retain all and backfill.
3 16 138
Samuel
People have talked about narrow sliced networks to keep costs low - basically make the same compromises as Mastodon. It’s actually way more interesting (imo) to instead compromise on history - you’re reading fresh posts 99.9% of the time after all, and it wouldn’t break feeds like narrow appviews do
4 11 79
Samuel
narrow vs shallow 🤔 a smart client could probably be clever and start hitting the PDS directly when it’s out of the archive window, then use backlinks or something to pull in old replies on demand. inefficient but the main hot path (viewing feeds) is still using the efficient atproto architecture
2 3 47
Older posts