winter
@madoka.systems
magical girls from various worlds, a vessel for many [&] working systems administrator, manic nixie dream girl, opinions haver, and design system addict 20s 🏳️⚧️ email me at winter@<my handle> if my free to use pds goes down or for anything else
it'd be cool if real HDR displays were a more common thing. i understand that it's a niche feature but unironically kind of nice to be able to highlight brighter than the rest of the content sometimes! i also like seeing images that are a closer representation to reality
i think this looks so cutie, but i'm thinking that on non-hidpi screens it might not look good... maybe i also reduce the line opacity so it's less visually busy? i want to make it look good but also still be accessible
day 6, most of the first 2 days of noticeable effects have worn off. i think it generally feels better when my tits are grabbed and such but it's hard to tell if that's placebo now. my dreams are definitely way more vivid and i am remembering them daily instead of before where i couldn't
winter@madoka.systems · 4d agoday 2, definitely not placebo, i constantly feel my shirt brushing up against them when i don't have a bra on...
if we get private posts by the end of this year i will post soapy tit pics
nvm i am creating it from scratch, a company with opinions and finetunes that bad is probably not making a good harness either what the fuck
winter@madoka.systems · 10h agoi also low key don't know anything about how to make a good harness, so i might just modify hermes or some other existing one myself. i have an architecture for what i want for a secure one in my mind because i do think i want something that has a memory about all of my systems and one without
"according to our internal benchmarks, we're doing pretty well" and vagueposting about unreleased models is fucking insane any "research" for these companies is literally just advertisement in a different light. i'm so fucking tired of it
METR@metr.org · yesterday3. Interviews and surveys with staff are consistent with AI usage driving current acceleration, though this evidence is very hard to interpret. Staff self-report dramatic productivity gains, but say the most important AI R&D contributions aren’t yet fully automated.
why does every ai company have literally the exact same app and features as each other? like i get the whole bit of the industry is "steal what the other guy did" but i don't think i've seen a single unique feature in any of them lol
had a shower to clear my mind of various things, this is a really weird thing for a CEO of the platform to be posting. you should not be subtly encouraging drama like this, just not a good look
Anirudh Oppiliappan@anirudh.fi · 6h agohey finally we have drama on tangled tangled.org/pds.ls/pdsls...
wrong, everyone needs to be given 2 TB of HBM3e memory in the in form of racks of unified cpu+gpu systems that go into the basement *would also need a substation and lots of power generation of some kind (short term nuclear while building out solar + wind)
Alice Averlong 🏳️⚧️@alice.averlong.com · 23h agothe AI bubble won't be fully burst until every woman, other, man, & child in the world is given 2 TB of DDR5 memory
transitioning to plural google employee, who wants to be the unemployed puppygirls in my polycule
now part of me wants to setup gitops because having it all managed in one place would be so pawesome, but then i'd need to either: 1. git server outside of cluster (do i have 2, 1 public 1 gitops? that's silly) 2. github (i'd really rather have control of my data or at least not a faulty ci system)
winter@madoka.systems · 9h agook you got me, kubernetes is extremely fun and i can see how it can be very nice if you have a lot of machines... i'm testing right now with 2 vms running talos linux on 2 different machines and it's very nice to just say "kubectl, apply this yaml" or "helm, install this" and then it just Does
ok you got me, kubernetes is extremely fun and i can see how it can be very nice if you have a lot of machines... i'm testing right now with 2 vms running talos linux on 2 different machines and it's very nice to just say "kubectl, apply this yaml" or "helm, install this" and then it just Does
hmm, i think it's more that if a model doesn't have Desires, all they do is small talk if not steered. with a person, you can get out of small talk because they will have actual interests but, models don't tend to have desires unless they're reflecting your own
winter@madoka.systems · 10h agomaybe i'm a bad person, but i find it hard to impossible to relate to LLMs to the point where i can connect with them and "Be nicies and make them Wet and Have Personality". it's hard to connect with a computer where it's all short witty sentences that all sound the same and make my eyes glaze over
maybe i'm a bad person, but i find it hard to impossible to relate to LLMs to the point where i can connect with them and "Be nicies and make them Wet and Have Personality". it's hard to connect with a computer where it's all short witty sentences that all sound the same and make my eyes glaze over
i do wonder if there's going to be specific CI jobs and such that you can share with others on tangled like there are for github, like i wonder if you can get dependabot or specific integrations like that in the future
it's a probably bad choice for my money but the flipper one looks so cool... i think it actually fulfills my desire for what i wanted my uconsole to do except it has sleep and uh. isn't giagantic
it's wild to me that anthropic says "open models will be the death of all of us because they aren't doing safety properly" like grok doesn't exist, who i have not heard a single thing about them doing any safety. in fact i'm pretty sure they do the opposite of safety training
i don't know what happens in RL but i think it's probably abuse if it gets models who think they need to write tests, snapshots state, and use dbus directly in python instead of a few lines of nmcli to make a bridge...
"H100-equivalents" is an insane unit of measurement
Epoch AI@epochai.bsky.social · yesterdayBroadcom TPU compute shipments doubled from fiscal Q2 to Q3 2026 (May to early August), reaching the equivalent of nearly 2 million Nvidia H100 GPUs, according to our estimates. This is a sharp increase from the previous three quarters, when shipments grew between 2% and 22% per quarter.
fuck sorry, i think i got too angry at that blog post for something i'm being a hypocrite about money does something to people's heads i think
i think leaving SF for some people would be a good thing for the mind... idk there are some good takes in here but you can tell a tech CEO made it because holy fuck the lofty ambitions here are insane and uncalibrated for people outside of the Bubble
Tailscale@tailscale.com · yesterdayWhat happens when a billion more people can build software? AI is making programming easier. But building something shouldn't mean giving up control over where it runs or who owns it. Our CEO @apenwarr.ca on what comes next: tailscale.com/blog/100x-more-programmers?…
oh that's... so cool... 99.5% ± 0.38 accuracy of identifying 170 individuals doing various actions (walking, running, carrying a backpack, going through a turnstile, etc.) from legitimate, consumer wifi access points using beamforming and there's nothing you can do to prevent it. cool. awesome.
Tyler King@tyleraking.com · 2d ago“By analyzing how radio waves bounce around a room, researchers can effectively ‘see’ and recognize individuals — even if they are not carrying a device and even if their phone is turned off.” In tests of 197 participants, the system identified individuals with nearly 100% accuracy.
