luna @l4.pm · 13/09/2026i think he is not happy nowadays (he has a point tho, nothing notable in my life happened from brics yet) 050
luna @l4.pm · 30/08/2026can i convince you to help me and @philpax.me tackle the many things we want to work on we want to pull modes to plugins/userspace so you can do whatever on top: github.com/lun-4/makima... 110
luna @l4.pm · 15/08/2026with help from @karashiiro.moe here's a bench on how different models perform as "auto mode" command classifiers anything more expensive than dsv4f-0731 (save for GPT-5.6-Luna) is not included 2111
luna @l4.pm · 25/07/2026fablepool is a really cool idea the problem is that most people dont have cool ideas 120
luna @l4.pm · 21/07/2026"so, we have Debug, ReleaseSafe, ReleaseFast, ReleaseSmall, right, right?" "now, what if, to integrate with Fil-C ABI, we made ReleaseSafer!" 160
luna @l4.pm · 19/07/2026it still (sadly) amazes me how Discover has a very bad hit rate how much extra data does it need?? 220
luna @l4.pm · 05/07/2026hello! lets' talk busines. wrld: Pedicta by Studio Imago #vrchat #vrchatphotography 000
luna @l4.pm · 30/06/2026doing some nagare work and got this a/b where the experiment is so one-sided it feels i struck gold but i need more time 000
luna @l4.pm · 09/04/2026hotter take: if you work in alignment, you should really watch Code Lyoko 2185
luna @l4.pm · 22/03/2026update for @atvouch.dev: a tangled bot link the repos you own and upon a pr from someone else the bot comments the vouch paths for them 2345
luna @l4.pm · 13/03/2026the fact this is behavior that is reachable thanks to this weird ass command design creates many wishes within me to make my own agent sadly, no oauth inference for non-cc 132
luna @l4.pm · 09/03/2026it happened again what is google cooking?!?!?!?!?? just purged its cache again, but issue persists 110
luna @l4.pm · 09/03/2026sorry for pinging yr personal but @null.namespaces.me which dns server are you using because the domain i verified yesterday already started failing now 210
luna @l4.pm · 20/02/2026wrld: The Way Back by luoyvzhuxiong link: vrchat.com/home/world/wrld_4e4dbfa1-0521-4259-a7a9-e3b34ba79d5f/info #vrchat 130
luna @l4.pm · 05/01/2026the language problem is composed of 3 parts: - acquisition - filtering - ranking acquisition can be somewhat automated via user likes but its not perfect. in practice the testers are clamoring for settings panel, so i shall give one of the first real knobs for nagare for lang setting. WIP UI: 100
luna @l4.pm · 31/12/2025made a little "grid environment" with lovr.org that i enter when i put my vr headset on. df output on one side, network vibe checks on the other its cute and works 140
luna @l4.pm · 07/12/2025i did get somewhere with a scoring model, at least don't know if that is better than random shuffling, will have to test further later 000
luna @l4.pm · 19/10/2025there has been innovation released in the vrchat avatar eyes department! cross-avatar system, reactive on lighting!! so cool!!!!!! iskaur.booth.pm/items/7416098 030
luna @l4.pm · 10/10/2025ran multiple a/b tests, ~300 cast votes across multiple different models for nagare the answer is: everything kinda sucks (v1 on both charts is the original code, what i want to move away from), it's brutal 110
luna @l4.pm · 05/10/2025for context's sake, this is the original shown on bsky dot app's front page: 240
