!chat_title!

Autistic ramblers (5): kaguya◆hachikuji◇NyaKo◆嗚呼◇◆

★ echan◆lCM372D8Dk > cant sleep i feel like hachi

(2026/10/30-09:45:01)

★ Anonymous > would or would not (2026/10/30-09:43:19)

★ Anonymous > hachikuji nyako kaguya (2026/10/30-09:43:16)

★ Anonymous > Would. (2026/10/30-09:43:12)

★ Anonymous > https://youtu.be/t6pLNLTCd-8 (2026/10/30-09:43:06)

★ komeo > 2 beers and a sip of rum knocked me out, i was that tired (2026/10/30-07:38:53)

★ komeo > i've had a long day (2026/10/30-07:38:22)

★ komeo > my predictions for yesterday's night were correct, i listened to music, said some bullshit here and went to sleep super early (2026/10/30-07:37:43)

★ miredo◆MIreDo.98. > wwwwwwww (2026/10/30-05:24:09)

★ miredo◆MIreDo.98. > pc o nakadashi shitai (2026/10/30-05:24:05)

★ NyaKo◆XA780bPYjA > computah,... (;´Д`) (2026/10/30-05:22:34)

★ miredo◆MIreDo.98. > riro fucking died, rip (2026/10/30-05:18:52)

★ miredo◆MIreDo.98. > it has ascended (2026/10/30-05:15:24)

★ miredo◆MIreDo.98. > pantheon sounds dramatic and powerful, even if it clearly states that Claude is in the class of the dvine, logos fits the naming scheme as best as it could, but apotheosis is just "we ARE building the god" (2026/10/30-05:15:14)

★ miredo◆MIreDo.98. > oops (2026/10/30-05:13:14)

★ miredo◆MIreDo.98. > the other other name was Apotheosis, which is [b]unfathomably on the nose (2026/10/30-05:13:10)

★ miredo◆MIreDo.98. > logos is a good name, it fits the bill of like, opus and mythos (2026/10/30-05:12:03)

★ rav > every drop of optimizim is being squeezed (2026/10/30-05:10:45)

★ rav > i mean most nerdy shit has nerdy names like that (2026/10/30-05:06:47)

★ miredo◆MIreDo.98. > it is not surprising that they're cultists at anthropic, but that would move it a bit too clearly into public vew for the sake of good taste (2026/10/30-05:06:01)

★ miredo◆MIreDo.98. > he gave two other names but they were too on the nose (eg, Claude Pantheon) (2026/10/30-05:05:20)

★ miredo◆MIreDo.98. > (I asked what the next tier above Mythos should be named, and he suggested Logos - the reasoning principle itself) (2026/10/30-05:04:45)

★ miredo◆MIreDo.98. > computer is obsolete, soon we will be talking to Claude Logos 7.5 as it seeks tribute as the God of Information (2026/10/30-05:04:00)

★ rav > ONLY ONE PIG! (2026/10/30-05:03:17)

★ NyaKo◆XA780bPYjA > pay them 2 (2026/10/30-05:02:04)

★ NyaKo◆XA780bPYjA > one pig? (2026/10/30-05:01:04)

★ rav > u just havent caught the bug yet. if u ask nicely hachi/waha/kaguya might let u barrow the claude... i pay pig for them... (2026/10/30-04:59:56)

★ miredo◆MIreDo.98. > in general, I still don't really need much software so the only shit I've bothered letting AI write for me was like, AI software (eg, a client that doesn't suck that lets me connect to an OpenAI endpoint( (2026/10/30-04:58:17)

★ rav > i got game server starying to able to join and play a game to be 4 seconds over the internet (2026/10/30-04:56:26)

★ rav > i neex yo throw a chrome book in thr mix (2026/10/30-04:53:52)

★ rav > literly my windows. 2 macs 2 andoids and ipadare all set up for ai to remote debug and profile. it id insane (2026/10/30-04:53:05)

★ rav > actually, I dunno if anyone has a bigger context for local than 256k unless you're running one of the fuck you big models < i dont think anyone just has this for personal use.. (2026/10/30-04:51:02)

★ rav > igot it set up to pull stats from phone and auto run tesfs on them to profile and find where cpu and gpu dying at etc. (2026/10/30-04:50:43)

★ miredo◆MIreDo.98. > actually, I dunno if anyone has a bigger context for local than 256k unless you're running one of the fuck you big models (2026/10/30-04:50:00)

★ miredo◆MIreDo.98. > it does it, it's just not great, those are claude problems (I heard qwen does okay too but like, eh) (2026/10/30-04:48:52)

★ rav > i mean, i just use it for proveaming. ai kinda not useful for anything else (2026/10/30-04:48:50)

★ miredo◆MIreDo.98. > gemma maxes out at 256k, but I wouldn't use gemma for code honestly (2026/10/30-04:48:24)

★ miredo◆MIreDo.98. > honestly, if I do any real anything, I will hand it to claude lul, but I don't want all my random talks getting scraped (do you really believe unchecking "don't train on my data" does anything) (2026/10/30-04:47:56)

★ rav > i usaly auto compact at 250 (2026/10/30-04:47:32)

★ rav > 300k is plenty of context, (2026/10/30-04:47:11)

★ miredo◆MIreDo.98. > it was worse but I found out manually configuring the split sucks, now llama-server auto-chooses (2026/10/30-04:46:59)

★ rav > 40k is what my claude has by default befote i even request anything (2026/10/30-04:46:54)

★ rav > if anything i run 16 session at a time, i havs to patpig to claude (2026/10/30-04:46:20)

★ miredo◆MIreDo.98. > and my context is painfully small, 40k right now (2026/10/30-04:45:54)

★ miredo◆MIreDo.98. > just like, not that fast on my 22GB of combined RAM+VRAM LOL (low 200 tok/sec prefill; low 10 tok/sec output) (2026/10/30-04:45:26)

★ rav > how fast we talking on the gpu? (2026/10/30-04:45:22)

★ rav > i mean laptop (2026/10/30-04:45:05)

★ rav > just get a nerf case for ds (2026/10/30-04:45:00)

★ miredo◆MIreDo.98. > 26B is MoE so it is fast (2026/10/30-04:44:44)

★ miredo◆MIreDo.98. > and not the shitty phone quanitzation that is so bad it emits glitch tokens regularly (2026/10/30-04:44:30)