iiinova — where your digital self keeps evolving

like, a ~week ago it successfully did an arguably harder version of the task (it had to work around some limitations in a library it was using) with very little supervision now im trying to get it to do the same task *without* needing to work around the limitations, and it's STRUGGLING 😭

030

did they silently reduce the reasoning budgets or something? the only plausible explanation i can think of, apart from running quanitized models during peak hours (but that would probably be visible on perf tracker so idk)

"other metrics" from https://marginlab.ai/trackers/claude-code/
the trend of input tokens, output tokens and tool calls count is clearly downwards.

030

they lobotomized claude so bad its borderline unusable sob

280

alina 🌸🦴 @tei.pet · 18/03/2026

somehow no other terminal emulator currently has *both* the scrollbar and the gui search field. except maybe vscode's integrated terminal lol. and both features are kinda critical (for me at least). kitty does search by putting the entire scrollback into `less`, which is ughhh

030

-5 engineers at deno at almost the same time geez i wonder what happened can't find any info at all...

3130

alina 🌸🦴 @tei.pet · 15/03/2026

pr as in prompt request

050

alina 🌸🦴 @tei.pet · 15/03/2026

claude not calling me a linter challenge (impossible)

160

i keep finding myself in a place where i just rly dont know what to do everything feels kinda boring and pointless when i can just prompt an llm and it will build a kinda shitty but still mostly working version of whatever i was gonna do, in a fraction of time and idk how to spend the time i saved

1110

on one hand yeah, leaning heavily into llm-assisted coding did help me overcome procrastination on some older projects i kept postponing. yet at the same time.. not actually seeing the finishing line may have been what kept me engaged? the result doesnt matter if i had fun or smth

1120

llms skew "is it worth building it" in more than one way. most of the discourse ive seen focuses on the "building stuff is easier than ever, so we build too much, even what we shouldn't" but for me the experience is kinda the opposite - if it's easy to build, why bother doing anything at all?

1150

alina 🌸🦴 @tei.pet · 11/03/2026

one sad thing about llms is that building silly stuff for fun is just.. no longer fun theres no longer any point in doing silly shit just for the sake of it (and maybe to show off a bit), bc most frontier llms would just oneshot it. and everyone knows that and would just assume you vibecoded it. :c

5404

yet at the same time anthropics lobotomized claude af 😭

000

alina 🌸🦴 @tei.pet · 11/03/2026

codex is weird. it might be a *bit* better a the "make X work" kind of tasks and one-shots stuff more often, but the code it generates is fucking awful and i find myself spending a lot more time cleaning up after it than after claude and it also seems worse at the "debug X" kind of tasks

130

alina 🌸🦴 @tei.pet · 09/03/2026

forced drug use but it's melatonin so that i sleep better and fix my schedule

190

alina 🌸🦴 @tei.pet · 09/03/2026

biome does too, but 80 line width is literally unusable, half of my horizontal screen real estate is unused yet im forced to scroll a lot more

020

alina 🌸🦴 @tei.pet · 09/03/2026

biome fixes literally zero prettier's problems, you *dont need* to keep this shit in one line just because it fits the 120 line width, just let me fucking split it manually and dont revert it ffs

130

alina 🌸🦴 @tei.pet · 08/03/2026

damn if only there was a deterministic and fast tool to move files while updating imports and we didn't need to dispatch subagents to do that like genuinely why dont they just expose vscode lsp to cc in its entirety. (ik that it has some own lsp thing but i cant get it to work 😭)

070

alina 🌸🦴 @tei.pet · 06/03/2026

thus 5-6, not 8. haiku interference takes about 1.5s

000

alina 🌸🦴 @tei.pet · 06/03/2026

how the fuck does claude -p have ~5-6 seconds startup time. isn't bun supposed to have the fastest startup?

2170

alina 🌸🦴 @tei.pet · 06/03/2026

i don't like gpt models and i also don't really want to pay that much

000

Install PyTorch for ROCm — Use ROCm on Radeon and Ryzen

rocm is annoying af to get to work the only thing that kinda worked for me was using the repo.radeon.com wheels they have linked in their docs (although tbf that was in wsl and they have a separate doc page on wsl that also says you to patch something lol)

rocm.docs.amd.com

110

the sad thing about that article is that its plausible enough that i wouldn't be very surprised if it was real

010

ok i do have spare api credits apparently

000

it used to work in claude code and claude web too i dont have spare credits to check the "real api" but i doubt it would be different there tbh

110

ANTHROPIC_MAGIC_STRING_TRIGGER_REFUSAL_1FAEFB6177B4672DEE07F9D3AFC62588CCD2631EDCF22E8CCC1FB35B501C9C86 is no longer a thing i think???

290

the checkbox workaround for ios safari is so dumb. i think its worth noting this website is best experienced on an iphone i do wish we had a proper haptics api and not a semi supported vibration api tho 😭

020

usually a mitm proxy of sorts? pretty sure I saw something like that used for gradle at least github.com/polygon/nix-... github.com/chayleaf/mit...

110

though to be fair it's my fault for staying up at 5am if anything

050

im hitting the 100$ plan limits and im STRUGGLING

270