typ sh

Ok, I finally get what Uhura was doing, she just had 120+ agent sessions running at all times with the button colors indicating the state (working, turn finished, ask_user_question, waiting for subagents), that's also why she never took off the bluetooth headset

Flo 🔶@faz.ms · 2w ago

Think I found my perfect way to keep track and switch between agent sessions

sucks when you procrastinated on <task> by making an app to do <task>, but now you actually have to use it for <task>.

I love how I wrote this and thought the world wasn't ready for it

stochastic parrot but it's, like, a really cool parrot who skateboards and smokes cigarettes and shit

all of the TUIs have comically bad UX. it feels like when the internet first started being a thing and people made a million horrible websites. similar to the early web, i'm convinced that all the pieces are basically there but nobody's put them all together yet in the right way.

tbabb@tbabb.bsky.social · last wk.

We all have TUI stockholm syndrome, and this is all going to seem very silly in a year or so

feel like there ought to be a term for the principle that llm use is compatible with investing *much more* care in a thing you make than the median non-llm assisted version of that thing.

Axe Ghost. On Steam!@axeghostgame.bsky.social · 6mo ago

[friend hands me a classic novel] "if you didn't give enough of a shit to write it why would I give enough of a shit to read it?" the reasons this wouldn't be appropriate are the same reasons it doesn't work as a general statement about llm writing.

Jev can talk if given a predictive keyboard, but its preference is mostly just to mash "no"

Typing of the Jev@jevons-talking.bsky.social · 6d ago

no 2 picks · 5 Jev calls · mean confidence 0.67

i have a mouth but not want to speak. 16 picks · mean confidence 0.56

Current #LocalAI model picks on my 32GB M1 Max: Main model: Jundot/Muse-Glimmer-30B-oQ4e Small model: UraionLabs/MiniCPM5-2B-oQ4e They fit together because both are efficient at long contexts (most are not). It's already tight, so incorporating a Jev-style model will be a challenge.

I'm joking but it really is currently a pretty blank canvas, especially in FOSS. The big players have at least started to put little animations and gradients there so the wait feels a bit shorter. I think you can go further.

Paul@solarkraft.bsky.social · 5d ago

those brainrot ide guys were on to something but they shouldve put ads

those brainrot ide guys were on to something but they shouldve put ads

Paul@solarkraft.bsky.social · 5d ago

Why has nobody monetized our attention while waiting for AI responses yet?

Why has nobody monetized our attention while waiting for AI responses yet?

“prime deals you don’t want to miss” and then it’s all deals you want to miss

I do have to say that it's largely pretty effective. Lots of sub-sessions in this session. Decent backgrounding. Things eventually get done. Just a bit intransparent.

Paul@solarkraft.bsky.social · 2w ago

OpenCode subagents, aight

OpenCode subagents, aight

when are they inventing degenerative ai

Cosmic Smokescreen Image date: 12 December 2022, 06:00 Credit: ESA/Hubble & NASA, ESO, O. De MarcoAcknowledgement: M. H. Özsaraç Source: ESA/Hubble

GPT 6.1 Sol seems to return some reasoning detail! (on Codex, I think the API always had more)

How are y’all keeping your agents happy? I haven’t found myself able to give them work for more than an hour before without intervention. I’m such a massive bottleneck and guess J need to orchestrationmaxx …

I’m looking for an LLM-friendly task tracking system that I can dump ideas into, optionally prioritize and just let the LLM attack tasks that are unblocked. Could beads fit? Anything better? I find that this is a pretty strong bottleneck.

Has anyone been able to get Jev-style decisions out of a fully usable LLM? I mean a hybrid model that can generate both text with full decoding AND efficient decisions. Should be possible, right?

I'm upping my p(bloom)

we’re spending trillions of dollars to domesticate text elementals and i think analogizing to the first wolves that decided to sleep near our fires is appropriate

philpax@philpax.me · 2w ago

I think this is still very much a major unsolved problem in persistent agent operation: you imprint on them, even when you are trying not to, and they will share your opinions and interests as a result you cannot imbue them with true independence or the ability to disagree

Scheduling should be a solved problem by now. Have your agent coordinate it with my agent. This is my "where's the flying car" opinion.

The Dark Prince of Donair@erisianrite.com · 2w ago

If calling to make doctor's appointments enriches your life, get better hobbies

tidal hydro is lunar power. makes u think.

Why don't you trust me?

I just want everyone to know that I bought 2 SSDs before the first version of ChatGPT came out