kedom

Interested in all things computer

[kedom @kedom1337]

@PaulPlayStudio @LoresongGame @theo it absolutely does when one has before spend some time creating a good environment for it to work in. it's not perfect but we are now significantly closer to the extreme of "needing no oversight at all" then we are to the extreme of "needing oversight for everything"

[kedom @kedom1337]

@MattiaManzati @EffectTS_ Isnt tsrs usage blocked due to upstream type issues with Typescript 7.1?

[kedom @kedom1337]

@DavidKPiano @mr_mig_en Exactly, with AI it just does not make sense to "make it work then make it good" anymore besides prototypes. Also making it right from the get go makes sure models don't get lost in the bad decision space of the earlier implementation

[kedom @kedom1337]

@RyanCarniato @ryan_t_brown It's still the same ecosystem story/problem it has always been. With other libraries it often starts (and stops) with something as simple as there not being a really good headless UI library. Then you look at React land and your only problem is choosing which one to use

[kedom @kedom1337]

@joelhooks @pidotdev do you have codemode set to "only" or just "on"?

[dax @thdxr]

what people have to contend with is there is a limited resource and they're competing with people who are way more productive than them for it

it is bizarre though, i look at our levels of spend and it's basically impossible that an early stage startup can spend that

so they're at a huge disadvantage

[kedom @kedom1337]

@willdepue @zaygranet isn't that just because Astra is the first model that stopped emitting CoT on every forward pass and now only do it every few?

[kedom @kedom1337]

A lot of people are talking about how the new models do everything with Python now and if that is even a bad thing...Fable just wrote a Python script to do a simple file edit, malformed the code with an infinite loop, and OOM'ed my machine. At least it apologized I guess

[kedom @kedom1337]

Opus 5 is easily the most unpleasant SOTA model I have used so far. It just does not follow instructions at all

[kedom @kedom1337]

Claude's issue with over-commenting is seriously getting on my nerves. Can't even fix it with instructions either, it just ignores them. And on top of that it is HORRIBLE at trimming/getting rid of them afterwards. pls fix @AnthropicAI

[kedom @kedom1337]

@pidotdev Here are the ones I have built so far
https://t.co/XchKqWDDC4

[kedom @kedom1337]

@TimZolleis @EffectTS_ no not at all, i track it using ccusage and cache hit rate is basically the same

[kedom @kedom1337]

Been adding more plugins to my pi-agent collection. A few of the newer ones:

subagent: minimal subagent implementation without all the usual bloat

usage: shows your Claude and Codex plan limits

checkpoint: keeps `/tree` and the files on disk in sync

all built on @EffectTS_ <3

[kedom @kedom1337]

@EffectTS_ https://t.co/XchKqWDDC4

[kedom @kedom1337]

Built a little collection of plugins for the pi-agent. Two so far, more coming.

The main one is a web_fetch tool, which was almost too easy to build thanks to @EffectTS_'s HTTP client.

Plus every plugin ships with proper Nix/home-manager support ❄️

https://t.co/XchKqWDDC4

[kedom @kedom1337]

@BenjDicken I personally found just having nvim with the copilot.lua extension for light auto completion, the lazygit plugin from snacks.nvim for diff review and then the agent of your choice in a separate terminal tab/tmux pane outside of nvim to be the best approach.

[kedom @kedom1337]

I really feel like the the system prompts in a lot of these agent tools out there are making the models worse. They seem to come from the "prompt engineering" era where models where not that smart yet. I am getting MUCH better answers using @badlogicgames 's PI than @opencode

[kedom @kedom1337]

@ryanflorence Exactly how I have been feeling. I also think people underestimate how fast skills atrophy. One of the reasons for that feeling is probably just me having gotten too lazy to "tradcode" but then I have also probably gotten worse, which increases the mental barrier.

[kedom @kedom1337]

@paulpopus @payloadcms That's great to hear! Is there any sort of rough guess about when v4 might come out and what other goodies it will come with?