Posted by karagenit 13 hours ago
https://github.com/getpaseo/paseo
It works really well, has the nice code diff view which what I used to open vs code for, and it lets you drive from your phone.
All the beefy work happens on my machine at home with big hardware and .env secrets, etc. And from my phone I'm able to run multiple threads with multiple models/thinking modes.
I think a testament of how useful it is, is everybody on my team switched over to Paseo pretty much in a week.
Maybe trivial, but changing the font colors, agent status placement & colors increase my productivity 5-10%. I have more ideas like thumbnail previews, but these are specific to the project being worked on.
If you take pi and its extensions, plus herdr desktop, you have an agent that can run the way you want it to, and an app that shows the info you need in the style and way you want it to. I don't think there is one single superior workflow to dominate everything.
The downfall to this is LLMs being trained for their agents. Opus on Claude Code, GPT x.x on Codex. There are certainly strong echoes to vendor lock-in in the past if you look at all of the APIs we used to have with the social web. You could combine this with AI licensing and a national identity verification system and make the whole thing fucking miserable, with locked down hardware & software like iOS. There's a fair chance this could happen, especially in the EU. That and OpenAI + Anthropic merging are two scenarios that would suck.
the tool descriptions and system prompts are notoriously overcooked but you can work around it with careful config; my startup context in CC is 10k, and i could have pruned it way more if i had wanted to be aggressive about it.
Does using Pi saves you so much token overhead that my maths are wrong?
I built OpenArtifacts (https://open-artifacts.com) to fix this exact problem. It’s free, open source and trivially self hosted on your own Cloudflare account. I’m releasing a hosted version in the coming days. I’d love to hear what other think.
You can run larger flash models on a modern PC using strata. These models perform pretty good even on low quants.
Furthermore, next generation open weight models will swap some compute for storage, making them more accessible for masses.
Not trying to diss on anything or anybody. I want to know.
Sonnet is great for simple work, such as subagents running a program and summarizing the output. Opus is just that much better at the more challenging work I find.
You can probably get Sonnet to reach the same goal, but Opus needs much less hand-holding and overwatch.
If you assume 1 million DAU and 100M requests per day, trimming 1,000 tokens off of [the over-long system prompt] could literally save millions of dollars in inference costs per year