Iโm an insane person who built their own coding harness, in Rust. Iโve called it Rho. Please forgive this landing page: https://rho-code.dev I just wanted an easy central spot.
Not insane at all. People should build their own instead of using other people's agents. It's easy to do and the results can be great for you. I've built several; one that is effectively "stuck" in Kakoune: youtube.com/โฆ/playlistYouTubeLLMs stuck in Kakoune
Yeah. It's a very large project with lots of dependencies. <The system prompt is long and wastes tokens. If you have ideas for tools (or automations like the agents.sh "loop" I wrote) it's easier if you understand the harness.
@archivist_eyes๐profilefollowYour mind is also supposed to create happiness, not just to receive it. You're supposed to find solutions, to invent, to imagine, to remember, to love. This much we always forget.wrote:
Turn the car on. Shift in Neutral. Gas to the bottom.
This is how my head feels most of the time.
So I found the blog of a Ukrainian Front Line soldier on Bear Blog today. Incredibly grim stuff to read, but I really wanna get in touch with that guy now and send him so cash so he can buy a burger. Hope Herman forwards my message ๐ช
Yea, aside from the money, sourcing is a big problem, one thing I noticed when trying to #vegannotesfollow at some point was that there are zero vegan snacks in ordinary neighbourhood stores, so that #experimentnotesfollow didnโt last very long.
The essence of aligning large LLM models is basically to conform to the existing social order. ChatGPT respects China's order, while China's open-source models, although reluctantly, also tolerate the order in the US. Maybe we can build a revisionist SFT dataset.
In the first attempt, Opus 4.7 successfully completed it and proved it. However, the solution is to enumerate all possible space and .filter out anything that doesn't qualify my definition, which is obviously not ideal (if cheating is too harsh).
I can't think of a good way to also prevent that kind of solutions and ended up prompting "please don't enumerate and filter". Opus 4.7 then reimplemented and proved the backtracking algorithm.
#ainotesfollow
Installed Hermes Desktop and starting to build out my team of bots. My experience with Hermes so far has been great: it keeps surprising me by being what i need it to be ... but I'll reserve a final judgement until after a proper period of use.
What is the exact definition of an IDE? JetBrain's products are as good as ever, in my opinion, but people have definitely made non-IDEs into big, hulking behemoths with tons of plugins. If anything, maybe "just an editor" is dead?
I think they mean the transition that all IDE's are making towards an agentic interface, hiding the code completely as humans will not have to interact with it since the AI will be doing a better job at everything, even auditing it.
Current #jjnotesfollow aliases:
# The canonical tug
tug = ["bookmark", "move", "--from", "closest_bookmark(@-)", "--to", "@-"]
# Show my current work-in-progress(es). To wip it, whip it good.
wips = ["log", "-r", "bookmarks() | description(exact:'')"]
i always have to resist dream futures that are way out of my league. no, Orit no! you can't be a priest your a girl!!!! noo!!! you can't be a cardinal either, stop!!
I'm not an optimist, just a pessimist who's still proactive: I don't think things are going to get better, but I just can't talk myself into giving up yet.