Open-source dev. Airdrop farmer. Building with my LLM bros. Rustifying things. @doubletop_io Farming notes (RU) → https://xitter.cf/t.co/ohBFWgSUKl
Joined October 2020
- Tweets398
- Following404
- Followers149
- Likes281
I counted how often I swear at my coding agents.
Not swearing in general, only when it's aimed at the model.
Opus 5.5: 4.9% of my messages. Astra: 18.2%, almost 4× more.
No surprise which one I enjoy working with.
Both come from cctrail, a terminal app I made: time, tokens and cost per project and per model from your local agent history. Each request is priced at the API rate of its own day, so a later price cut doesn't rewrite your old costs.
It reads Claude Code and Codex, and since this release @pidotdev and fx by @vercel too.
Try it: cargo install cctrail
github.com/vladkens/cctrail
And here's how they built it: 22 minutes of coding, testing, and tweaking at 32× speed.
Music in both videos: Kevin MacLeod — “Fluffing a Duck” / “Cipher” (incompetech.com).
Claude Desktop got animated hearts, so I wanted them in my terminal too.
Adapted the effect for Ghostty with Codex. Works with any coding agent in any window.
Completely unnecessary. Keeping it.
Shader + setup instructions: gist.github.com/vladkens/12b…
Credit to @localhost_5173: the heart ripple math comes from MyGo’s wind demo. Astra adapted it for Ghostty with cursor ripples and some rendering optimizations.
Coding agents took the routine off my hands, so I finally did what I kept putting off: making macmon lighter.
Its own CPU, M2, 1s refresh:
- TUI + process list: 5.5% → 2.9%
- TUI alone: 1.4% → 0.9%
- pipe: still 0.5%
Out in 0.10 with 9 new features: github.com/vladkens/macmon/r…
wavo 0.3 is out: local speech-to-text in pure Rust, now faster than transcribe.cpp on every GPU I tested – Apple M2 & M5, NVIDIA RTX 4060, even an Intel N100 iGPU. Whisper ~3× faster on NVIDIA, and on M5 it uses the new GPU neural accelerators.
github.com/vladkens/wavo
wavo got its first user. Their feedback: "I don't get CLIs, make it dead simple."
Done. Now it's just: wavo -y talk.m4a
It downloads the model by itself and prints the text. 25 languages, on your Mac's GPU, pure Rust.
github.com/vladkens/wavo
Remember the rabbit hole? It's called wavo now.
Speech-to-text in pure Rust on the GPU: Parakeet, GigaAM and Whisper, a 7-minute video in ~7 s on an M2. Pulls models like ollama.
cargo install wavo
Added kill to macmon. Select a process, press k, terminate or force kill.
It started as a power meter for local LLMs, and somehow it turned into the thing I keep open all day.
I also can’t stop coding with AI agents lately. I recently released a new version of macmon with a process list.
And now I’ve added multi-process filtering, so you can compare several processes at once — and it’s amazing!
OpenAI says Astra is an autonomous agent designed for long-running tasks, blah blah blah.
But how can an autonomous agent work if their own harness can randomly die at any moment and needs a user to poke it to continue?
And of course, the company that supposedly solves complex math problems and so on can't even set up agents to scan their own GitHub issues, pick out real bugs, and fix them.
github.com/openai/codex/issu…
github.com/openai/codex/issu…
cc @thsottiaux
My feed is full of posts about apps people have built recently, often just different implementations of the same idea.
In the AI era, how is it even posseble to choose software when installing and trying someone else’s app takes only slightly less time than building your own from scratch?
Checked it after half a day of mining — 0 NFTs.
How it went: I asked Codex Astra to extract the mining logic, parallelize it across rented GPUs, build the control plane, and deploy everything. The software ran fine, but by the time I got it all set up, GPUs were almost impossible to rent – and now the mint math is already bad.
Main takeaway: I spent too much time thinking about code and not enough about execution. With agents, the bottleneck is increasingly not coding – it’s defining the task well and handling everything around it.
Going back to RB LP farming
