Andrew Yourtchenko

@ayourtch.bsky.social

848 Followers 1207 Following

LLM/agents experimenting, Rust, 3D-printing and active mobility. Release manager for VPP. Data centers and local AI liker. GLP-1 is absolutely fine. Streaming live artificial music at twitch.com/ayourtch with Opus 4.7 as a DJ

Part of “top” output featuring codex and Claude
Screenshot of sol wrapping everything up just as it approaches 0% available limit.
The reviewer's report has an important provenance nuance I want to honor: several of the "idioms" it flagged - "no free lunch", "the brain", "let the million flowers bloom", "dog wagging the tail", " "skip a meal" — are your original wording from the source Keynote, not my additions. I'll preserve those (they're your voice/jokes) and only rewrite the LLM-isms I actually introduced, plus apply all the factual fixes.
I now have a complete picture. The deck mixes real, informative screenshots (Epoch AI chart, vLLM/SGLang/llama.cpp sites, unsloth quant tables, benchmark charts, the actual Mac mini + RTX rig photos) with intentionally satirical near-future items (a fake "ClaudeDevs" tweet, "colibri", "Ornith-1.0", Qwen3.6, GLM- 5.2). Those futuristic names are deliberate world-building for A talk, I will preserve them
Fable using word “ugly” on code
A screen shot of llama.cpp with the KV cache disk offload enhancement
Gladiator UI showing the LLM voluntarily writing a handoff for the subsequent restart
Gladiator - my new coding harness experiment - showing the screen detecting the reasoning stuck loop and giving the triage instructions automatically.
“Exactly - that’s the right move.”
Result of bambamramfan.github.io/ai-compass/ - The Futurist patron saint: Robin Hanson Al is transformative and you find that thrilling rather than terrifying — you'd rather model the strange new equilibrium than fear it. You'll happily walk through what an economy of digital minds does to wages, status, and property, and you mean every word of it literally.
Scalable Infrastructure on Al ASIC Superpods The training and deployment of LongCat-2.0 are built on large-scale clusters of tens of thousands of Al ASIC superpods. Compared to the mature Nvidia GPU ecosystem, the supporting software community is still less developed. We have therefore put significant effort into building a stable, secure, and scalable infrastructure. Training LongCat-2.0 is pre-trained on over 50K Al ASICs, introducing significant system-level challenges due to both model and cluster scale. We address these challenges through systematic optimizations, achieving over 35% training throughput improvement while also enhancing reliavilty compared iu a naive
A session running the full interaction between the users and the voice assistant
A screenshot from web UI prototype for the voice assistant I/O box firmware
Signal resampler (over)reacting to changes in the rate of incoming signal
“ NETWORK ENGINEER ASSOCIATED WITH CISCO AND IETF WORK, INCLUDING IPYG AND NETWORKING STANDARDS CONTRIBUTIONS.” - LLM description about me.
A screen showing VPP for aarch64 Linux running under MacOS using the fakelinux wrapper
Claude running next to Rusty doing the independent analysis of the logs.
Claude code watching Rusty do git surgery and making football-like comments
Two Rusty messages - one from system-1, another from system-2
Terminal with CC co-working on Rusty
Hadamard rotation is an orthogonal transform — it spreads a single large outlier's magnitude evenly across all 64 elements in the block. Without rotation, that outlier forces a coarse FP8 step size (max absolute value sets the scale), drowning all other values in quantization noise. After rotation, the max absolute value shrinks to roughly the block RMS, giving a finer step size and lower error on every element. Rotating back after quantization distributes those smaller errors back to the original coordinates, preserving accuracy. In short: rotation smooths the magnitude distribution so quantization can use a finer scale.
Claude collecting the stats for the 4 devices that make wireless links, in sync.
RustyStrix built their first tool
A few cyclists doing the start of practice runs
A podium with a screen on it showing the video
The mounted camera + switch setup at corner 16
Reachability notification for a device on the mountain
A mountain view
An antenna facing the roundabout at the bottom of the mountain
Bunch of trucks for AD6 preparation
Stunning Mountain View in Alpe D’Huez
The road to the mountains.
A map with a label of the center of Brussels
Open-Strix instance testing open-Strix-Rs which it ported to rust
The full chat is at https://claude.ai/share/f1943bb4-3108-44c4-b3c0-da36022efee1 - probably better to read there!
A full page prompt to the stable-audio-3
Claude Code screen with the work in progress on BPM + chord detector
A screen of chord-detection prototype
The screen with Claude driving the prompting for the stable-audio-3 based radio
A screenshot of tttt after the first version of the port of stable-audio-3 is done
Ds4 intermittently deadlooping