176B on two desktops: running Qwen3.8-Flash-Next across a pair of GB10s
A 176B model does not fit on one desktop. It fits on two — if you let them share a cable, and if you are prepared to distrust almost every number you measure along the way.
// ARCHIVE
All transmissions
BattleLab v0.18.0: see what your AI coding agent changed
BattleLab v0.18.0 adds a FILES panel and a GIT tab to every AI agent session: browse the working tree and read the diff without leaving the terminal.
Handing a live coding session to a different AI agent
BattleLab hands a live coding session from one AI agent to another — Claude to Codex, Codex to opencode — carrying an editable brief of the work so far, so the next agent starts with the context already in front of it instead of a cold terminal.
Forgehand: the harness that lets AI agents ship my code
Agents plan in issues, an AI reviewer gates every PR, approved work merges itself, and deploys are watched until the change is verified live. A tour of my AI development harness — released as forgehand, a generic MIT-licensed kit for your own Forgejo.
BattleLab: one deck for every AI coding agent
A self-hosted, mobile-first command center that puts every AI coding agent — Claude Code, Codex, opencode, Gemini, Antigravity — on one deck: organized by project, summarized by AI, and reachable from your phone through a blind relay.
Teaching an AI to Write Escape Rooms That Don't Cheat
Gameplayce can now generate a cooperative escape room from a prompt — a story, interlocking puzzles, a final code. Generating one was easy; the week went into stopping the model from quietly cheating.
Getting ACE-Step to make actual music on a home RTX 3090
How I got ACE-Step v1.5 XL Turbo to stop making broken almost-music and start producing usable local audio on an RTX 3090.
Running NVIDIA's SANA-WM video world model on a home RTX 5090
A build log: getting NVIDIA's SANA-WM video world model running on a single home RTX 5090 — what fits in 32 GB, what doesn't, and the gotchas that cost real time.
Running MiniMax-M3 — a 400B model — on two desktops
How we serve a ~400B mixture-of-experts model across two NVIDIA GB10 desktops with llama.cpp's RPC backend — stats, gotchas, and a downloadable config.