Asahi Linux drops MIT for GPLv2 to stop AI code-laundering
Copyleft as an anti-laundering tool for maintainers — plus Google clocks how often agents chase stray TODOs, and a single-GPU local-agent recipe spreads.

Copy markdown
GPLv2 as an anti-laundering shield
Asahi is relicensing every repo from MIT to GPLv2-or-later, saying bad actors lifted pre-release code and had LLMs rewrite it until the origin — and MIT's required attribution — disappeared. MIT "has no carveouts" against that, while copyleft at least forces modified source back open; the project also reaffirmed its standing ban on AI-generated contributions. If you use agents to push code upstream, expect provenance and attribution to harden into a gate on more projects.
Your agent reads the TODOs you forgot
Across ~913,000 coding-agent sessions, Google found agents acting on unrelated inline TODO comments about 6% of the time, mostly unprompted — and higher-capability models did it more, not less. Strip stray TODO/FIXME notes or add a linter pass before pointing an agent at a repo, or it may wander off and "fix" things you never asked for.
A single-3090 coding agent, fully boxed in
A detailed field guide doing the rounds runs a private coding agent on one used RTX 3090: llama.cpp in router mode with draft decoding for speed, the Pi harness sealed in Docker with no credentials and a single project folder, and a remote web UI to steer it unattended. It drives local Qwen3.8-27B and Gemma-4-31B fine-tunes — a realistic weekend build if you want an offline agent, plus the local-vs-cloud cost math to decide if it's worth it.