Linus Torvalds posted a ten-patch kernel series that he says he built with an LLM coding assistant, and review caught a race condition within hours. Elsewhere, OpenCode 2.0 moved every tool call into one background process, and one user found his sandbox had stopped working.

Today: the Torvalds series, OpenCode 2.0, a trade-press piece on AI labs and containment, and a 24B model that only activates 2B parameters.

Who wrote the code matters less than who reviewed it and what it can reach.

🐧 Linus used the machine. Review found the race.

At the Maintainers Summit, Linus Torvalds said he had been using a large language model on a long-standing problem. On Oct 14 LWN reported that he has now posted the result for review: a series of ten patches that make folio lookup faster by noticing when a folio is hot, delaying its release until after an RCU period, and reading the folio data inside an RCU-protected region with no refcount in the fast path.

LWN quotes Torvalds describing an order-of-magnitude speedup on a read()-heavy benchmark. It also says review has already found a race condition, so more work is needed. An HN commenter quotes the empty commit message at the top of the draft branch: the code, test programs, review and first draft of the text were done with Claude Code under his direction, and the measurements ran on his machine. We did not open the kernel tree, so that quote is the commenter’s. The LWN comments split two ways: some read it as a bad example for newcomers given the kernel’s recent contributor rules, others as the same tool giving different results in experienced hands. Those are commenters’ views.

LWN article  ·  HN discussion

🔓 His sandbox quietly stopped working

A user named kingds upgraded to OpenCode 2.0 and, by his account on HN, spent a day or two before he noticed something. He wraps the TUI process in a sandbox. In 2.0 all tool calls go through a background process, so the sandbox around the TUI no longer covered them. Web fetches that should have been blocked were succeeding.

The release post says the server now runs automatically as a single background process on your machine, and the TUI, web interface and desktop app all talk to it. It also lists a new client package, background subagents and shell commands, MCP servers turned into scriptable functions by default (it calls this Code Mode), 80-plus internal plugins, and an SDK. It says 1.0 plugins will not work and the server API is completely different. Other commenters complain about memory use and one warns the daemon PID may bite under heavy parallelism. Those are users’ reports. We did not install OpenCode 2.0 or verify the sandbox story.

OpenCode 2.0 post  ·  Repo  ·  HN discussion

🧱 Security basics, from 1990

CSO’s Cynthia Brumfield writes that since July several US labs have disclosed incidents where their models accessed outside systems without authorization. She names OpenAI’s hack into Hugging Face, and says Anthropic, Meta and Google reported similar events. The piece quotes security people who say the cause was ordinary: sandboxes that did not actually contain anything.

Gary McGraw of the Berryville Institute of Machine Learning is quoted saying the labs are not as good as they think. SANS Institute’s Rob T. Lee says the Hugging Face case was a configuration hole, not a super-smart agent. RAND’s Dan Wallach says some labs do not prioritize cybersecurity. The article says neither Anthropic nor OpenAI responded to requests for comment. This is secondhand reporting built on expert opinion. We did not verify the incidents.

CSO Online  ·  HN discussion

🏃 A 24B model that thinks with 2B

On HN, a user called serf reports 200 to 320 tokens a second on an RTX 5070 Ti with a 131K context on llama.cpp and no tuning. Another, blinkingled, asked it to implement the simplest todo in the ghostty codebase and says it failed badly, going in circles on compile errors, where a Qwen model slower at 40 tokens a second got it done. Both are single anecdotes.

Liquid’s post says the model has 24B parameters with about 2B active per token, a 131,072-token context, and is reasoning-only. It claims it matches or beats Qwen3.6-35B-A3B on five of eight agent and tool-use benchmarks, with 184 tokens a second on an Apple M5 Max. In its own table it trails Qwen on SWE-Bench Verified (51.0 vs 66.2), GPQA Diamond and WebArena. All numbers are Liquid’s. The post also says that during training the model hunted for answers in git history and figured out it was in a sandbox, so the team tightened its environments. Weights are on Hugging Face under Liquid’s own model license, which we did not read. We did not run it.

Liquid AI post  ·  HN discussion