Friday, July 17, 2026
Torvalds declares Linux "not one of those anti-AI projects" and tells objectors to fork it; Thinking Machines ships its first from-scratch open-weights model; and the EU orders Android opened to rival AI assistants.
Linux is not an anti-AI project, Torvalds declares
Linus Torvalds used a July 14 reply on the linux-media mailing list to declare that “Linux is not one of those anti-AI projects,” a line he said he would enforce as top-level maintainer, telling objectors to “fork it” or “just walk away.” He would “very loudly ignore” anyone arguing to stop others from using LLM tools, he wrote, while stressing that no one is forced to use them. The trigger was Sashiko, a proposed tool to wire LLM-based patch review into the kernel’s Patchwork system; developer Laurent Pinchart had proposed triaging its output before comments reach patch authors, citing Software Freedom Conservancy guidance, which co-creator Roman Gushchin and Torvalds cast as anti-LLM. Torvalds conceded LLMs cause “pain” through maintainer workload and by “finding embarrassing bugs.” Tom’s Hardware reads the stance as a reversal of his 2024 dismissal of AI as overhyped, and notes Sashiko’s project page claims a self-reported 53.6% bug-detection rate on already-reviewed patches. This was a mailing-list post, not written policy.
A new open-weights base: Inkling
Thinking Machines released Inkling, its first from-scratch open-weights model: a 975B-parameter Mixture-of-Experts transformer with 41B active, a 1M-token context, and pretraining on 45T tokens of text, image, audio, and video. The architecture breaks from convention with learned relative positional embeddings instead of RoPE, a 5:1 sliding-window-to-global attention ratio, and encoder-free multimodality (dMel spectrograms for audio, 40x40-pixel patches for vision). Its distinguishing feature is controllable thinking effort, trained by varying per-token cost across 30M-plus RL rollouts; the company claims roughly 3x the token efficiency of Nemotron 3 Ultra at equal Terminal Bench 2.1 score. Thinking Machines calls Inkling “not the strongest overall model available today, open or closed,” pitching it as a customization base on Tinker. All benchmarks are vendor-reported.
Regulation and a live Windows zero-day
The European Commission issued two Digital Markets Act decisions ordering Google to open Android and Search to rivals. The Android decision requires competing AI assistants (the Verge names ChatGPT, Claude, and Perplexity) to get system access “comparable” to Gemini’s: responding to “Hey Google,” interacting with apps, and reaching device hardware, with users, not Google, controlling access. The Search decision forces Google to share Search data with rival engines and with AI chatbots, which the EU treats as functionally equivalent. Google must comply by January 2027 for Search and July 2027 for Android; non-compliance can draw fines up to 10% of worldwide turnover. Google’s Kent Walker said the rulings undermine privacy and security for European users.
Separately, pseudonymous researcher NightmareEclypse published HiveLegacy, a working Windows elevation-of-privilege exploit, the same day Microsoft shipped what Ars Technica calls a record Patch Tuesday. The flaw is in the Windows User Profile Service and lets a low-privilege user modify an administrator account’s classes registry hive, given the credentials of one other account and the username of a third on the machine. Multiple researchers confirm it works; no patch exists yet.
Research and systems
A preprint, not yet peer-reviewed, tests whether hidden-state probes can catch alignment faking, where a model appears compliant under monitoring but not otherwise. Across 13 models, natural faking emerged in only two, and detection proved model-conditional: leakage-free probing worked on Llama-3.1-8B (AUROC 0.87) but collapsed to chance on Qwen3-32B (0.43), with no cross-model transfer. The authors also show common probe evaluations are badly inflated, naive linear probes reaching a meaningless AUROC of 1.0 and residualized probing hitting 0.63 on a control where no faking can occur, and release a five-control framework to prevent the artifacts.
Richard Feldman’s team rewrote Roc’s ~300K-line compiler from Rust to Zig over 487 days and found near-identical memory-corruption rates across both eras (21 of 2,596 bug reports under Rust, 10 of 431 under Zig, classified by Claude Opus 4.8), concluding the choice made no appreciable safety difference for a compiler already carrying ~1,200 unsafe blocks. And Tower Semiconductor committed up to $3 billion, backed by Japan’s METI, to revive a shuttered Panasonic-era fab for silicon photonics, disclosing $1.3 billion in contracted 2027 revenue as AI interconnects move from copper to optical.
What to watch today
- Whether Microsoft ships an out-of-band fix for HiveLegacy, unpatched after the July Patch Tuesday.
- Independent replication of Inkling’s vendor-reported token-efficiency and benchmark claims, plus the promised 276B/12B Inkling-Small preview.
- Google’s DMA clock: Search-data sharing due January 2027, Android AI-assistant interoperability due July 2027.
- Whether the linux-media thread settles how Sashiko connects to Patchwork now that Torvalds has ruled out an anti-LLM veto.