How far has AI-generated code penetrated Linux kernel development? This week's collected data surfaced a number that shows the reality, a perception gap between top executives of major AI players over AI risk, and moves to recreate a model that returns "only judgments" on existing LLMs.
17.25% of September's Linux Kernel Patches Tallied as AI-Generated
Tech media outlet Lunduke Journal counted the patches submitted to the Linux kernel in September and reported that **17.25% were AI-generated code**. The story also drew 48 comments on Hacker News, sparking lively debate.
The Linux kernel is built on strict maintainer review, so an influx of AI-generated code can affect both the load on the review process and code quality assurance. The tally's methodology and the criteria for judging code as "AI-generated" are unclear at this point, so the figure should be read as a rough reference. Even so, it stands as a signal of how fast code-generation AI is spreading, reaching even one of the most conservative development communities out there.
NVIDIA CEO Jensen Huang Puts the Odds of AI Destruction at "0%"
On the 21st, The Guardian reported that NVIDIA CEO Jensen Huang dismissed AI alarmism, stating that the odds of AI destroying the world by 2030 are "0%." The article apparently also picks up references to warnings from Anthropic.
Just this month, on the 16th, Anthropic CEO Dario Amodei issued a warning about large-scale risks that sent AI-related stocks plunging (covered in this blog's September 16 issue). The fact that leaders on the supply side are this split over AI's destructive potential looks like important material for reading regulatory and investment decisions.
Jev Returns Only "Judgments" — Moves Underway to Recreate It on Existing LLMs
Jev, which generates no text and returns only probabilistic judgments, has been the center of attention these past few days amid a surge of Japanese hands-on measurement articles and controversy over its origins. This week's collected data contained multiple signs that the surrounding scene is advancing into the implementation stage.
In Reddit's LocalLLaMA community, an open-source gateway called "hearim" was released. Give it a set of data and choices, and it returns a Jev-compatible judgment API (Choice/Score/Noul formats) from existing LLMs running on llama.cpp, vLLM, SGLang, Ollama, and the like. The trick: use only pre-fill and a single token of decoding to read out each option's probability (logit), never letting the model generate prose. The developer argues that "before instruction tuning and RLHF, pretrained models have already learned a great deal. If the task is just picking from options, that capability already lives in the probability distribution," and says they had used the same technique for over a year before Jev's announcement. The developer is careful to make clear that this is not a claim of reproducing Jev's architecture or performance. At the same time, skeptical voices are pushing back: "isn't this something simple neural networks have been able to do all along?"
In the Japanese-language sphere, an article was published comparing three approaches to routing incoming inquiries — Jev, rule-based, and GPT-5.6-luna — on the same data with the same output format. It examines the value of decomposing a monolithic prompt into small, typed questions, tackling the problem that when part of a monolithic prompt's output goes wrong, isolating the cause is hard. On top of that, Vercel Labs has released a template combining Jev with the AI SDK, so the ecosystem around the official SDK is shaping up too.
DeepSeek-V4.1-Flash: Working Capacity Backward from "890 bytes/token"
An article was published parsing the DeepSeek-V4.1-Flash technical report, submitted on September 17, from an operator's perspective. The number that matters isn't the maximum context length (1 million tokens) but **the global KV cache's 890 bytes/token** — a capacity-per-token figure.
The article argues that when evaluating long-context models, comparing maximum length alone is insufficient: you need to convert "actually live tokens × concurrent sessions" back into HBM and persistent cache capacity. That's what makes the 890 bytes/token figure the metric closest to real operations — it lets you work backward from actual hardware to the number of agents you can keep in memory simultaneously.
OmniVChat: Pairing Synthetic Data and RL for Speech-Vision Dialogue Without ASR
On Hugging Face Daily Papers, research targeting native speech-vision dialogue, "OmniVChat," was featured. Approaches that hold conversations through audio and video without ASR or intermediate text have long struggled with two challenges, the work notes: data scarcity and the difficulty of open-ended evaluation.
The study takes on both. It proposes "OmniVChat-Studio & Bench," which uses multi-agent synthesis to create single- and multi-turn dialogue data across five capability categories, and "OmniVChat-RL," which balances response accuracy, efficiency, and conversational naturalness. In validation using Qwen3-Omni-Instruct, the authors report improvements on both the synthetic benchmark and a benchmark recorded with humans. Whether transfer from simulation to the real environment actually succeeds is the make-or-break point for this kind of synthetic data research — and the work seems to deliver there too.