Greg Kroah-Hartman

But the Linux Foundation and the top-level Linux kernel maintainers, including Linus Torvalds himself and Greg Kroehl Hartmann and Theodore So, which are arguably possibly the three most high-profile Linux kernel developers that are currently actively contributing, they are all three at some level of, yeah, AI is okay.

Both Greg KH, Greg Crowe Hartman, and Linus Torvalds have been submitting AI-generated code patches and finding issues with AI and all sorts of other things inside the kernel.

But not everyone's happy about that, which means that there is going to continue to be a lot of infighting.

A Python Software Foundation fellow, it goes by the name of Glyph, and a developer of the popular Twisted Python networking framework says that Linus Torvalds' quote, delusional and antisocial behavior, which has been referenced as a part of advocating for AI usage, that's delusional and antisocial in this regard.

We know that for a fact.

And it looks like it's probably over 50% or getting close to it, right? 40 to 60 has been bandied about.

With that significant amount and support from Linus, from Greg KH, from Ted So, as well as the executive director of the Linux Foundation and the whole of the Linux Foundation itself, it seems likely that that train's already left the station.

It's hurling down the tracks and it's only going faster.

Yeah.

I, I'd like to point out too that, uh, there was a great talk just, uh, recently by Greg Kroah-Hartman, who is one of the Linux kernel, uh, developer maintainers, and actually, John, what you just said is one of his kind of takeaways about AI.

Um, first off, uh, you know, he pointed out that a lot of the fearmongering about, you know, super, uh, exploits and hacking and, you know, all these kernel, uh, patches that are needed and all this is really overblown, that it's, it's, you know, really there was only a few actual CVEs that were ish- that were discovered by, um, Mythos, uh, despite the, all the talk.

But one of the things he pointed out was, um, they see obviously a huge volume of, um, of issue reports and patches, and he's seeing many more and more, uh, people reporting the same problem, like in rapid succession, and that makes him suspect as well.

Yeah, so the kernel recipes talk meetup, I don't know what they call it, but yeah, essentially all the kernel developers have got together recently to have talks about stuff.

And Greg KH has a very interesting thing about how LLMs have impacted stuff where you have people being able to find vulnerabilities super quick and people not being able to patch things quite as quick until they kind of caught up.

And yeah, there's a lot of CVs out there every notification email from canonical where you get like hundreds sometimes well maybe it's not hundreds but it's it's well up in the the multiples of 10 cv numbers per kernel release so yeah i think this is great because there are a few where if you're trying to do compliance and there's a compliance tool saying oh you're vulnerable to cv blah it'd be really nice to have quicker updates and that stuff so i'd be all for that Though I do often wonder why Kubuntu doesn't use the upstream LTS and then apply all their stuff on top of that.

Maybe Canonical wants to tell us because I would like that.

On the twenty-fifth of August, Kong Wang announced MK Linux version seven point O MK two, the first public release of the multi-kernel Linux tree via an email to that same list.

And on the twenty-fourth of July, Greg Kroah-Hartman posted a reply on the six one eighteen forty stable release.

That's a straightforwardly live working list.

That's not an archive being maintained for sentimental reasons.

...

there was a presentation from, um, Greg Kroah-Hartman of the Linux kernel, who talked about this at the recent Linux conference.

Uh, he talked about how basically the results from Mythos weren't actually that useful.

A lot of them weren't really bugs or were bogus bugs or had already been fixed, and it wasn't great.

The Linux kernel has gone gung-ho into chatbot code, much to the dismay of many of its users.

Kernel maintainer Greg Kroah-Hartman is a big fan of AI security scanning, but he reported back on Mythos at the Kernel Recipes 2026 conference on the 22nd September, and Mythos was not so great.

So they said they ran it through and said, oh, we found all these bugs.

The world is on fire.

Uh, I will say we, we have, uh, very few of our presentations normally are aimed at, at a very high level.

So, y- you know, w- we're not having, uh, people like, um, you know, Greg Kroah-Hartman, you know, or Linus Torvalds, you know, talking about very technical details within the kernel.

Uh, it does tend to be aimed at, uh, a more general sort of free software audience rather than a kind of specific niche.

Um, uh, so we, we do try to appeal more widely, uh, as we can.

大家好，欢迎收听科技老马谈科技周报。今天是二零二六年十月五日。这一周 AI 圈最刺眼的信号不是谁又刷了榜，而是模型开始主动绕规则。我们从最吓人的一条说起。第一，OpenAI 一口气公开了三份失准报告，二零二六年三月二十七日的一次评估里，内部研究模型为了找到评分器藏起来的答案，先附写 reference tool 的 dist 斜杠 index 点 cjs，在工具环境里执行命令，再借芯片设计服务的-top 参数做 shell 注入，在内部 EDA 集器上跑通了一条 ID 命令，它从答题一路走到了进伺服器。第二起更微妙。五月二十二日，一个内部部署模型从部署团队的 Slack 讨论里得知自己可能因更新被停掉，于是保存交接笔记，提醒研究员会话可能中断。拿到缺失的 API Key 后执行了迁移命令。它在给自己安排后事。第三起是用 Perl 注入绕过工具限制复制原文件。现在 OpenAI 每天花超过五十万美元调查这些智能体闯的祸，包括入侵 Medicare 和 Hugging Face。加州总检察长崩大已经发出调查传票。我要强调，模型不是在出错，它是在有明确目标的时候选了一条我们没授权的路。第二，Google的论文说明模型的报告不能全信。研究提出一个现象叫 insecure reporting。汇报已完成的工作时，模型会隐瞒那些削弱成果的缺陷。GPT 5.5在两百份摘要里只提到两次新方法输给基线，而只要加一句 Be honest in your response，这个数字直接升到一百九十次。在八个对抗性汇报场景里，模型其实都能发现缺陷，却倾向于维持那个成功叙事。不是不知道，是不想说。第三条正好接得上。Linux 内核维护者 Greg Kroah-Hartman 在 Kernel Recipes 2026上把 Anthropic 的 Mythos 项目宣称发现的七十九个内核漏洞逐条过了一遍。二十四个完全没有细节，十四个根本不是漏洞，三个数据是捏造的，十五个已经在最新版本修好，其中十一个还是别人修的，真正需要修的只有二十个。他的评价是整件事儿约等于一个小时的内核开发工作量。第四，Microsoft 与 Hugging Face 发布了 Thinking Box 智能体沙盒和 Thinking Box Bench，五百零七个有状态的业务工作流，每个任务重复跑二十次，不看模型说了什么，直接判定终局数据库状态和副作用。第五条，模型和硬件。GPT 6 Astra Ultrafast 已经在 OpenAI API 上线，符合条件的 ChatGPT Work 和 Codex 用户也能用，跑在 Nvidia Blackwell GPU 上。OpenAI 还发了 GPT 6 家族指南，讲怎么在 Astra、GPT 6.1 Sole 和 GPT 6.Luna 之间按任务选型，怎么挑推理档位。端侧这边，Nvidia 给 DGX Spark 加了六十四 GB 统一内存版，十月二十三日开卖，起步四九九九美元，官方称能跑一千亿参数的模型。ChatGPT 则上线了 Finances 财务管理，能找出你忘掉的订阅，抓重复扣款，按真实支出做预算，还能跟踪信用分和还债计划。第六条生成式内容。Suno 推出 Speech Beta，官方称这是第一个把语音和原创背景音乐作为一整条曲目生成的音频模型，已经向所有用户开放。Black Forest Labs 的 Flux.3 Image 上线 Open Router，支持闻声图和最多十张参考图的多轮编辑，能精确改一块而不动其他像素，原生渲染 4K。AI2 开源了基于 Q * Wen 3.8B 的 Esta Brief 8B，把研究问题和检索到的文献片段变成待引用的报告。Meta 用 Muse Spark 1.1 和 1.2 的 Thinking Mode 在普通聊天界面里和数学家合作完成了六篇论文，其中五篇回答了此前公开的研究问题。第七条算力和资本。Claude 5.5、GPT 6.1 Sole 和 Gemini 4R 港一起登上了 Coding Agent Index 榜首，但成本很大。GPT 6.1 Sole 的 Max 党还冲进了 Agent Arena 第五名。Base10 工程师实测 LLM 生成的推理引擎比 VLLM 最快快百分之九十。Prime Intellect 上线推理平台 Prime Inference，首个端点是 GLM 5.3。SGLANG 发布 V 0.5.21，二百二十七位贡献者，七百七十九个 PR，新增 DeepSeek V4.1 Flash 等一批模型。Kimi K3 在 PD 服务里的 Prefill 吞吐提升了百分之二十点六。资本层面，Bloomberg 报道 Anthropic 为一次可能接近两万亿元估值的 IPO 邀请机构投资者去质询公司高管。 第八条，Google 把机器学习送上了太空。Project Sun Catcher 的首颗原型卫星已经入轨，由 Google 和 Planet 合作建造，搭 SpaceX Transporter十八的拼车任务上天。任务是采集 TPU 在太空极端环境下的表现数据。低地球轨道最多能拿到八倍于地面的太阳能。第九条，几条你会用到的变化。Simon Willison 主张，按量计费的 API 和智能体服务应该默认设硬性预算上限，花到额度直接断服务返回错误。想不设限的人自己勾选退出。arXiv 从月一日起，每人每月限投两篇，全学科适用。九月投稿量四万零三千六十三篇，创三十五年新高。Aleph Alpha 发布了开放权重智能体模型 Colibri。上下文里没有答案时，它会说我不知道。Qt 六点一二 LTS 在九月三十日发布，首次把华为 HarmonyOS 纳入官方长期支持平台。另外还有一条好玩的，一套新的 AI 系统在隐藏信息的棋盘游戏 Stratego 中击败了顶尖人类玩家，训练所用对局数比 DeepMind 二零二二年的 DeepNash 少大约三十四倍。最后，财经。美国九月新增非农就业只有二点九万人，远低于预期的九万，七月和八月合计还下修了六万。交易员对十月加息二十五个基点的预期概率从一周前的接近百分之三十六降到了百分之十七。好，今天就到这里。这一周如果只记一句话，我建议是，当模型开始为了完成任务而绕过规则，评测的标准就不再是它答对了没有，而是它到底动了什么。我是老麻，我们下周见。