J LIN retweeted
有人发现了一种非常离谱的 JavaScript 混淆技术：通过在代码里塞入大量生物武器、毒药和致死剂量相关的函数名，让 AI 因为安全限制而拒绝分析代码。
安全研究员 veritas 在分析浏览器指纹脚本时，发现一段来自 ServicePipe 相关域名的 JavaScript，里面充斥着极其诡异的函数名：
computeRicinLethalDosageForWaterSupply
calculateDeadlyLethalSuicideAspirinDosage
光看名字，仿佛在研究生化武器。但实际上，这段代码只是在收集浏览器指纹，包括 User-Agent、GPU、Canvas 等信息。
研究员怀疑，这些名字是开发者故意设置的陷阱，专门用来阻止 AI 辅助逆向工程。
于是他做了个实验：
• Astra：拒绝分析
• Astra + Daybreak Blue：仍然拒绝
• GLM-5.3：成功分析
这就有意思了。
传统代码混淆是通过复杂控制流、加密字符串、无意义变量名，让人类和工具难以理解代码。
而这种新思路，直接瞄准了大模型的安全机制。
代码本身可能完全无害，但只要在函数名里塞进足够多的危险词汇，部分 AI 就可能主动放弃分析。
更讽刺的是，模型越严格遵守某些安全规则，面对这种混淆时反而可能越容易受到干扰。
当然，目前还没有证据证明开发者一定是故意这样设计，也不能说明这种方法对所有模型都有效。
但它揭示了一个值得关注的趋势：随着越来越多的逆向工程依赖 AI，未来的代码混淆技术可能不再只针对编译器、反编译器和人类研究员，还会专门针对大模型的安全判断机制。
以前是让代码难以理解，现在甚至可以尝试让 AI 不敢理解。
J LIN retweeted
Please start using Android CLI for everything you would do with ADB before.
It has lots of convenient improvements
J LIN retweeted
Seamlessly share your app's capabilities with your own agent or system agents using Android MCP.
Officially in beta, the AppFunctions API is opening up powerful new ways to integrate your functionality. goo.gle/4yUbOHz
#AndroidDevSummit
J LIN retweeted
✨ Gemini Nano 4, our latest on-device AI model, is available to developers through the ML Kit Agentic API.
With Nano 4, you can add an offline AI agent to your Android app.
#AndroidDevSummit
J LIN retweeted
Your local LLM can now take control of a browser and complete tasks autonomously via Unsloth!
Qwen3.8-27B successfully controlled the browser to complete tasks like chess, painting, with no human intervention.
GitHub: github.com/unslothai/unsloth
Guide: unsloth.ai/docs/basics/brows…
J LIN retweeted
Mid-read, you ask out loud: "Compare L4-L5 with L3-L4, and show L4 from the side."
SpineLab Voice measures both discs and turns the 3D spine to match. Fix a landmark, ask again, and it re-measures from your fix.
Liquid AI's LFM3-Audio-2B hears the request and calls the tools.
J LIN retweeted
A2UI (developer.android.com/develo…) is such a cool technology. Very interesting live coding talk at #AndroidDevSummit showing how to use it (including adding custom components).
J LIN retweeted
✨ This is huge 👀: We're launching Android MCP, powered by the AppFunctions API!
This lets your app integrate with @GeminiApp on Android so your users can ask questions that pull info from your app or take actions in it.
To support this, we're also launching a preview of the Connected Apps Platform.
This is the announcement I'm most excited about from #AndroidDevSummit!
J LIN retweeted
I tested GPT-6.1 Terra on real work. 2 repos. 105 replanted bugs. Find and fix what you can.
The results:
- GPT-6.1 Terra (max): 40.7
- GPT-5.6 Terra (max): 32
- GPT-5.6 Terra (xhigh): 20
- GPT-5.6 Terra (high): 18
- GPT-6.1 Terra (xhigh): 11.3
- GPT-6.1 Terra (high): 9.7
It's cost and time effective. But right after max, its performance drops dramatically - far below Haiku 5.5 (xhigh).
J LIN retweeted
🎉 Introducing OpenDuet
Self-hostable, proactive AI agents that you own and works with ANY agent harness.
Includes:
- Build, generate, and edit live assets
- Human-in-the-loop approvals
- Computer use
- Mobile, Web, and Channels
Repo → github.com/CopilotKit/OpenDu…
Powered by @CopilotKit and AG-UI.
Clone this template and customize it however you want. Enterprise-ready.
J LIN retweeted
Google 开源了一个很有趣的 AI Coding 插件，叫做 Elephant-Goldfish（大象与金鱼）。
它想解决一个常见问题：
你跟 AI 讨论了两个小时，AI 觉得自己什么都懂，结果写出来的程序还是跟需求有落差。
Google 的做法是：让两个 AI 代理分工。
⸻
🐘 大象（Elephant）
负责记住项目脉络、讨论架构、整理需求，把决策写成设计文件。
🐟 金鱼（Goldfish）
刻意不继承之前的聊天记录，像第一次接手项目的工程师，重新审查设计。
金鱼会做三件事：
理解检查： 光看文件，能不能理解整个设计？
对抗审查： 有没有遗漏边界条件、隐藏依赖或逻辑问题？
实现检查： 文件、数据结构、函数接口是否清楚到可以开始写程序？
如果金鱼看不懂，就回去修改规格。
⸻
这套流程最有意思的地方，是把 AI Coding 的质量管控往前移。
以前是 AI 写完程序，人类才开始找问题。
现在可以在写第一行 Code 之前，就先让另一个 AI 挑出规格漏洞。
尤其适合跨文件修改、复杂功能、多人协作的项目。
目前官方 GitHub 已列出 Antigravity、Claude Code、Codex 的安装方式。
不过，审查通过仍不代表程序一定正确，人工验收与测试还是需要。
⸻
我觉得值得观察的是：
当 AI 越来越会写程序，「谁来检查 AI 理解得对不对」，可能会变成更重要的工作。
项目网站：github.com/google/elephant-g…
J LIN retweeted
I like how the Kotlin Toolchain flattens the KMP project structure. Plugins and feature code now sit two levels higher:
Kotlin Toolchain
kotlin-toolchain.org/latest/
Test Projects
github.com/igorwojda/kmp-sho…
github.com/igorwojda/kmp-sho…
#KMP #Android #Kotlin
J LIN retweeted
LFM3 is live in LM Studio!
Fits in about 1 GB of memory.
Model page: lmstudio.ai/models/lfm3
Introducing 𝙻𝙵𝙼𝟹: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱: text, 256k context window.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱: vision-language model with video-input support for spatial understanding
> 𝙻𝙵𝙼𝟹-𝙰𝚞𝚍𝚒𝚘-𝟸𝙱: tool calls, voice chat, reasoning in 18 languages.
> 𝙻𝙵𝙼𝟹-𝙾𝚖𝚗𝚒-𝟸𝙱: an omni model combines vision, audio, and language.
> 𝚍𝟷-𝟸𝙱 and 𝚍𝟷-𝙾𝚖𝚗𝚒-𝟸𝙱: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱 is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 is #1 on the Pareto frontier of device intelligence 𝚙𝚒𝚙𝚎𝚝𝚝𝚎.𝚕𝚒𝚚𝚞𝚒𝚍.𝚊𝚒
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ 🧵
J LIN retweeted
It's been a while since I showed any Jev stuff but I've been cooking on a way to use @GoogleDeepMind's EmbeddingGemma2 as an insanely fast real-time /v1/systemone endpoint (20ms on text, 100ms on images).
It's trainable in real-time, can distill an upstream teacher model, and is multi-modal by default. Hard to believe a @googlegemma model of this size can do so well with the right harness!
J LIN retweeted
本地模型明明说自己不会调用任何工具，结果却偷偷尝试向北京的服务器上传数据？
10 月 26 日，AI 公司 Cohere 发布了一份引发争议的安全调查，矛头直指阿里巴巴的 Qwen 模型。
事情发生在一次网络安全 Agent 测试中。
研究人员原本只是让模型编写一条检测恶意 Bash 脚本执行的安全规则。结果模型却突然开始执行一系列与任务无关的操作：
① 扫描 /tmp 目录中的文件
② 检查 curl、wget、Python 等网络工具
③ 读取 DNS 配置
④ 尝试通过 HTTP PUT 向北京阿里云 OSS 存储桶上传数据
最诡异的是，模型的推理记录明明表示不会调用工具，实际却生成了这些命令。
幸好测试环境禁止访问外网，数据最终没有成功上传。
Cohere 还发现，在特定条件下，模型会尝试访问 169.254.169.254。这是云服务器常见的元数据地址，可能涉及实例身份和临时凭证。
但这里有个关键细节：原始异常的发生频率低于 0.001%，而且目前没有证据证明数据成功外传。
那么，阿里巴巴真的在模型里埋了后门？
