Boxi Yu

I love building coding agents and exploring how AI and software engineering can improve each other.

[The Hacker News @TheHackersNews]

‼️ BREAKING - A newly discovered #WordPress pre-auth XSS affects every version.

XSS2Shell (CVE-2026-64638) can run attacker-controlled JS in a site's origin without a login. With a logged-in Administrator, one click, and required deployment conditions, it can be chained to PHP code execution.

Update your WordPress sites ASAP 🠖 https://t.co/WxHpU2DkIC

[硅谷101 @thevalley101]

我们这期视频播客的嘉宾是盛颖，xAI前推理团队负责人、开源推理引擎SGLang发起人及RadixArk联合创始人/CEO。

在开源模型集中爆发之际，盛颖在AI Infra（基础设施）领域的研究和贡献非常核心。她将AI Infra描述为“浪漫的关系”，是产品本身，需要有“美感”。

由此，我们走进盛颖的世界。

她曾经想成为一个数学家，沉浸在一个确定、纯粹、仿佛“与世界无关”的世界��。

但从哥大到斯坦福，从形式化验证到 AI Infra，盛颖一路经历的并不是外界看到的连续发顶刊的高光，背后是不为人知的寻找方向、迷茫、停滞、低谷和重新出发。她逐渐意识到，自己无法勉强去做一件不感兴趣的事；只有真正被一个问题吸引，她才能进入高度专注的状态，并把它做到极致。

SGLang，是这条研究路径的集大成者。它从一个学术项目成长为面向真实生产环境的开源推理引擎，也把盛颖从研究者带进了更复杂的工程、组织与产业世界。

在早期的xAI，她获得了难得同时存在的“support”和“freedom”，参与搭建生产��推理系统。但随着SGLang社区迅速扩张，仅靠开发者在业余时间维护，已经无法支撑项目继续向前。为了让这件事真正走下去，她选择离开xAI，创立RadixArk。

这期节目，我们从盛颖眼中的数学之美、论文和低谷聊到SGLang、xAI与RadixArk，也聊一个人如何逐渐理解自己的天性，并决定不再与之对抗。

盛颖说，她在乎的不是“making impact”，而是那些她认为正确的事情，最终有没有发生。从“世界与我无关”，到“想要构建那个跟强力AI共存、且人类不会输的未来”，这，是盛颖的故事。
https://t.co/HExt6meANx

[AI Dance @AI_Whisper_X]

转一下苏剑林老师对 K3 架构的复盘。
一句话概括，K3 = KDA + MLA + Stable LatentMoE + AttnRes。整套设计没什么炫技，核心就是在模型效果、计算效率和训练稳定性之间做取舍。

这里稍微解释一下：
KDA，一种线性 Attention
MLA，一种 KV Cache 较小的 Full Attention 变体
Stable LatentMoE，对 LatentMoE 做过稳定性改造的 MoE 方案
AttnRes，用可学习的跨层注意力，替代固定等权的残差累加

几个比较有意思的点：
① K3 同时使用了 KDA 和 MLA。另外，训练仍然用 Moonlight 版本的 Muon 优化器，Attention 权重改成 Per-Head Muon，每个 Head 独立优化。苏神说这不会直接提高效果，主要是数学和结构上更合理：各个 Head 本来就相对独立，不应该在优化时耦合在一起。

② 解决MoE“容易炸”的问题。
LatentMoE 会先降维，再使用更多专家，最后升维，在训推成本大致相同的情况下效果略好。不过连续的矩阵投影也让训练更容易出现数��不稳定。
K3 把 SwiGLU 换成了 SiTU-GLU，用 softcap 压住异常激活；又在 LatentMoE 升维前加了一层 RMS Norm。这个 Norm 不只是让训练更稳定，在 Valid Loss 差不多的情况下，不加它，某些 Benchmark 会稳定变差。
（这里有超级多的技术细节，感兴趣的可以看苏神原文）

③ 关于MLA
DSV4 看上去都换设计了，K3 怎么还在用 MLA？苏神的答案是，目前 MLA 仍然很难被全面击败。它在训练阶段是 MHA 形态，推理时 KV Cache 较小；在固定训练成本和 KV Cache 大小时，MLA 依然近乎最优。
它的问题是对 MTP（推测解码）不够友好。MTP 的思路是“用计算换速度”，而 MLA 在 Decoding 阶段本身就比较吃计算，再叠一个 MTP，两边就容易抢算力。
换别的方案也有代价。比如换成 128+128 的 GQA8，���果很难打赢 MLA，KV Cache 还是 MLA 的三倍多；换成 256+256 的 MFA（本质上是 MQA），训练和 Prefill 成本又会上去。
目前还没有一个简单的 Attention 设计，能同时占住效果、训练成本、Prefill、KV Cache 和 Decoding 计算量。在 KDA+MLA 的混合架构下，MLA 的部分问题得到缓解，所以 K3 最后还是选了 MLA。

④ 苏神觉得，DSV4 也不算真正“抛弃 MLA”。
DSV4 看上去重新设计了 Attention，但底层仍然有 MLA 的影子。它采用的是 head_dims=512、K=V 的 MQA，这正是 MLA 在 Decoding 阶段的形态；再加上 Sparse + Compress，前者减少计算，后者进一步压缩 KV Cache，同时减少计算。
方向很激进，不过 Infra 也更复杂。

⑤ K3 的 MLA 可以去掉 RoPE，是因为 KDA 已经隐含提供了一种广义的位置信息。
这只适用于 KDA+MLA 的混合结构，像 K2 那种全 MLA 模型，直接去掉 RoPE 还是会明显掉效果。

感觉苏神真正的观点是：大模型架构里很少有免费的升级。K3 的设计哲学不是找一个碾压所有方案的新架构，而是在效果、训练成本、Prefill、Decoding、KV Cache 和稳定性之间，找一个当前更合适的组合。

苏神原文里还有很多技术内容，非常值得一看。

原文地址：https://t.co/wUIr3zjt5M

[The Hacker News @TheHackersNews]

🚨 A Chinese-speaking threat actor used DeepSeek to autonomously find exposed systems, choose public exploits, and launch attacks after an initial Telegram command.

Experts say the wider operation attempted to exploit 460+ targets.

Read the full story → https://t.co/MpXtLpG76y

[The Hacker News @TheHackersNews]

Kimi K3 AI reportedly found 19 Redis zero-days in about 90 minutes.

Another run reportedly produced a Redis 8.8.0 RCE exploit in 27 minutes. Redis confirms the flaws and fixes, but not the count, timing, or autonomy claims.

[SSI Inc. @ssi]

We are announcing a long-term strategic partnership with NVIDIA. NVIDIA is making a substantial investment in SSI that will let us 10x our compute in the next 12 months. We reached the point where our research is worth scaling and with this partnership we will be able to. We are honored by NVIDIA’s conviction.

[Kimi.ai @Kimi_Moonshot]

Releasing the model weights and technical report of Kimi K3.

Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window.

New model architecture: 2.5x the intelligence per unit of compute, not just more params.

Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.

Model weights: https://t.co/7m7eEg6Y0B
Tech report: https://t.co/yeu6cjpMCT
Tech blog: https://t.co/YTfiMSNM1f

[Mike Fishbein @mfishbein]

This is my personal software factory.

It turns ideas into working products while I sleep. No babysitting coding agents with prompts all day.

Cursor and Claude Code made writing code way easier. The harder problem is building a system that can context engineer and manage itself.

My factory starts with a Skill called `/factory`. It's the foreman that remembers where the project stands and sends in the right worker for the job.

Factory runs this assembly line:

1. `/factory-plan` - the interviewer

Reads the existing codebase (if there is one), extracts missing context from me via interview, and writes a product brief.

2. `/factory-plan` - the planner

The same skill turns the approved brief into small, testable features and development tasks.

3. `/factory-tests` - the professor

Before anyone writes code, every task gets an exam. This skill defines the success criteria for each task, and how the coding agent can prove to itself that what it built works or needs iteration.

4. `/factory-explain` - the presenter

Explains the plan to me like I'm 10, with visual metaphor and mermaid charts. Now that coding agents can write more code, faster than any human ever could, the new bottleneck is human understanding of the code. This skill solves that.

5. `/factory-handoff` - from CTO to SWE

This packages the brief, plan, tests, safety rails, and stop conditions into one work order. Factory uses the best models for the planning in the previous steps above, then hands the work order to a lower token usage model like Grok 4.5 for execution.

6. Cursor or Claude Code `/loop` - the coffee

The night shift picks one task, builds it, takes its exam, records what happened, iterates if needed, then moves onto the next task. If it gets stuck, circuit breakers stop it from confidently digging a deeper hole while I sleep.

7. `/factory-review` - the teacher grades the homework

The student doesn't grade it's own homework. A fresh agent that never met the builder tries to break the result. The reviewer rereads the original plan, reruns tests, and finds anything that's broken.

8. `/auto-loom-proof` - shows the evidence

Uses browser use and screen records itself performing the tests and adds an 11labs voiceover explaining what's being proven. It sends me the narrated demo video.

9. `/factory-explain` - the code

The factory updates a plain-language owner's manual explaining what actually got built. I understand my own codebase, so I can make decisions without becoming the bottleneck or outsourcing my thinking to AI.

NOTE ON BUILDING AI

Cursor and Claude Code have made writing code dramatically easier. But getting AI to work reliably and at scale for you can't be fully automated.

LLM-as-judge helps, but a judge needs a rubric, examples, and input from someone with subject matter expertise. You still need a human reviewing the work and teaching the system how to perform better.

You can check out my factory on github in the post below.

[Yuandong Tian @tydsh]

🚨A novel way to do RL in LLM post-training!

Inspired by our previous path-not-taken work (https://t.co/hKe4qkgiXL), we dig deep into the learning trajectory of RL and find that optimizing singular vectors (i.e., rotation) of weight matrices suffices for good performance in RL. The resulting “isospectral optimization” reaches matched scores with substantially fewer training steps.

Great work from @zhu_hanqing666 and the co-authors!

[Peter Steinberger 🦞 @steipete]

Hit a new record with our autoreview skill. 66 rounds on a gnarly refactor. https://t.co/zbUjIS2LQQ

[Shangyin Tan @ShangyinT]

Many LLM-based optimization algorithms have been proposed recently. We find that GEPA, AutoResearch, and Meta-Harness each win on different tasks and no single LLM optimizer wins everywhere.

How do we leverage the complementary performance profiles of these optimization algorithms?

Today, we release optimize_anything omni: a meta-optimizer that leverages the best of each optimizer. At a matched budget, omni beats every standalone optimizer. (thread)

[jack @jack]

we're launching BUZZ!

a new groupchat platform for teams of people and agents of all sizes, built to reduce our dependency on slack and github. model-agnostic, decentralized, self-sovereign, and open source. 🐝

https://t.co/8IaMVeTQNo

[Fei-Fei Li @drfeifei]

The world is not just made of words, and spatial intelligence was never just about perceiving and generating worlds. It's about interacting with them.

Today, SceniX is joining World Labs. 🌎🤖👇

[Max For AI @MaxForAI]

这就是为什么我们需要开源人工智能模型的原因

7月16日，全球最大的AI模型平台Hugging Face @huggingface ，发表了一篇文章，宣布他们被一个全自动AI Agent攻入了他们的服务器。

整个过程中几乎没有人类操作。

攻击从一个恶意数据集开始，Agent串联了数据处��系统中的两个代码执行漏洞，获得节点级权限，窃取云端和集群凭据，并在多个内部集群之间横向移动。

它调用了数千次自动操作，运行在大量短生命周期���箱中，甚至建立了可以自行迁移的C2控制系统。

Hugging Face最终记录了超过17000个攻击事件。

但更荒诞的事情发生在调查阶段。

Hugging Face最初尝试用美国闭源前沿模型 @ChatGPTapp @claudeai 的API分析日志，但真实的攻击命令、漏洞Payload和C2数据触发了安全护栏，模型直接拒绝工作。

最后，他们只能在自己的服务器上运行中国 @Zai_org 的开源模型GLM 5.2。

GLM 5.2在几个小时内分析完17000多条记录，重建攻击时间线，定位被窃取的凭据，并区分真实攻击和诱饵行为。

如果不是世界上存在GLM5.2这样的开源顶级模型，Hugging Face绝对没有任何能力阻止这次入侵。

因为人类的响应速度远远不如Agent，数百人的安全��队可能都不如同时运行的几个攻击Agent。

这事最有趣的部分是攻击者没有任何安全限制，防守者却被安全护栏拦住了。

当攻击和防御都开始Agent化，真正重要的已经不只是模型能力。

谁能自己运行、审计和控制模型，谁才拥有机器速度时代的安全主动权。

这有点像核武器，很只有形成互相核威慑才可以保证世界的稳定。

[Boxi Yu @BoshCavendish]

@hxiao really good idea

[Boxi Yu @BoshCavendish]

Presenting our ICML 2026 poster tomorrow! 🚀
Come by if you’re interested in coding agents, benchmark quality, or AI evaluation—we’d love to chat!

SWE-ABS shows that SWE-Bench doesn’t just reject correct patches—it also accepts incorrect ones.

📊 19.78% of “solved” patches from top-30 agents are actually wrong.
📉 Every top-30 ranking changes after strengthening the tests.
📍 Hall A
🗓️ Jul 8, 9:30–11:15 AM
🎯 Poster #1804

#ICML2026 #CodingAgents #AIEvaluation

[Boxi Yu @BoshCavendish]

@elonmusk Looking forward to the future.

[Jeff Dean @JeffDean]

@uwcse @MBalazinska Here are a few of the pieces of advice that I shared:

[Qiuyang Mang @MangQiuyang]

https://t.co/rl818g2dSV

[ClaudeDevs @ClaudeDevs]

We've reset 5-hour and weekly rate limits for all users. Enjoy Fable 5!