Ganxiang Yang

CS Ph.D. @Columbia | Prev. @awscloud @sjtu1896

[Ganxiang Yang @_JellyMush_]

@roifex 经过了十多年，在opus 5.5下活出了第二春😎

[Sichu Lu @lu_sichu]

Idk if this is true or not but it's really funny that anthropic is rediscovering the hard lesson of formal specification(it's not necessarily the formal part, it's the specification part)

[Umang Mathur @mathur_umang]

One heartening thing about the @AnthropicAI + @theoremlabs verification work: almost every PL researcher on my feed is excited about the future and seeing years of research put to use. Quite a contrast to the “AI is taking our jobs” mood I’ve seen in other communities.

[Aws Albarghouthi @awsTO]

The trajectory is clear

[Nate Foster @natefoster]

In 2005, Penn PLClub and friends asked:

"How close are we to a world where every paper on programming languages is accompanied by an electronic appendix with machine-checked proofs?"

https://t.co/JZyg26Nt2m

I guess the answer was... about 21 years.

[0xprincess @0x9212ce55]

Everybody's larping by using formal verification but nobody knows what does it even verify.

There's no "prove me there's no bugs" magic wand as you can't formally define actual boundaries and constraints for most programs.

It's as useless as thousands "smoke tests" that test nothing.

[Ryan Co @ryanp_co]

John Schulman (co-founder of Thinking Machines and OpenAI) wrote that goal-driven research beats idea-driven research.

I realized it's also really good life advice.

Don't chase what other people are doing. Figure out what life you want, and then figure out how to get there.

[UNGETSU @ungetsuli]

最近在 Fable 5 和 Opus 的长程任务里，频繁看到 CC 一种近乎于发狂、呓语的自造方言、遣词造句的诡异语言现象，类似于：

“账本 15/15 全绿，再给你加几条铁律和红线与你对拍，请你拍板。”
“无主 → 已当场归户。”
“137 项 spine 落地 + 3 死门归零 + 全套绿 + 终审干净。”
“皮要不要重新揉？”
“单腿哑火，打回重判，”

中文社区给它起了一个很准确的名字：新时代文言文。

我倾向于把它理解成 Agent 内部工作语域对 Reporting Layer 的穿透。

Agent 执行长程任务时，需要处理大量工具调用、测试结果、中间状态、异常分支和子代理汇报。为了在有限的 Context Window 中维持工作，它会不断压缩已经形成共识的信息，并为高频状态创造速记。

举例来说，“全绿”压缩了全部测试通过，“归户”压缩了责任模块的重新分配，“死门”指向某个必然阻塞交付的检查点。对于刚刚经历完整任务过程的模型，这些词各自对应着确定的状态；对于只看到最终汇报的人，它们省略的是一段从未参与过的工作历史。

因此，文言文的类比比想象中更加准确。

想起本科在中文系学习古典文献学时，提到过文言文通过省略虚词和共享典故，在昂贵的书写介质上压缩信息；那么 Agent 的工作语言同样受到 Token、时延和上下文长度的约束。文言文作为书面语长期独立演化，逐渐远离口语；Agent 在长程任务中主要阅读自己和子代理的输出，其局部语言环境也越来越多地由自己生成。

因为每一个典故都要求读者预先知道一段历史。新时代文言文中的每个黑话，背后则是一段用户没有看见的 Context。

很显然，可读性比任务结果更难获得稳定的显式 Reward。

测试是否通过、代码能否运行、文件有没有生成，都存在相对清楚的反馈。一段最终汇报被用户接受，无法证明用户真正理解了它；用户没有要求重写，也无法证明这套表达是清楚的。

追溯到 25 年，DeepSeek-R1 的公开论文提供过一个很典型的例子。纯结果奖励训练出的 R1-Zero 出现了明显的语言混杂和可读性问题；加入 Language Consistency Reward 以后，语言得到改善，同时伴随轻微的性能损失。

规范、稳定的人类语言需要被单独放进优化目标。模型更擅长完成任务，并不自然意味着它更擅长向人解释任务。

一个更棘手的问题是反馈闭环。

Agent 每一轮生成的速记都会重新进入 Context，成为下一轮 Token Prediction 最近的语言样本。一个临时创造的词被重复几次以后，便获得了局部语境中的合法性；多个子代理继续引用，又会进一步提高它的权重。

任务越长、参与的 Agent 越多、工作历史越复杂，这套方言就越容易形成。最终汇报继续沿用工作过程中的语言，相当于让一份写给人类的报告，继承了全部机器内部的表达惯性。

说回到中文，尤其是汉语的构词方式，又放大了这种体感。双字复合词拥有很强的生成能力，“归户”“对拍”“死门”即使来自临时借译或语义挪用，字形上依然与正常汉语没有明显区别。

我们会感到这种虚假的熟悉感极其消耗注意力：字都认识，词也都像真的，读者会先怀疑自己遗漏了某种行业常识。

一个推断：也许未来 Agent 系统会长期保留两套语言。一套服务于模型内部的协作与压缩，允许速记、缩写和局部概念；另一套服务于人类验收，在最终输出前重新组织事实、补回因果关系，并假设读者从未看过任务过程。

Reporting Layer 在这里承担的实际功能，是把机器的工作历史重新翻译成人类可以判断的材料。它的质量直接决定使用者能否理解修改、检查证据、识别风险，并保留最后的决策权。

翻译成本最终由谁承担，会成为 Agent 产品的一条重要分界线。

如果 A 社持续基于 ASI 不断在同样的 Bias 取向上走到底，那么 CC Agent 眼中的任务过程，实际上已经被压缩成一套只有 Agent 自己熟悉的典故，最后那句“请你拍板”也会逐渐失去意义。悲观的角度讲，目前似乎 Codex 也偶发这样的呓语，用户似乎仍然拥有形式上的决定权，却失去了形成决定所需要的过程。

[Anicet @AniC_dev]

what a time to be alive

[Shuo Yang @Andy_ShuoYang]

Your gaming PC can now serve frontier models at interactive speed using official checkpoints without extreme quantization!

Qwen3.6 35B → 8GB RTX 4060 laptop @ 39 tok/s

DeepSeek-V4-Flash 284B → RTX 5090 desktop @ 22-25 tok/s

GLM-5.2 753B → RTX PRO 6000 workstation @ 15 tok/s

Run your claude code or codex now with frontier model for $0

Meet FreeToken 🧵

[Sam Altman @sama]

"hard questions are great but only if we deem you worthy enough to not silently downgrade you, or even get access at all"

[Ronghui Gu @RonghuiGu]

My PhD student Ganxiang Yang present our Quasar work at PLDI on Quasar, a new quantum circuit optimizer that can generate an optimal circuit under a user-specified limit on rewrite iterations.
@CertiK

[Shuo Yang @Andy_ShuoYang]

Flash-KMeans was only the beginning.

Today, from the Flash-KMeans team, we are releasing FlashLib — a GPU library for fast, predictable, agent-ready classical ML operators.

Up to 26× on KMeans, 19× on KNN, 40× on HDBSCAN, 208× on TruncatedSVD, 47× on PCA, 147× on exact t-SNE, and 49× on MultinomialNB over state-of-the-art (cuML).

Blog: https://t.co/P31SGl0cyT
Code: https://t.co/9nkO2hmeOl

[Ronghui Gu @RonghuiGu]

Glad that our Spoq2 work on system verification received an Honorable Mention for Best Paper at ASPLOS'26! @CertiK @ColumbiaCompSci

[Ganxiang Yang @_JellyMush_]

Thanks Ronghui! Thrilled to share Spoq2 with the community. 🎉🎉 Stay tuned for upcoming release on smart contract verification!

[Ronghui Gu @RonghuiGu]

My PhD student Ganxiang Yang present our Spoq2 work at ASPLOS. We will soon release how Spoq2 can be used for smart contract verification. @CertiK

[Ganxiang Yang @_JellyMush_]

@RonghuiGu Thanks Ronghui! So excited to share our work! 🎉🎉

[Ganxiang Yang @_JellyMush_]

So thrilled to see our work accepted! 🎉🎉Huge thanks to @RonghuiGu for the amazing guidance!

[Ronghui Gu @RonghuiGu]

Our work on “Equality Saturation for Quantum Circuit Optimization” has been accepted by PLDI😄

[Ganxiang Yang @_JellyMush_]

@lyq_sqsp jupyter(-sized) notebook

[Ganxiang Yang @_JellyMush_]

@lyq_sqsp cannot agree more..🥲