Xin Yan

Research Scientist @ ByteDance Seed | Prev. @UWaterloo @MITIBMLab @01AI_Yi @WHU_1893

[EchoXu @EchoXuAI]

怀念在 GPT-3.5 时代写代码的日子。

那时，LLM 还只是一个朴素的对话框，但我却每天都能学到新东西，感觉自己在实实在在地变强。不再需要自己翻文档、读报��、反复调试。它把那些繁琐、枯燥、耗神的部分统统挡在外面，只把思考和设计的乐趣留给我。看着它在浏览器里一行行吐出的代码和解释，我常常带着一种近乎孩子气的满足。那时的 AI，既提高了效率，也守护了创造的快乐。而如今的 coding agent，把效率提升了十倍甚至百倍。它几乎接管了整个流程，却把我从创造的核心里彻底摘除。我的思考速度已跟不上它输出、执行、测试、修复的节奏。望着那些命名规范到近乎冷酷、结构整齐得像被算法亲自梳理过的配置文件和代码，我知道自己完成了更��的工作，却再也找不到曾经那种「我正在亲手把想法变成现实」的喜悦。我得到了大量产出，却仿佛失去了创造本身。

[JUNDE WU @JundeMorsenWu]

I’m an AI PhD at Oxford wanted to move to the US for years, but it never worked out. Maybe I can share why.

To me, it feels like the US is making it harder and harder for Chinese talent to come or stay.

First, Trump signed PP10043, which meant I couldn’t go to the US for any graduate school. That’s why I had to turn down my Stanford offer and come to Oxford.

Later, I qualified for an EB-1A (the extraordinary ability green card). But because I was born in China, I’m stuck in a country-specific backlog that could take another 4–5 years.

Meanwhile, on the other side, I’m getting 3–4 emails from China almost every week. They offer high salaries, free housing, generous research funding, and even offered an astonishing amount of money just for me to come back and have a conversation.

I’ve always loved the American spirit. I truly believe my abilities could have a bigger impact there. But honestly, the whole process has been incredibly frustrating.

All of this makes me wonder whether I should just give up.

Why am I spending so much energy trying to go to a country that doesn’t seem to want me, when another country is doing everything it can to bring me back?

[Xin Yan @cakeyan9]

@masoncaiws 想要这么个东西很久了，试试看

[Xin Yan @cakeyan9]

DeepSeek and Moonshot have been my two favorite Chinese startups for over two years.
The key is their research taste.
Congratulations!

[Kimi.ai @Kimi_Moonshot]

Introducing Kimi K3: Open Frontier Intelligence

🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal
🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts
🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost
🔹 Built for long-horizon agentic coding and self-evolving workflows

Kimi K3 is now live on on https://t.co/zrk6zZxZUo, Kimi Work, Kimi Code, and the Kimi API.
Open Weights by July 27, 2026.

🔗 API: https://t.co/XCrgjXAqMw
🔗 Tech blog: https://t.co/YTfiMSNM1f

[Xin Yan @cakeyan9]

@zxytim K3 is superb! Congratulations

[Xin Yan @cakeyan9]

it's a good model

[Thinking Machines @thinkymachines]

Today, we are introducing Inkling.

Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available.

https://t.co/Ghebq5mG30

Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵

[zappy @velocizapkar]

@nrehiew_ For convolutions on KV, check out dconv from the Primer paper and Canon B layers from Physics of LLM

[BytePlus @BytePlusGlobal]

Thrilled to see Seedream 5.0 Pro debut at #2 on @arena ’s Multi-Image Edit leaderboard.

It’s especially encouraging to see the leap from #11 with Seedream 4.5 to #2 with Seedream 5.0 Pro. Independent community evaluations like Arena help push everyone to build better models.

Congratulations to the incredible research and engineering teams behind Seedream, and thank you to @arena for including us in the benchmark. Looking forward to what’s next.

[Xin Yan @cakeyan9]

@ShanningZhuang 检验你的harness功底

[Xin Yan @cakeyan9]

已严肃学习

[Owen Tian Ye @tiny85114767]

这个周五发现了 AgentTank，一玩就停不下来玩了两天。我搭建了一套系统，让我的 Agent Tank 在不到 24 小时内从 0 分冲进全球 Top 10。
我把完整过程和关键经验写成了一篇博客，希望能帮助新玩家快速上手！
https://t.co/RZVHLrqzlo

[Arena.ai @arena]

Seedream-5.0 Pro enters the Multi-Image Edit Arena at #2 with 1415 pts.

Strong improvement compared to Seedream-4.5 (#11 -&gt; #2)!

[ben @contraben]

Can an image be "Too Perfect"?

Creatives judged @BytePlusGlobal's Seedream 5.0 Pro blind against:

>@OpenAI's ChatGPT Images 2.0,
>@GeminiApp's Nano Banana Pro, and
>@bfl_ml's Flux 2.

It won photorealism outright.

But, it got dinged for feeling too perfect.

Some insights 🧵

[Xin Yan @cakeyan9]

Muse Image is really good! Can't wait to see a battle with our Seedream 5.0 Pro ⚔️

[Yushi Hu @huyushi98]

We built the entire Muse-Image/Video post-training stack in just a few months, and we are amazed everyday by how well everything scales for multimodal generation, beyond language. 😎

Super grateful to work in such a talent-dense team!

https://t.co/FTUlYQgKup

[Xin Yan @cakeyan9]

Excited to release our Seedream 5.0 Pro. Everyone should give it a try!

See Blog here: https://t.co/IfJg2Fi5fm

[BytePlus @BytePlusGlobal]

Dola Seedream 5.0 Pro API is now available on BytePlus.

AI image generation is evolving beyond creating a single image.

Edit with precision. Visualize complex information. Render realistic images and portraits. Create across languages.

Create production-ready visual assets for enterprise workflows.

[Xin Yan @cakeyan9]

@turingou 可以试试2k分辨率

[Xin Yan @cakeyan9]

@ShanningZhuang 这么爽

[Xin Yan @cakeyan9]

Jack Lindsey的文章一直都挺有趣的
不过国内大厂似乎不太关心这个方向...

[[email protected] @Jiaxi_Cui]

“为什么 latent space 和自然语言空间之间没有其他表示层？”

Claude 的 J-space，某种程度上回答了这个问题。

我之前说 text CoT 更像当前阶段的工程折中：好监督、好验证、好被 reward model 评估，也好被用户、媒体和投资人理解。

所以它能 scale，但不代表它是终点。

J-space 把问题往前推了一层。

模型不是全程在自然语言里 thinking，也不是完全在不可读的 latent space 里 thinking。更准确地说，它内部存在一个可语言化的 latent workspace。

大量底层计算发生在 latent space；但那些需要被复用、反思、审计、对齐的中间思想，会进入这个 workspace。

所以它在科学上未必神秘。steering vector、logit lens 之后，大家早就知道 residual stream 里能塞下大量未说出口的语义、倾向和中间状态。

但它在工程上非常恐怖。

传统 RL / RLHF / RLAIF，本质上是：

prompt → model output → reward model 打分 → 更新模型

但如果 J-space 走通，训练范式可能变成：

prompt → model internal J-space + output → critic 打分 → 更新模型

这不是训练模型最后说什么。

这是训练模型在关键时刻“应该意识到什么”。

以前我们训练模型的嘴。 现在可能开始训练模型的内心独白。
这��是 Anthropic 这篇最值钱的地方。

它不只是 interpretability，而是可能影响 RL、对齐、长期记忆、continuous learning、fine-tuning 本身 甚至更多这两三年我们习以为常的操作。

知识库是外挂，微调是重训。 但 J-space 如果成熟，可能会出现第三种东西：直接塑造模型在某类场景下会激活什么概念、记住什么经验、调用什么原则。

而值得注意的是这一篇工作似乎诞生于 Sonnet4.5 时代，Anthropic 比 OpenAI 更让人震撼。

我甚至能想到前段时间大量 Researcher 批量加入 Anthropic 的场景：
"""
X: "我不喜欢你们，你们是一家非常喜欢封号的公司"

Claude: “要不你看看我们半年前研究的这些工作再说，但这些论文都��近 close 了，你进来肯定挂不上这些名了，但你可以自己探索新方向”

X: "我什么时候可以入职！"
"""

早期 OpenAI 迷人的地方，是 GPT-2、GPT-3、CLIP、Whisper 那种朴素但强大的信念：

用数据、规模和统一范式，打破之前学术界“架构优先”的惯性。

那时候的 OpenAI 很像一个有明确科学直觉的组织。它知道自己相信什么，也知道自己要把世界往哪个方向推。

但现在的 OpenAI 越来越像产品和算力公司，强是强，却越来越难感受到一套属于自己的科学理念。

Anthropic 反而还在追问更底层的问题，J-space 不是证明模型有灵魂，它想证明模型的“内心独白”可能可以做成 API。

一开始我还以为 J-space 只是随手起的名字，后来仔细看才发现这个 J 是 Jacobian。

这就���浪漫了。

Jacobian 把多元复杂函数在局部近似成可操作的线性函数；J-space 则像是在把模型内部高维、复杂、混沌的 latent space，投影成一组可以读取、干预、训练的方向。

这种命名的美感，让我想到 SpaceX。

其实我从高一开始关注到 Musk 的，那会我每周都会买很多杂志，在其中一期看到了他完整的创业记录，只是近几年所有人都认识他了，所以也不怎么提了。

当时令我印象深刻的倒不是今天熟悉的星舰，而是 2008 年 Falcon 1 的第四次发射。前三次全部失败，第四次几乎凑出了最后所有的钱做发射，包括 Tesla

那一次 Falcon 1 入轨，SpaceX 才活了下来。

后来他们把海上火箭回收船命名为 Of Course I Still Love You。这���字出自科幻小说，不是冷冰冰的工程代号，而像一句情书，去接住一枚从天上返回的火箭。

所以迷人的地方是极端硬核的工程、生死边缘的赌局，最后被一种近乎浪漫的命名接住。

科学、工程和浪漫，在这里突然接上了。

Author List 中 Wes Gurnee @wesg52 和 Jack Lindsey @Jack_W_Lindsey 构思了整套方法，以及语言表达与意识访问之间的联系，伟大无需多言。

有两位华人 Rowan Wang 和 Runjin Chen 参与了 Reflection 训练部分

[Jack Morris @jxmnop]

my paper won an award at icml 😁

some thoughts:
• this work was rejected from NeurIPS. i cleaned it up a small amount and it got great reviews from ICML! don't give up
• ICML received 24k submissions and only gives out 7 awards, which is crazy. feeling grateful
• i distinctly remember sitting at my desk two winters ago wondering if i would ever finish this project. most of all this is the product of sitting down and forcing myself to keep working for several months straight. the results emerged from running the experiments over and over and fixing a long sequence of tiny details. eventually, the curves looked like that 👇
• also happy that the insights in this paper are becoming more widely accepted: 3.3 bits/param, thinking about capacity "LLM as flashdrive" mentality
• the method here is used successfully for selecting midtraining data at least one frontier lab, which is cool!
• i am grateful to my collaborators, but Meta is no longer a great place for academic research imo and this almost never got published for a number of reasons. i shall not elaborate further
• for future work, i think analyzing the implications of on-policy algorithms on capacity, as well as LoRA and things like it, are fruitful potential research directions
• sadly i'm not in Korea but am following the conference online from california and happy to chat!

a nice end to one phase of my research career :)

[OōEli.eth @ooeli_eth]

是的，靠投资美股实现财务自由的概率远大于 在字节混到 4-1 职级。各位加油！