Rick

AI Infra, DRAM/SSD | PhD

[SK hynix @SKhynix]

@SKhynix has released its second-quarter 2026 financial results.

The quarter was marked by another record performance, continued progress in HBM4, and expanded long-term customer partnerships.

🔗https://t.co/5kJbeW78sb

#SKhynix #AIMemory #HBM #AIInfrastructure

[Kimi.ai @Kimi_Moonshot]

Kimi K3 (open weights, coming soon)

[SK hynix @SKhynix]

AI innovation is built on collaboration.
@SKhynix is proud to collaborate with @NVIDIA on next-generation AI memory technologies, helping build the foundation for the future of AI infrastructure.

#SKhynix #NVIDIA #Partnership

[NVIDIA @nvidia]

SK Group and NVIDIA today announced a $500 billion-plus initiative spanning AI factories and next-generation memory.

➡️ @SKtelecom is building a 2-gigawatt NVIDIA Vera Rubin DSX AI Factory in Korea to serve global compute demand.
➡️ @SKhynix and NVIDIA will codevelop next-generation AI memory, including HBM.

Korea is ready to become a global AI powerhouse with world-class networks and data centers, leadership in chip technology and vast industrial scale.

[Jensen Huang @JensenHuang]

For my first post, I’m sharing a letter @NVIDIA signed on why open models matter.

AI will transform every industry, power every company, and be built by every country.

Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.

The world needs both frontier closed models and frontier open models.

https://t.co/AUKzoQ5Ikb

[Simons Foundation @SimonsFdn]

During today's #ICM2026 opening ceremonies, the IMU announced the 2026 Fields Medals recipients:

@UChicago's Yu Deng, @stonybrooku's John Pardon, @UofT's Jacob Tsimerman, and Hong Wang of @nyuniversity and @Institut_IHES.

Read more: https://t.co/oJARAI8I8j

[Rick @rickatny]

两位新晋菲尔兹数学家到底研究了什么

数学研究常常看起来离生活很远，但很多重要问题都可以用简单的画面理解。王虹和邓煜研究的是两个不同的问题：一个研究大量不同方向的线如何占据空间，另一个研究大量粒子或波如何形成整体规律。

王虹：很多方向的线，能挤在多小的空间里

想象一根长度固定的针。

我们希望找到一个区域，使这根针可以在其中朝向任何方向。针的位置可以移动，只要每个方向都能放进去就行。

直觉上，如果一个区域能容纳所有方向的针，它应该占据很大的空间。但数学家发现，在二维平面上，可以构造出一种极其复杂的集合。它几乎没有面积，却包含朝向所有方向的线段。

这就带来一个更困难的问题：

在三维空间中，包含所有方向线段的集合，能不能仍然像一张很薄的纸，或者像一团复杂的骨架？

王虹的工作证明，在三维情况下，这种集合必须具有完整的三维结构。它可以非常零碎，体积也可能很小，但从数学上的“维数”来看，它不能被压缩成二维或更低维的东西。

这个问题为什么重要？

因为在数学和物理中，波经常被看成沿某个方向传播的细长波束。光、声音和量子波都可以分解成许多不同方向的波。研究这些线段如何重叠，也是在研究不同方向的波能否集中在很小的区域中。

王虹的结果说明，大量方向不同的波无法无限集中在一个过于简单的空间结构里。

邓煜：为什么复杂的微观运动会形成简单的宏观规律

邓煜研究的是另一类问题。

想象一个盒子里有数量巨大的气体分子。每个分子都在运动，并不断与其他分子碰撞。

如果只看一个分子，它的运动可以用牛顿力学描述。但现实中，一个盒子里可能有数以亿计的分子。我们不可能逐个计算它们的位置和速度。

因此，物理学家通常使用一个宏观方程，只描述气体整体的温度、密度和速度分布。

问题在于：

我们能否从每个粒子的微观运动，严格推导出描述整体气体的宏观方程？

这件事非常困难。一个粒子可能先撞到第二个粒子，再撞到第三个粒子。第三个粒子又会影响其他粒子。时间越长，碰撞关系越复杂，像一棵不断分叉的巨大树。

邓煜和合作者发展了一套方法，用来分析这些复杂的碰撞历史。他们证明，在合适条件下，大量粒子的微观运动确实会形成稳定的统计规律，并由玻尔兹曼方程描述。

他还研究了类似的波动问题。

例如海面上有大量小波。每个波都有自己的方向、频率和相位。它们不断相互作用，能量会在不同频率之间传递。

邓煜的工作说明，当波的数量足够多时，这些复杂的微观相互作用可以被一个更简单的宏观方程描述。

两项工作的共同点

王虹研究的是空间中的复杂性。

她回答的问题是：

大量不同方向的线和波，最多能集中到什么程度？

邓煜研究的是时间演化中的复杂性。

他回答的问题是：

大量粒子或波经过长期互动后，为什么会形成稳定的整体规律？

两人的工作都在处理同一个更深层的问题：

当系统中包含数量巨大的简单对象时，整体会出现什么样的结构？

王虹证明，方向足够丰富时，空间结构不能过于简单。

邓煜证明，微观互动虽然极其复杂，整体仍然可能服从清晰的数学规律。

这也是现代数学的重要作用之一：从看似混乱的现象中，找出隐藏的结构和规律。

[Rick @rickatny]

DeepSeek’s 4-hour meeting mapped the path: CoT → agents → continual learning → model self-improvement. Open weights and low API prices compress model margins. Cheaper inference expands compute demand. China’s real constraints are chip supply, cluster efficiency, and talent.

https://t.co/KQTulFlQRQ

[Andrew Conner @connerdelights]

Terence Tao posted his ChatGPT session trying to understand the Jacobian conjecture counterexample. It's so lovely reading a slice of how his mind works, the connections he's making, etc.

https://t.co/wu6CcAk3V6

[Rick @rickatny]

Wistron Celebrates Grand Opening of First U.S. Smart Factory Marking Milestone in Global Smart Manufacturing Strategy https://t.co/qm4jgZENkA

[Elon Musk @elonmusk]

@huaijiangzhu It is pretentious and poisonous for for-profit companies to call themselves “labs” when they are actually corporations and to create a lords & peasants culture between (looksmaxxing) “researchers” & engineers when they are both, on their better days, actually engineers

[Rick @rickatny]

Spent some time this weekend for the post.

TL;DR: Pretraining is no longer a pure scaling race. Frontier advantage now comes from better data, efficient memory and communication, reliable training systems, and lower lifetime inference cost.

https://t.co/amviOHB3as

[Rick @rickatny]

AI hardware cycle is still intact

[Rick @rickatny]

Amazing

[Arena.ai @arena]

Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5.

This is a 17-place jump from Kimi-k2.6 (#18 -> #1).

In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5.

The full model weights will be released by July 27.

Congrats to the @Kimi_Moonshot team on this major milestone!

[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. 🧵

[Rick @rickatny]

Why a Busy GPU Is Not a Useful GPU in LLM Serving https://t.co/dvtrsqv8BX

[Silicon Data @Silicon_Data]

When it rains, it pours. Since Jul 2, our H100 GPU rental curve has risen yet again with even more providers hiking prices. Our perspective below on "meta selling compute" also seems to have been affirmed by today's news. Acc to our data, compute market continues to tighten.

[Ben Pouladian @benitoz]

SK Hynix just got bid 7x oversubscribed for its Nasdaq ADR

The same week everyone shared the "HBM is a mistake" bear note

We checked the JEDEC specs. The stack is fine. The tax is levied on the system, and it names who gets paid next

New, with @rickatny Free
https://t.co/PAMmvYpH7Y

[Lilian Weng @lilianweng]

new post on harness engineering for AI self-improvement: https://t.co/ZYvGfVs61k

It is hard to forecast how much the future of RSI will rely on harnesses. Likely harness engineering will evolve in the direction of self-improvement and enable auto-research, and, in turn, smarter models keeps harnesses simple.

Even when many harness improvement get eventually internalized into core model, the need to specify goals and context will not disappear.