Vinh VC

AI Researcher/Scientist, Building the next top-tier Deep Learning/Neural Network engine

[Nadun Ranawaka @NadunRanawakaA]

Real robot data is expensive. Real robot evaluations are slow.

Excited to share SimFoundry - a system that turns real scenes into sim-ready worlds for training and benchmarking robots at scale -

✅Automated Scene Reconstruction with asset generation
✅Handles clutter, articulated objects, multiple robot embodiments
✅High Correlation Real-to-Sim Evals
✅Zero-shot Sim-to-Real
✅Generates diverse digital cousins

Less manual environment authoring, more scalable feedback for robot learning.

🌐https://t.co/jmCno7eslU

🧵1/9

[me @twetsfyp]

Mito Claude is back in an insane way

This is a 16-minute tutorial on how to create $50,000 cinematic websites (step by step) with Claude Fable 5

[Mikel Artetxe @artetxem]

We are releasing a new benchmark to evaluate world models!

[Francois Chaubard @FrancoisChauba1]

New paper coming soon.. teaser..

no transformer, no backprop, no problem!

Zero Order CAN pretrain!

very exciting.. stay tuned!

[Alex Imas @alexolegimas]

This is an incredibly cool paper.

[Stéphane Deny @StphTphsn1]

"Neural Fields as World Models"
by Joshua Nunley (Feb 2026)

🔗https://t.co/KMnrwRk8OM

[DFRobot @dfrobotcn]

An open-source robot vacuum you build yourself — Raspberry Pi, ROS 2, 2D LiDAR, Home Assistant, 3D printed chassis. No cloud, fully local.

oomwoo is early stage and building in public. The community can contribute modules in parallel — from SLAM navigation to dust bin design.

https://t.co/ip0HWptZg0

#ROS2 #RaspberryPi

[PatSnap Eureka @PatSnapEureka]

14 live IP MCP servers you can connect into your AI workflows, covering novelty search, global patent FTO, design patent infringement search, legal status, landscape analytics, and advanced patent search.

[Ziad Obermeyer @oziadias]

AI can help us *do* useful things in medicine – great!

Below - one such thing, from our @nature paper:
- AI can flag people at high risk of dropping dead
- and help us decide who gets implanted defibrillators

But what can we *learn* from AI?

https://t.co/InNKVcSEtc

[deep Manifold @BetaTomorrow]

Title: Reasoning as Attractor Dynamics: Latent Memory Retrieval via Gibbs-Weighted Energy Minimization
Author: Kanishk Awadhiya

#DeepManifoldInterpretation
** The central limitation of the Hopfield formulation is not its definition of equilibrium. Equilibrium as an unchanged state is mathematically standard. Its foundational hypothesis is that memory and neural computation should be represented by convergence under a fixed operator toward static, stable states. This may describe restricted associative retrieval, but it should not be generalized into a theory of dynamic neural inference or reasoning. **

From the Deep Manifold view, the relevant object is not necessarily one static equilibrium. It may instead be a prompt-conditioned, dynamically generated fixed-point class reached through multiple stochastic iterated-integral pathways. The operator, boundary and local geometry co-evolve, so “the neural state no longer changes” is too restrictive—and perhaps the wrong definition of successful inference.

For the paper we are discussing, the problem becomes even greater: it borrows Hopfield’s equilibrium interpretation without even possessing Hopfield’s recurrent state dynamics. It has neither a fixed neural-state operator nor demonstrated convergence. The supposed “attractor” is created by externally weighting completed trajectories.

[David Bessis @davidbessis]

As a 17 yo, I genuinely didn't understand how one of my classmate, a Will Hunting-type mathematical genius, could outsmart all of us without putting any effort, while sleeping and snoring throughout the lectures.
He went on to win an IMO gold medal.
This felt like magic and triggered my interest in mathematical cognition. I was quite good myself, but very far from him. I persevered, discovered effective techniques, made some progress, did a reasonably good PhD and went on to become a tier-2 mathematician.
Then, in my early 30s, my self-confidence increased a lot, I found ways to increase my focus and intuition, which triggered a success feedback loop where I became capable of working much harder, for much longer periods of time, with much better techniques.
I ended making stellar progress and proving first-rate results between 32 and 34 before I quit math.
My interest for mathematical cognition continued, which explains my book, Substack, and public engagement here.

[Aleph @alephneuro]

There are only a few ways to image the brain without surgery: electricity, magnets, light, sound. All existing methods either produce blurry images or require a room-sized machine.

Our image is captured with ultrasound transmitted through the skull, with a contrast agent.

[Aleph @alephneuro]

We recently obtained the highest-resolution 3D images of the human brain ever taken from outside the skull. This is the first look.

Introducing Aleph, a research lab building brain interfaces for the telepathic future. (1/n)

[Jonathan Richard Schwarz @schwarzjn_]

PhD apps in 2026 are wild: Every successful applicant to our new lab @tri_fair_lab has multiple 1st-author papers, incl. several orals at NeurIPS, etc.

Feel deeply for students who'd have earned a spot 5–10 years ago — same talent & drive, just a different moment in time.

[Phoenix Yin @Phoenixyin13]

我认为这是ICML 2026里的上乘工作。

传统 Transformer 的 Attention 机制，本质上是点对点匹配，把输入切成一堆 token，即离散点，然后算 Query 和 Key 的相似度，再加权 Value。

这在 NLP、图像上很管用，但对连续函数、场，比如物理模拟中的 PDE 偏微分方程、3D 几何、流体、材料力学等来说，就有严重的问题：

1. 连续的东西被强行离散化成 token，忽略了全局函数结构。
2. 对分辨率敏感，换个采样率就容易崩
3. 计算是 O(n²)，n 很大时很贵。

所以，Functional Attention应运而生。

这次，FuncAttn 把函数本身当成第一公民，不再用 softmax 算点对点相似度，而是用 结构化的线性算子，来建模两个函数空间之间的函数对应关系。

它的灵感，来自几何处理里的 Functional Maps，把复杂 3D 形状对应问题变成简单线性算子，实为巧妙。

结果是更紧凑（k×k 算子，k << n）、分辨率不变，捕捉全局依赖更好、在 PDE 求解、3D 分割、回归等任务上达到或超过 SOTA，还在 OOD泛化上表现突出。

未来，Functional Attention对 AI 应用的实际用处很大，尤其在科学计算 + AI和需要处理连续、物理世界数据的场景中。

它会成为Neural Operator与科学 AI 领域的强力升级组件。

这篇论文，会让 AI 向擅长连续物理世界迈进一大步。对科研加速、工程模拟、具身 AI、BCI 等前沿应用极度有价值。

这正是我认知神经愿景的核心基础设施之一。

[Xiuyu Li @sheriyuo]

aiming to push Deepseek-OCR one step further 🙌

[Mathematica @mathemetica]

Neural networks for dynamical systems map high-dimensional fields to low-dimensional manifold latent spaces, where dynamics evolve before the decoder reconstructs the predicted field.

The encoder, consisting of convolutional layers and multilayer perceptrons, compresses the initial field into latent coordinates w and z on the manifold.
The upper part visualizes the encode-project-decode process with example point cloud manifolds.

It is used to accelerate long-term predictions of complex fields in scientific computing applications like fluid dynamics simulations.

[Arkil Patel @arkil_patel]

Can we release all the weights of an LLM but still provide differential access to privileged users?

Yes! We introduce: 𝗧𝗶𝗲𝗿𝗲𝗱 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗠𝗼𝗱𝗲𝗹𝘀 (𝗧𝗟𝗠𝘀). Define access tiers corresponding to different computation graphs over the same set of LLM parameters!

[Guandao Yang @GuandaoYang]

There are three reasons why it's called "Functional Attention."

1⃣We make functions, instead of tokens, as first-class citizens for attention module.

2⃣It was inspired by the geometry processing concepts of functional maps.

3⃣It works!

[Hirokatsu Kataoka | 片岡裕雄 @HirokatuKataoka]

“Replacing Labeled Real-image Datasets with Auto-generated Contours”

A year after our "without natural images" paper, we showed that Vision Transformers pre-trained on formula-generated images can match or even exceed the recognition performance achieved with large-scale real-image datasets such as ImageNet-21k.