X.Z.

PostDoc @Penn & Ph.D. @ MMLab-CUHK & Prev. @Nvidia. Working on Machine Learning and Computer Vision.

[Kording Lab 🦖 @KordingLab]

Come to a brewery to night, have a cold one, and talk ai with Bhuv Jain and yours truly. AI is awesome and it also sucks. Let’s talk about it. https://t.co/vzR8jm9SIw

[Joey Rudoler @JRudoler]

Wrote a new, modern stats curriculum.

Teach about probability and sampling via computational examples / simulations with real data. It's unbelievably helpful for intuition. Everything else follows.

Online and open-source: https://t.co/uXeRCqsnEW

[Yixiao Ge @ge_yixiao]

https://t.co/YZQdDo6wUh

ARC-Hunyuan-Video-7B is open-sourced! 🎉 It crushes real-world video understanding—e2e AV reasoning, precise grounding, deep thematic analysis, for short/long videos.

Available on HF with transformers support, vLLM deployment, online demo & API. Try it!

[Math Cafe @Riazi_Cafe_en]

Stanford's "Parallel Computing"

Youtube: https://t.co/AXV5VDSGcs

[X.Z. @XZhang_Vision]

@KordingLab If you mean the second picture, I think it’s warning about going the wrong way?

[Michael Black @Michael_J_Black]

Many people are in the middle of the @CVPR deadline. So I'm sharing my guide to writing a CVPR paper (or any paper). My students have had this for years but I haven't shared it publicly before. I hope you find it useful and write a great paper. #CVPR2025 https://t.co/RAvnQFnuLQ

[Frederik Gieschen @FrederikNeckar]

Jim Simons: "My algorithm has always been: you get smart people together and you give them a lot of freedom. Create an atmosphere where everyone talks to everyone else. Provide the best infrastructure. And make everyone partners. That was the model that we used in Renaissance."

[Kording Lab 🦖 @KordingLab]

Everyone in science has to write abstracts. See how I edit abstracts.

[Guillaume Lample @ NeurIPS 2024 @GuillaumeLample]

Today, we are releasing Mistral Large, our latest model. Mistral Large is vastly superior to Mistral Medium, handles 32k tokens of context, and is natively fluent in English, French, Spanish, German, and Italian.

We have also updated Mistral Small on our API to a model that is significantly better (and faster) than Mixtral 8x7B.

Lastly, we are introducing Le Chat (https://t.co/CCMpILcmFy), a chat interface (currently in beta) on top of our models.

[Soumith Chintala @soumithchintala]

Based on all the user-request videos that @sama's been posting, it looks like sora is powered by a Game Engine, and generates artifacts and parameters for the Game Engine. 🤔

[Song Han @songhan_mit]

It's time to systematically study the fundamental principles of GPU-accelerated AI computing: https://t.co/GxGN9N6oFX

[X.Z. @XZhang_Vision]

Perhaps the tech should be on transferable causality? I read many papers on AI interpretability and noticed that people define it differently. Some think it's about using experiments to explain how NNs think and make intuitive sense. Others use math to define and analyze what these networks really understand. Most of these just work on seeing correlation, but the true goal should be about finding causality. However, causality is challenging to establish, whereas correlation is relatively easier to demonstrate. Then considering that discovering causality often requires high-quality data, there's an interesting situation: communities with large-scale, high-quality data, like in vision and NLP, don't have a strong demand for interpretability or causality. In contrast, fields that need interpretability more, like medicine, finance, and sociology, often lack quality data. This might suggest that a tech on transferable causality could be a valuable tool in the future, much like a foundation causal model, bridging the gap between data-rich and interpretability-needy domains.

[Chase Lean @chaseleantj]

Meta just launched their new AI image generator.

But how good is it?

I compared it to Midjourney, DALL-E 3 and Adobe Firefly across 10 image categories.

Here are the results:

[Greg Brockman @gdb]

GPT-4 for radiology. Far from perfect, but state-of-the-art performance on some tasks: “Surprisingly, we found radiology report summaries generated by GPT-4 to be comparable and, in some cases, even preferred over those written by experienced radiologists”
https://t.co/bi6RwqgeHw

[X.Z. @XZhang_Vision]

@pranavrajpurkar What if I were on VS Code on the right-hand side of the screen, with a LubMed article open on the left-hand side of the screen, where should I be?🤔

[Brandon G. Neri @nerijs]

NVIDIA released new drivers and a AUTO1111 extension for using TensorRT with SD 1.5 and SDXL

https://t.co/jyotFEvAr1

Works like a charm! Conversion on a 4090 is ~2mins and support multiple resolutions, batch sizes and loras!!

[Kording Lab 🦖 @KordingLab]

I will probably onboard one PhD student for science of science and one to fit big models to neural data/ neuroAI. Relevant prior experience and love for nerding out/ building algorithms and math will be needed.

[X.Z. @XZhang_Vision]

Well well well...

[Gautam Kamath @thegautamkamath]

A realistic simulation of what it's like to be a grad student? I was only clicking buttons but the frequent rejection was triggering 🙃
It took me just under 7 years to graduate, can you do better? https://t.co/hCMpQmOFrG

[X.Z. @XZhang_Vision]

@junting9 Lol cannot work out together this week, bro

[X.Z. @XZhang_Vision]

@_akhaliq Dropout is for better learning😆