It was so great to talk about how the DeepSpeed team has been driving the frontier of LLM model training and how we will thrive under PyTorch Foundation. Kudos to the DeepSpeed core team Tunji, Masahiro, Minjia, Jeff, Logan, Stas, Guokai and Sam who continue driving the hard

Many posts have been discussing how #claudecode and other AI toolings significantly changed the software engineering profession, and how this profession might come to an end. Here is my perspective: AI models will and should outperform human beings on tasks that leveraging

Brilliant paper!

As 2026 approaches 🚀, it is a fitting moment to reflect on the achievements of 2025. This year has been exceptionally productive, and I am truly grateful for the opportunity to collaborate with outstanding teams and partners on a range of cutting-edge AI research initiatives.

Turkey grants Chinese ordinary passport visa-free access starting from Jan 2, 2026. It seems like Turkey can be one of the good places to host top ML conferences (e.g. #ICLR, #NeurIPS and #ICML etc.) given that it grants visa-free access to all US/Canada/European countries as

Check out our work on AHA: Audio Hallucination Alignment for Audio Large Language Models. In this work we built the diagnostic benchmark, proposed explicit taxonomy for audio hallucination, and also provide preference alignment datasets to help us beat SoTA performance on Audio

One controversial point: We should stop doing any live coding interviews for tech companies. Or at least let's not test candidates any tricky data structures & algorithms questions. At this moment, we are wasting everyone's time doing this, because it's no point to memorize all

As a researcher who works on both ML Systems and Multimodal models, 3D modeling and rendering is always a challenging field as it demands substantial resources and manual effort when scene editing is performed in the traditional manner. Despite recent progress in VLM-based

Check out our recent work on Efficient Knowledge Distillation of Large Reasoning Models (LRM) (arxiv.org/abs/2512.21002). Knowledge distillation (KD) of LRM involves distilling over lengthy sequences with prompt (P), chain-of-thought (CoT), and answer (A) sections, which is

Our paper "Reasoning Models Can be Accurately Pruned Via Chain-of-Thought Reconstruction" has been accepted to ICLR 2026! Kudos to Ryan Lucas, Qingquan (QQ) Song Shao Tang, Rahul Mazumder and Kayhan Behdin, and all the PCs/ACs/SPCs who put tremendous amount of work in it. See you in

🚀 Our work on "Scaling up Large Language Models Systems for Semantic Job Search" has been accepted to #MLSys2026 ! This paper highlights our years of science and engineering effort to make LLM serving more efficient, which can accommodate high-throughput, low latency online

I personally went through the experimental classes in China and grind all my way to get admission to Peking University. I ranked top 5 in my province and worked super hard. I feel one thing I appreciate Chinese education is that you don’t have too much distractions. If you like

Saw Jerry Tworek ‘s post about the necessity of coding interviews. I feel we should probably change the live coding interview to “code reviews” interview. After all it’s going to be what we need to do at jobs. During the interview we just show the candidates some piece of code,

It’s crazy, by far the most capable Video Generation model.