Shan Chen

Teams are already post-training Nemotron 3.5 Super VL for tasks in document intelligence, cybersecurity, finance, scientific experimentation, and travel. The chart below shows how accuracy improves when the model is post-trained for each task.

I'll be on a panel at the @COLM_conf DAIH workshop on Friday! We're also hiring @sprinterhealth for Research Scientist, Applied Scientist, and Data Scientist roles. Stop by and say hi!

Come to the only health AI workshop @ COLM! — LLM/VLM Deployment Opportunities and Risks in Healthcare SF | 📅 Oct 9, 2026 COLM format research papers & 4 pages deployment case reports from companies and hospitals! ⏰ Deadline: June 23, 2026 🔗 daih2026.github.io/index.htm……

Grok Bot can literally run Grok Build inside its own computer So make Grok Bot install Grok Build and let it delegate tasks directly to the CLI Yes, Grok Bot can open and operate Grok Build for you...and since the task is actually running through Grok Build, you can use the Grok models with the highest reasoning level Basically: Grok Bot = orchestrator Grok Build = execution engine And since your Bot and Build are connected to separate usage pools, you can distribute workloads between them too This setup is insanely useful actually

randomly wonder when people work on some field + AI, do you usually think of it as building tools for people in that field, or building toward AI eventually doing what they do?

1/ We fine-tune a lot of customer models, so we decided to systematically try and figure out some best practices for finetuning. SFT isn't sexy, but it's still important. We vary one SFT lever at a time across 2 model families, dense + MoE to 235B, on 4 real-world customer datasets. What makes this clean is that each dataset is paired with an eval that took weeks to build with the customer, and the training outputs were generated to pass that eval. So the supervised target and the thing we measure downstream are the same criterion, which strips out the usual confounders

(1/n) Large Reasoning Models are trained with RLVR to improve their reasoning capability. But beyond improving accuracy, what else does RLVR change in model reasoning? Can it make reasoning more monitorable—so humans can reliably audit model behavior? 🤔