We share Code-Guided Synthetic Data Generation: using LLM-generated code to create multimodal datasets for text-rich images, such as charts📊, documents📄, etc., to enhance Vision-Language Models. Website: yueyang1996.github.io/cosyn/ Dataset: huggingface.co/datasets/allen… Paper:

Claude 3.7 one-shotted this 3D room— everything from layout to geometry and texture. This is always a fun way to test a model’s spatial reasoning ability

🧩 Built an MCP that lets Claude talk directly to Blender. It helps you create beautiful 3D scenes using just prompts! Here’s a demo of me creating a “low-poly dragon guarding treasure” scene in just a few sentences👇

🧑‍🎨 The future of creative tools will look very different. 🧠 Imagine an AI control-centre for orchestrating complex tasks using just prompts. 📽️ Demo: Prompting an evil dragon with soundtrack using a single control centre (Claude). It uses both Blender MCP and Ableton MCP.

we trained Kevin-32B = K(ernel D)evin using GRPO on KernelBench it's to our knowledge the first open model trained using RL on writing CUDA kernels it beats o3 & o4-mini in correctness & performance! shoutout to Carlo Pietro Marsella Ben Pan!! x.com/cognition_labs…

How to scale visual affordance learning that is fine-grained, task-conditioned, works in-the-wild, in dynamic envs? Introducing Unsupervised Affordance Distillation (UAD): distills affordances from off-the-shelf foundation models, *all without manual labels*. Very excited this

Check out this work led by Tianyu Hua! We believe code generation is the future but we need to start with better benchmarks

🤖 Household robots are becoming physically viable. But interacting with people in the home requires handling unseen, unconstrained, dynamic preferences, not just a complex physical domain. We introduce ROSETTA: a method to generate reward for such preferences cheaply. 🧵⬇️