Imagine a robot chef that learns to cook not through tedious programming, but by watching videos and reading recipes online—just like a human! That’s the culinary dream being cooked up in a new research paper. Researchers explored how to teach robots cooking skills using readily available internet resources like videos and large language models (LLMs). Traditional methods struggle because they lack the physical nuance of cooking, such as knowing how much pressure to apply when scraping a cutting board or the subtle wrist movements of stirring. This new approach tackles this challenge by giving the robot a library of basic cooking actions. The robot uses internet data to choose which of these pre-programmed actions best matches the recipe steps. One surprising discovery was that LLMs, despite lacking visual information, did a decent job selecting appropriate actions based on recipe instructions. Even more impressive, a system trained on video motion (optical flow) significantly outperformed a model trained on millions of video clips—demonstrating that recognizing fine motor movements is crucial for cooking tasks. By combining the LLM’s recipe understanding with the optical flow’s sensitivity to motion, the researchers achieved a 79% success rate across diverse skills like cutting, peeling, scraping, and stirring. While there's still work to be done before robots replace human chefs, this research brings us closer to a future of automated cooking. The next steps include incorporating even richer multimodal AI models and developing a system that can learn by observing human chefs in action, potentially even caching prior executions and modifying them for new similar recipes. Bon appétit, future robot chefs!