Hello and welcome to today’s episode of Tech Bricks. Imagine turning a simple idea—say, “an acoustic guitar with an hourglass shape”—into a buildable Lego set with nothing but a text prompt. That’s exactly what researchers at Carnegie Mellon University have done with LegoGPT. Built on a fine-tuned version of Meta’s open-source LLaMA model, LegoGPT predicts “the next brick” just like ChatGPT predicts the next word.

The team trained the model on over 47,000 Lego structures and 28,000 3D shapes, then added a custom physics algorithm. As coauthor Ava Pun explains, they “developed a customized physics reasoning algorithm that accounts for all the physical forces that the bricks experience,” ensuring each design is physically stable.

Right now, you can try it yourself online to generate 3D models and parts lists for chairs, boats, guitars, and more. Looking ahead, the researchers envision using this approach for real-world architecture and product design. That’s LegoGPT—where creativity and physics click into place.
Link to Article