Researchers from around the world gathered virtually and in person for the third annual hackathon organized by Ben Blaiszik, a machine learning specialist at the University of Chicago and former materials scientist. Held in Berlin and across multiple global sites, the 48-hour event brought together over 1,200 participants who used large language models (LLMs) to tackle challenges in materials science and drug discovery. The competition, fueled by pizza and innovation, saw more than 100 teams submit 2-minute video presentations showcasing AI-driven projects ranging from predictive models for material behavior to tools for drug supply chain tracking. The hackathon served as a proving ground for customizing LLMs—such as OpenAI’s ChatGPT and Google’s T5—for scientific research. While these models excel at generating text based on vast datasets, they require fine-tuning to be useful in specialized domains. Teams worked to adapt LLMs to tasks like predicting how crystal structures relax to their lowest energy states, a key step in understanding material properties. One team, including a physics PhD student with no prior LLM experience, used AI to help write code for data processing, completing in two days what would have taken weeks. The student now sees AI as a valuable collaborator rather than a threat to scientific creativity. Other teams developed AI agents to solve real-world problems. One created a system that tracks the origin of ingredients in pharmaceuticals and materials, enabling faster recalls if safety issues arise. Another built a “co-pilot” chatbot that uses existing chemistry AI tools to generate research hypotheses. These projects highlight the growing need for tailored data pipelines and domain-specific training to turn general-purpose models into reliable research assistants. Central to many efforts was NOMAD, a decade-long project led by Humboldt University’s Pepe Márquez and others to create the world’s largest open repository of materials science data, housing over 19 million entries. Several teams designed AI tools to help scientists navigate NOMAD or automate the upload of experimental results into the database. Ana Velázquez, a data steward at Helmholtz-Zentrum Berlin, emphasized the importance of making data workflows seamless. “Once we have everything set up, they just drag and drop a file, and everything is like magic,” she said. With 32 judges from academia and industry reviewing submissions, the event aims to identify promising prototypes for further development. Blaiszik plans to connect top teams with mentors and investors to help turn ideas into real-world tools. But beyond the prizes and products, the hackathon underscored a vital lesson: collaboration. As software developer Kutlualp Tazefidan noted, “The danger of AI is it allows us to do so much without ever collaborating.” When a teammate struggled after a sugar crash, Tazefidan simply said, “I’ll just ask ChatGPT.” The moment captured the event’s spirit—AI as a partner, not a replacement, in the pursuit of science.