Teaching an LLM a Little Jiu Jitsu (maxdemarzi.com)

Researchers at Relational AI have developed a large language model (LLM) specifically for Brazilian Jiu Jitsu (BJJ), aiming to assist beginners by providing accurate answers to their questions without frustration. The project leverages a vast library of resources, including over 30,000 instructional video segments, 14,000 competition clips, and various written guides. By training the model to understand not only the terminology but also context and physical mechanics demonstrated in videos, this AI coach can deliver precise, relevant responses tailored to the user’s inquiries. The significance of this project lies in its innovative application of advanced AI techniques to niche domains. The model employs a sophisticated Knowledge Graph to map synonyms and categorize terminology across different schools of thought within BJJ, ensuring comprehensive understanding. Additionally, it incorporates multimodal visual search capabilities for better context retrieval from silent video demonstrations. By utilizing fine-tuned cross-encoder re-rankers, the model ensures that responses are not only contextually aware but also computationally efficient, making it a pioneering example of how LLMs can be customized for domain-specific educational tools without incurring high operational costs.