Limits of Current LLMs

Edan Meyer discusses the limitations of current large language models (LLMs) and the challenges of relying solely on in-context learning and retrieval-augmented generation (RAG). He explains that while in-context learning allows models to use relevant context to perform tasks, it has critical shortcomings. Firstly, the necessary context may not always be available, especially for niche or novel problems where documentation or references are lacking. Secondly, the scope of what a model can learn in context is limited by its pre-training data. For example, a model trained primarily on code may not perform well on tasks like poetry generation. Meyer emphasizes that in-context learning comprises various skills dependent on the model’s training data. He argues that to solve complex, cutting-edge problems, models need continual learning to adapt and learn new information beyond their initial training. This is why his startup continually trains models in production, despite the higher costs and slower process. Meyer concludes by inviting collaboration on his research project aimed at addressing these limitations.