Compromise between scale and cost for LLMs

Room A008

This talk will review the state-of-the-art about LLM scaling laws, both for training and inference, and will contrast this requirement to scale to achieve the best performance with the computational cost constraints that must also be considered in practice. Several research tracks and ideas that are currently being investigated in our group to reduce these costs will be briefly presented.

The talk will not be technical but will stay at the conceptual level, hopefully triggering future more focused technical talks.