Adapting Foundation Models to Any Budget (Reuse & Retrieve, Don't Retrain or Recompute)
The cost of foundation models does not end with pretraining. It continues through adaptation, inference, and real-world deployment. Today, the default is to pay much of that cost again for every new task, query, and environment. This talk explores how to make models as efficient as possible after pretraining, from adaptation to deployment. The key idea is simple: avoid repeating expensive computation when we can reuse what has already been built. We apply this principle across three parts of the stack. We adapt models by combining existing capabilities instead of repeatedly training from scratch. We speed up inference by compressing reusable context once and serving it across future queries. We deploy new capabilities through retrieval while keeping the core model fixed. Across all three, the goal is the same: build once, reuse broadly, and scale to different budgets without rebuilding the system each time.