In the last several years, the gap has been widening between the sensor fusion approach used in traditional ADAS versus the approach taken by L4 AV systems. Common reasons cited for this gap included limited computation power in ADAS, as well as ADAS being dominated by distributed computation architectures, versus the high-power, centralized architectures used in AV systems. These differences have led to the widespread adoption of Foundation Models, the algorithmic equivalent of a Swiss Army Knife, in the AV space, a technique that has not yet made its way to the ADAS world. In this presentation, we will share what we have learned from investigating bridging the gap between these two types of systems. These lessons span a wide breadth across the development process, from estimates of required data volumes at various stages, intermediate/hybrid types of network architectures, and the surprisingly high-quality results that can be achieved even with models that will fit in limited compute environments.