Building AI chips is a complex endeavor that requires anticipating future workloads, which is notoriously difficult due to the rapid pace of change in the field. By focusing on breaking down workloads into fundamental primitives and applying inversion as a mental model, clarity can be gained on what features to prioritize for long-lasting chip design. Interestingly, some elements, like linear algebra operations, are likely to remain constant, providing a foundation for future development.
