The speaker talks about the use of compiler intrinsics in Sickle in DPC++ for writing single source heterogeneous programming API, which can run on different GPUs, such as NVIDIA, Intel, AMD, etc. They explain the process of converting Sickle cause to native implementations and how they use libclc for this purpose. However, they also discuss the need to redirect the process to use intrinsics for offload targets and how it can benefit from upstreaming the SQL implementation. The speaker calls for discussion about resolving the external dependency on the BC library and using intrinsics during the IR level pipeline. Finally, there are questions about MLIR and performance comparison with NVIDIA's closed-source PTX generator.