I'm looking forward to it. Good topic.
I'm my limited experimentation, there seems to be a notable quality gap in the output and "helpfulness" between the open models and the commercial offerings. Given how well the open models benchmark nowadays (as you pointed out), I suspect that difference is something that can be drastically narrowed with "ecosystem" tooling and configuration -- like finding good harnesses, crafting good system prompts, input processors, output guidelines, and maybe figuring out which specific models come pre-baked with specific trainings (RLFH) pair best with that kind of last-mile configuration.
See you all on Tuesday!