I think the bottleneck is testing. I want to build a replacement for Zwift, a virtual gym game for bike trainers and treadmills, but testing it could be difficult without a real person on real hardware. How does the LLM know about the hardware protocols and stuff like that.

Same way you’d do it without AI. Record sample data, test against that, generate more data, test IRL, record more data, loop until it’s good enough.

I don't have one. how through is this blog post reverse engineering it? https://www.makinolo.com/blog/2024/07/26/zwift-ride-protocol... ?