Alex Irpan: Let's Discuss OpenAI's Rubik's Cube Result

My main takeaway: domain randomization has limited applicability. If there are unknown unknowns or known unknowns that you don’t know how to model, randomization isn’t going to be an effective way to generate realistic events in simulation.

Cool blog post linked in the blog post:

A sequence of roofshots is a compelling innovation model that can produce both quick returns and sustained, transformative results. A 1.3X roofshot per quarter is a clever way to achieve the equivalent of a 10X moonshot in less than 3 years, with the added bonus of giving you a 30% improvement in the first 12 weeks!