You're massively overestimating the skills needed to build "competitive models". Researchers like Ilya were essential several years ago when OpenAI got started, because at that time nobody even knew what to build or how to go about it. It was all research, and OpenAI was solely focused on RL. But these days, LLM code is 95%+ data and infrastructure engineering and very little novel research. Even things like LLM speedups via new attention mechanisms are more engineering than they are research.
Above all else, what you need to build competitive models is a huge amount of money and access to compute. Money alone doesn't solve that either due to the global hardware shortage. You are competing for a globally limited pool of hardware with all other companies.