Yup, and also notable how China proceeded to invest in energy infrastructure before trying to scale out compute. The Americans put the cart before the horse here.

I don't think there's anything that fundamentally prevents AGI from being developed, but I strongly suspect that it's going to take more than LLMs to do that. In fact, frontier is already moving towards world models where a physics simulation becomes the basis for reasoning. Deepseek gunning for AGI is not in any way at odds with the current methods hitting a plateau. We already know that simply making models bigger doesn't scale. It's also unlikely that feeding more data into models is going to help either. The problem is with how the architecture works, and making any substantial leaps from here will require rethinking that.

And putting AI in charge of R&D is tangential to improvement of models themselves. Current capabilities are already very useful, and when you put a model in a feedback loop, it can indeed work autonomously and produce useful results. So, current crop of models will absolutely lead to a ton of breakthroughs and new types of automation that weren't possible before. China is also very well positioned to take advantage of all that being the factory of the world.

The question is what it will take to make models themselves significantly more capable though. Look at the progress that was being made like two to three years ago. Every few months we'd see some incredible breakthrough in capability. That's basically stopped in the past year, and progress has become very incremental now.