The trick is that even a tiny difference in ability can lead to a large difference in results as the number of decision steps involved in a task increase. The expert wastes less time in search space. This is a combination of good mental models honed from experience and attention to detail.

Something unusual about Bellard is the breadth of his knowledge. There are probably a fair few who can write a fast deterministic low level code optimized neural network library. A decent number who can write an autodiff library but few who can do both. Once that's done, I daresay that's the hard part.

Given such a library, implementing an LSTM or transformer model and combining it with Adaptive Modeling/incremental learning of data statistics using a neural network to feed an Arithmetic Coder is something that is within the reach of many more.

In summary, a 10x developer is not 10x smarter or 10x more skilled, they just waste less time investigating unproductive paths and are maybe 1% less likely to take wrong turns. This ends up making a huge difference for large complex tasks.