I think the biggest potential area for algorithm-driven improvement isn't so much coming up with clever ways to trace rays faster, but rather coming up with way

I think the biggest potential area for algorithm-driven improvement isn't so much coming up with clever ways to trace rays faster, but rather coming up with ways to generate better quality output within a given rays-per-second budget. Machine-learning-based denoising algorithms are a good example of that.