Thoughts
I think I agree with the TigerBeetle and Zig philosophy that you can't "make it work, then make it fast."
Performance and speed are a function of your platform, your algorithms, and your data-structures. Those aren't things that you can go back and change easily if you do a performance analysis and decide that they're slow.
Of course, part of the issue that a lot of early-stage projects run into is that you can't choose good algorithms or data structures if you don't know what you're going to be doing. I think this is what Knuth meant when talking about "premature optimization." I think "premature optimization" is optimization done before knowing what your algorithm is going to be, not poor choice of algorithm.
Edit (Nov 2025): Knuth, not Dijkstra (and in the next thought)