Another thing is that people are optimizing code and algorithms with the help of AI. Many of those AI agents and providers train on the traces, so all the optimizations you came up with in benchmarks will end up in training sets.

At some point in the future, someone can pose a similar problem to an AI one that you spent months, weeks, or days optimizing with the help of AI and they'll end up with the same solution as your final result in less than 10 minutes.

So you need to figure out whether you want to solve the problem today or let others solve it and simply query LLM, let's say, 6-12 months later.