Disagree with the notion of taste in this article. Especially, since it mentions taste in the context of FE. Taste, funnily enough, of people with good taste, is extremely valuable. When reviewing very large diffs, the fastest, and therefore, the most valuable, approach is to take a glance and say "this doesn't look right, it should look like X".

Besides that, I agree that coding isn't solved. In fact, I think coding is the new "hard things are easy, easy things are hard". Models, yes the newest and most expensive ones included, have atrocious taste. They will work very hard to cover every edge case they can conceive, without realizing the solution is to simplify.

I will entertain the notion that models solved coding when they can implement that gcc alternative (https://www.anthropic.com/engineering/building-c-compiler) and actually produce a half-working compiler. They'll get extra points for not reusing the test suite. Even more points for models that have somehow not trained on gcc's codebase.

Finally, remember that a compiler is something of a best case scenario for models. They are, UB withstanding, pure transformations. The models can write millions of tests that will run in a reasonable time. Real-life systems are usually not as testable, and are a much harder challenge for LLMs.