Unless you use anything that does JIT compilation (e.g. JVM), e.g. JetBrains IDEs are really slow onder Rosetta.
Another pain point is code that relies heavily on SIMD. AVX/AVX2/AVX-512 instructions are not translated, so in those cases you'll be using slower kernels. My machine learned models are ~10 times slower with libtorch under Rosetta. Though they are competitive when compiling libtorch natively.
(Source: I have used an M1 Air for a week, but decided to return it.)
