Comment by anuramat

"some model I don't get to use is much better at benchmarks"

pick one or more: comically huge model, test time scaling at 10e12W, benchmark overfit