Large reasoning models almost certainly can think