What is the research speed multiplier of the most advanced current LLMs? — LessWrong