Additional, Fan mentioned that OpenAI should have found out the inference scaling regulation a very long time in the past, which academia is only in the near past discovering. Nevertheless, he did level out that productionizing o1 is far more durable than nailing the tutorial benchmarks and raised a number of questions.
“For reasoning issues within the wild, how (the mannequin) to determine when to cease looking out? What’s the reward perform? Success criterion? When to name instruments like code interpreter within the loop? The way to issue within the compute value of these CPU processes? Their analysis put up didn’t share a lot.
OpenAI, too, in one of many weblog posts has mentioned that the brand new mannequin, which continues to be within the early levels of improvement and is predicted to bear vital iteration, doesn’t but have lots of the options that make ChatGPT helpful, resembling shopping the net for data and importing information and pictures.
