←back to thread

GPT-5.2

(openai.com)
1019 points atgctg | 1 comments | | HN request time: 0s | source
Show context
SkyPuncher ◴[] No.46235977[source]
Given the price increase and speculation that GPT 5 is a MoE model, I'm wondering if they're simply "turning up the good stuff" without making significant changes under the hood.
replies(2): >>46235986 #>>46236012 #
minimaxir ◴[] No.46236012[source]
I'm not sure why being a MoE model would allow OpenAI to "turn up the good stuff". You can't just increase the number of E without training it as such.
replies(2): >>46236953 #>>46236981 #
1. yberreby ◴[] No.46236953[source]
Based on what works elsewhere in deep learning, I see no reason why you couldn't train once with a randomized number of experts, then set that number during inference based on your desired compute-accuracy tradeoff. I would expect that this has been done in the literature already.