A recalled January 2024 pitch on pretraining and inference scaling

The author recalls pitching Ethan Perez in January 2024 on studying the “exchange rate” between pretraining and inference scaling, based on an Andy Jones paper. He says the pitch came nine months before OpenAI released o1 and takes that timing as a lesson to pursue convictions about the future. In the same thread, he says showing universal jailbreaks seemed difficult, if not impossible, for multimodal inputs.