“In the early days of OpenAI, there was significant internal resistance to the proposal to dedicate the company's entire compute budget of 10,000 GPUs to training a single large model based on the scaling hypothesis.”