Google and Meta may be the companies with the most detailed personal information. Yet they don’t manage to have the best model. As others have pointed out, the

Google and Meta may be the companies with the most detailed personal information. Yet they don’t manage to have the best model. As others have pointed out, the secret for the NN may be to just have as much parameters as possible. So if I had a ton of money to train a LLM, I'd spend that money on curating a corpus of text. First priority would be books and papers out there. Then, maybe quality information on the internet like Wikipedia, Stackoverflow, source code, documentation. After all, the low-level functions like grammar, sentence structure, word use can be learned from any good text. For higher-level stuff, I would want quality input in there. So it learns reasoning as scientists do in papers rather than some politically biased person on Twitter. (We've seem some old emails of CEOs posted online. Something like this could be of value because they discuss strategy. But these are the exception. I guess even for interesting public figures, most of the conversations wouldn’t help much. It would also put OpenAI in danger of leaking sensitive information.)