My theory with no insider information: it’s a little of all of the above, but mostly money. To some extent, you can dig yourself out of a data hole with RL and a lot of compute. And you can buy a lot of compute and some data with a lot of money. Big labs have been operating in this regime for a while and it’s one of the drivers behind their costs beyond just scaling the weights and doing the actual training. Mistral just doesn’t have access to this level of compute or the money to try and muscle their way in.