I can’t see how any of the current LLM-based ecosystem will survive once the money circle-jerk ends. And now there a huge bunch of pre-trained models out there anyone will be able to run them on their own infra (GPUs will be cheap as chips once the bubble bursts). It’s not like most corporate users make use of the “knowledge” part of the models - there’s far more call for performing classification and summarisation tasks than “who won the Mexico ‘86 World Cup?” general knowledge questions.