But op's point was about people talking about tiny LLMs failing to consider the amount of parameters used. This has nothing to do with the ability of autonomously collecting the training data.