I honestly don't get why companies need to train on prompts any more? Surely they've got PLENTY of training material, and the advancements these days come from architectural improvements?
Not at all. As companies refine their models to perform better at specific tasks, they need more data to train on that specific task. I know first hand that the frontier labs are reaching out to all kinds of companies asking to license their software to run so they can train on it. The stated reason is not to reproduce that tool, but to get better at operating it, though history tells us that reproduction will naturally follow even if unintended.
We have talked to many customers who specifically wonder if our software could learn to perform better for them by learning from their interactions - they are suggesting we train on their work! This tuned model would only be allowed within their enterprise, of course. It's understood that the models know how to do this generally, but every situation is different, and results could be better if taught how these specific situations work.
We have talked to many customers who specifically wonder if our software could learn to perform better for them by learning from their interactions - they are suggesting we train on their work! This tuned model would only be allowed within their enterprise, of course. It's understood that the models know how to do this generally, but every situation is different, and results could be better if taught how these specific situations work.