Probably aren’t seeing the promised productivity improvements of AI in terms of shipping production code and not just “super demos” that aren’t robust. So they want to see if the withers are really putting in the time or if the models struggle past a level of complexity that stalls or reverses early gains.