The Deepsky reaction to DeepSeek is weird. Noone who has worked with LLMs is remotely surprised when someone comes up with a more cost-efficient way of training these models. First build ginormous LLMs and then try to match their functionality with fewer resources. This is how progress is made.