TL;DR the article concludes that the only real viable way to training frontier models as your primary business is to provide agents built on top of them — everything else is a race to the bottom as tokens and intelligence become commoditized.<p>I'm curious what secret sauce the author thinks is going on with agentic development flows where one of these labs is going to be capable of delivering significantly better agents than their competitors. As I understand it, agents are prompts with control structures and memory. I guess that's the opportunity — come up with something gigabrain? As in literally AGI?<p>And somehow make the models good enough and accessible enough through your agentic workflows that your API clients don't compete with you by offering similar functionality?