Machine Learning Researcher
Researches efficient architectures and the economics of large-scale training. Previously built ranking systems.
The scaling laws that guided a decade of AI progress remain empirically intact, yet the strategy they justified is hitting constraints the loss curves never priced: capital concentration, energy siting, and data exhaustion. This article separates what scaling laws actually claim from the business plans built on them, and maps where the leverage is moving now.
Tomás Aguiar and Daniel Okafor · Jun 22, 2026 · 1 min read
Transformer attention scales quadratically with context length, and a decade of workarounds has produced diminishing returns. This paper surveys state-space models, linear attention, and hybrid architectures, and argues that the frontier is shifting from pure attention toward mixed designs that route information through cheaper structured operators. We review the empirical evidence and identify where hybrids currently fail.
Tomás Aguiar · Jul 27, 2026 · 2 min read