Efficient Intelligence Lab

Vision

Our work centers on a fundamental question:

In recent years, progress in AI has mostly come from scaling two things together: compute and data. We throw more FLOPs at larger models, run them on increasingly large GPU clusters, and feed them ever bigger corpora of text, code, and interaction data. It works, but it comes with costs that are inaccessible to most researchers and institutions. Open-source systems like Kimi and Qwen now approach the quality of proprietary systems such as GPT-5 and Claude, but running them still assumes expensive, power-hungry multi-GPU servers that most labs, products, and devices simply cannot justify. At the same time, the models we call “small” are often still at the billion-parameter scale, trained on massive curated datasets, and they remain far weaker.

The Efficient Intelligence Lab focuses on the sufficiency frontier: how much model, data, and compute are actually needed for different kinds of reasoning, and how those resources can be traded off. The human brain runs on tens of watts—roughly a single light bulb—and learns from a limited stream of experience and reading, while typical large-model servers draw thousands of watts when active, comparable to powering several homes or fast-charging an electric car, and are trained on orders of magnitude more data than any individual will ever encounter. This discrepancy isn’t just a curiosity; it suggests a genuine opportunity for a step change in how we design and train reasoning systems. Our goal is to develop principles that let smaller models learn more from less and retain much of the reasoning power of today’s largest systems, while operating within the power, hardware, and data budgets that the real world can actually afford.