Andrew and Emanouil identified three conditions any genuine intelligence must satisfy. It needs to be efficient, able to learn from limited data without demanding vast compute or power. It needs to be stable, scaling in capability without its outputs fracturing under pressure. And it needs to be grounded in the physical world through its senses, starting with vision.

The dominant thesis in the industry was that scale solves intelligence. More parameters. More data. More compute. The result is systems that are fluent without being accurate, confident without being right, and prohibitively expensive to run in exchange for outputs that cannot be trusted where it counts.

The path forward was physics, not text. While the industry ran the compute arms race, we are building in a different direction entirely.