Amara's Law: Too Much in Two Years, Too Little in Twenty
The same technology gets overestimated early and underestimated late. Both errors come from drawing a straight line through something that is slow first and rapid afterwards.
The same technology gets overestimated early and underestimated late. Both errors come from drawing a straight line through something that is slow first and rapid afterwards.
Chip design is spread across hundreds of firms. Making the advanced ones happens in very few places — and that single asymmetry explains the shortages, the politics, and the fragility.
Transformers won not because they were the cleverest idea, but because they were the most parallelizable one.
GPU datacenters look like a technology story. They're really a capital-intensity story with technology as the excuse.
The interesting number in AI is not what a model can do. It is what a unit of machine reasoning costs — because when that falls tenfold, whole categories of product become possible.
First post on the new blog.