This essay, written in October 2024, laid the intellectual groundwork for what became Jean Technologies. It argues that the real breakthrough in AI was deep representation learning, and that the logical endpoint is trusted infrastructure for embedding people.
The argument: as models scale, vector embeddings can represent any concept. Those embeddings become interpretable. And the most valuable thing to embed, ultimately, is not documents or images but people: their goals, their working style, the latent traits that predict who they will succeed with and what they will build.
Two years later, Anthropic's interpretability team published evidence that LLMs already maintain causally operative representations of emotional and behavioral concepts. The infrastructure thesis is moving from speculative to specific.