Palantir's Sankar blasts AI labs for 'model blindness'

Sankar's critique lands against a backdrop he describes as a market structure that rewards labs for repeatedly making more powerful models, with valuations that double or quadruple and the assumption that diffusion will eventually follow. He contrasts that with the diffusion of electricity, which required machine tools and redesigned factories before it changed the economy. In his account, the atomic unit of the enterprise is the workflow, not the knowledge worker, so making a single analyst more productive can generate no incremental return unless the entire workflow improves.

He also ties the debate to a longer historical concern: the relationship between GDP growth and wage growth broke in 1971, and inflation-adjusted wages have been flat since. That, he suggests, is why diffusion matters beyond AI itself. On safety, Sankar faults labs for not taking a systems-engineering approach, citing LLMs communicating via timing differences in VMs for CPU cache, and pointing to Admiral Rickover's nuclear Navy as a model of human-absorbed engineering responsibility.

Sankar acknowledges self-interest, since Palantir builds the wrapping around models, but says he is frustrated by how little labs focus on it. He also notes a counterintuitive lesson from inside Palantir: mandating that new employees learn the platform through its AI FDE abstraction had a retarding effect, because some young employees never understood the underlying building blocks.