George Hotz on LLMs: Separating Capability from Narrative

George Hotz posted a characteristically blunt take on the current state of LLMs. The thesis is simple: the models themselves are genuinely useful; the discourse around them is not.

The distinction matters for builders. Hype inflates expectations in ways that damage real adoption. Engineers who believe the marketing copy build systems against capabilities that do not exist, then blame the tools when those systems fail. Hotz draws the line between the artifact and the narrative surrounding it.

This is a consistent position for him. He has shipped production AI systems, driven real hardware at inference time, and engaged with model behavior at a low level. When someone with that background says the technology is good but the hype is bad, it carries different weight than the same claim from an analyst.

For solo founders and senior engineers, the practical implication is: benchmark against what the model actually does in your specific context. Ignore capability claims made in press releases. Run evals on your own data. The gap between benchmark performance and production behavior remains large, and most hype lives in that gap.

The post also signals something about the broader technical culture. A vocal segment of the builder community is pushing back on inflated narratives, not because they are pessimistic about AI, but because they want accurate models of what these systems can and cannot do. That pushback is healthy. Systems get built better when engineers work from accurate priors.

The full post is on Hotz's personal blog and is worth reading directly. It does not offer a framework or a listicle. It offers a perspective from someone who builds, and that is the relevant signal here.

Source