George Hotz Draws a Line Between LLMs and the Hype Around Them

George Hotz published a post titled "I love LLMs, I hate hype" on his personal blog in July 2026. The core argument: LLMs are useful tools, and the hype machine around them is a distinct and corrosive phenomenon.

This framing matters because it comes from someone with direct skin in the game. Hotz built tinygrad and comma.ai. He works with models at the systems level. His take is not a VC narrative or an accelerationist pitch. It is a practitioner drawing a line.

The distinction he is drawing is one engineers feel daily. The models themselves have compounding capability. The discourse around them — the claims about AGI timelines, the breathless product announcements, the benchmarks stripped of context — generates noise that makes it harder to evaluate what the tools actually do.

For builders, the practical implication is calibration. Overclaiming what a model can do leads to wrong architectural choices, wrong product scopes, and burned trust with users when the system fails at something that was never realistic in the first place. Underclaiming causes builders to avoid integrations that would ship real value.

Hotz has a track record of public positions that diverge from the consensus and then age reasonably well. He was skeptical of transformer scaling as the only path forward before that skepticism became common. His willingness to say something is good without endorsing the ecosystem around it is a useful signal.

The post does not argue against using LLMs. The argument is the opposite: the hype is the problem, not the technology. Builders who can separate those two things are in a better position to make sound technical decisions about where models belong in their stack and where they do not.

Source