The most productive stance on large language models right now is probably this: treat them as powerful but narrow tools, and ignore almost everything written about their broader implications. That appears to be the core argument in a widely-shared post by George Hotz, which drew over 400 upvotes and 255 comments on Hacker News — numbers that suggest it hit a nerve with working engineers.
Hotz, who built tinygrad and leads comma.ai, occupies an unusual position in this conversation. He uses LLMs heavily in real codebases, which gives his skepticism about the surrounding narrative more weight than the typical contrarian take. The distinction he draws — between the technology itself and the claims made on its behalf — is one that practitioners have been quietly making for a while, but rarely state this plainly.
The practical implication for builders is straightforward: you can extract significant value from LLMs for code generation, debugging, documentation, and rapid prototyping without buying into any particular story about where the technology is heading. Capability and narrative are separate questions, and conflating them leads to both over-investment in wrong bets and under-investment in what actually works today.
The Hacker News discussion is worth skimming for the engineering-specific pushback and use-case breakdowns. When a post like this generates 255 comments from a technical audience, the thread often contains more signal than the original piece — particularly around where LLMs are failing quietly in production environments.
Bottom line: calibrate your tooling decisions on observed performance, not on fundraising announcements or benchmark press releases. Hotz's framing — love the tool, distrust the hype cycle — is a reasonable operating posture for anyone building with these systems right now.