« I love LLMs, I hate hype » - geohot reminds the only rule that remains

Separate what models actually do from what is attributed to them. Discipline is lacking more than compute power.

The fact

George Hotz (geohot) published on July 12, 2026, "I love LLMs, I hate hype": a short, sharp post that distinguishes what the models do (compress, restore, interpolate) from what marketing attributes to them (reason, plan, understand).

Our reading

Geohot isn't inventing anything - he's doing the sorting that no one wants to do. In 2026, most "thinking agent" demos are demos of well-connected harnesses, not emergent reasoning. This doesn't make LLMs useless: it makes them usable, provided you know what you're connecting.

Under the hood

The argument boils down to three points: (1) LLMs excel at compressible tasks (code, summarization, translation) where the pattern is in the training set; (2) they fail outside of support (long-horizon planning, strict formal coherence, exact calculations); (3) the "reasoning progress" observed comes from the harness (test-time compute, RAG, tools), not the bare model.

So what

Practitioner: the best ROI remains on compressible tasks + tools. Decision-maker: always ask what part of the gain comes from the model and what part from the harness. The distinction determines whether you are exposed to the next model change.

To watch

Benchmarks that isolate model and harness. The next post from Lambert or Karpathy in response.