LLMs Are Useful Without the Hype

George Hotz's “I love LLMs, I hate hype” is refreshing because it refuses the two positions that dominate too many AI arguments. He is excited about language models, coding agents, self-driving cars, and video generation, while rejecting the claim that everyone is racing toward one sudden, inevitable technological break.

That combination lands with me. I think LLMs are becoming genuinely useful tools, especially when they compress search, translation, setup, and repetitive coding work. I also think usefulness is routinely smuggled into much larger claims about inevitability, monopoly, and machines taking over everything. Those claims need their own evidence.

Answer Snapshot

Question	My read
What happened?	Hotz published a July 12 essay arguing that LLM progress is exciting and practically useful, while fear-driven deadlines and superintelligence narratives are hype.
What problem does it name?	AI discussion often bundles present-day utility with speculative claims about sudden discontinuity and durable control by a few frontier labs.
Who benefits from a clearer frame?	Developers, learners, open-model builders, and buyers who need to choose tools based on observable value rather than urgency.
My thesis	The strongest case for LLMs is that they extend the computer revolution. They do not need to be magical, monopolistic, or destiny-changing overnight to matter.

Utility Does Not Prove the Mythology

The most useful move in Hotz's essay is separating the product from the story sold around it. He describes getting value from a local model and coding agent for a Linux setup task, and he revises his earlier skepticism about whether models can program. His updated position is modest: programming is changing, models can provide a boost, and learning to use them is itself a skill.

I find that more credible than either “AI is useless autocomplete” or “AI will soon own the future.” A compiler can transform programming without becoming a programmer in the human sense. Search, find-and-replace, documentation, and Stack Overflow all changed what developers had to remember and how quickly they could move. An LLM can be another large step in that lineage without every science-fiction conclusion following from it.

The linked 2016 talk “Superintelligence: The Idea That Eats Smart People” is useful context because it decomposes an intelligence-explosion argument into premises. Some premises are plausible; the full chain remains an argument, not an observed event. That is the standard I want applied consistently: capability demonstrations are evidence for those capabilities, not automatic proof of every downstream scenario.

The Productivity Evidence Is Messier Than the Sales Pitch

Hotz says he now gets some boost from models, but he also points to cognitive fatigue and the continued abundance of low-quality generated software. Public research supports that mixed view better than a universal multiplier.

A PACIS 2026 experiment with 24 developers found that an agent reduced mean completion time and reported workload during brownfield onboarding tasks, but code correctness did not significantly improve. The authors also observed a shift toward passive supervision and raised concerns about over-reliance and skill erosion. That is an encouraging speed result with a large sample-size caveat, not a final verdict on software engineering.

An Anthropic randomized trial with 52 mostly junior engineers found a different tradeoff. Participants using AI finished only slightly faster, a difference that was not statistically significant, while scoring 17% lower on a quiz about the unfamiliar Python library they had just used. The study was small and tested immediate mastery, but its most practical finding was that usage style mattered: asking conceptual questions and checking understanding was associated with better learning than simply delegating the code.

Hype Manufactures a False Deadline

The critique I find most persuasive is Hotz's rejection of “window closing” messaging. Urgency is commercially convenient. It can turn an uncertain technology choice into a status contest: move to the right city, join the right lab, buy the right model, or accept permanent irrelevance.

That framing is especially weak when the same tools are spreading quickly. Hotz's own example is a local open model, not private access inside a frontier lab. His larger economic claim is that AI may create enormous value without frontier labs capturing all of it. The essay does not prove that outcome, and self-hosting still carries hardware, operations, security, and utilization costs. But open weights and falling deployment costs make permanent scarcity a claim to demonstrate, not an assumption to inherit.

This is where I would temper Hotz's argument too. General computing progress does not mean every model is interchangeable, and commoditization is rarely uniform. Reliability, distribution, data, integration, and support can remain valuable even when core capability spreads. Rejecting a monopoly narrative should not become a mirror-image certainty that frontier work has no durable advantage.

My Bottom Line

I like Hotz's essay because enthusiasm without submission is a healthier posture than reflexive belief or reflexive dismissal. LLMs can be impressive compilers, search systems, translators, and coding collaborators. They can also create fatigue, hide mistakes, weaken learning, and produce a flood of software whose value is hard to find. Those facts are not contradictory.

The standard I want is boring and demanding: show the task, the baseline, the failure rate, the review cost, and who keeps the skill to repair the result. Then let competition among hosted and open models push the useful parts into ordinary computing.

AI does not need a closing window or a flash in the sky to be important. If it keeps making computers more capable and more accessible, that is already a profound story. I would rather build around that observable progress than let the hype decide what the tools mean.