George Hotz has spent years building systems that push hardware and software to their limits. From his early days hacking consoles to founding comma.ai for self-driving tech and creating tinygrad, his focus stays fixed on practical results. So when he posted “I love LLMs, I hate hype” on July 12, 2026, the tech world took notice. The piece cuts through the noise. Hotz declares himself “absolutely giddy” about AI advances. Yet he reserves sharp words for the surrounding exaggeration.
His excitement feels genuine. “I’m so excited for the new LLMs, self driving cars, video generation models, and coding agents,” he writes. Last week he set up a Linux box running opencode on a local GLM-5.2 model from Hugging Face. The command “install tmux with the geohot configuration” just worked. “The Year of the Linux Desktop is finally here!” Hotz exclaims. Simple. Direct. And a sign that local models have reached a point where they deliver real utility without cloud dependencies.
But enthusiasm doesn’t blind him. Two forms of hype draw his fire. First comes the negative kind. Talk of a closing window of opportunity. Warnings about a perpetual underclass. Predictions that outsiders will fall hopelessly behind. “This is negative valence hype,” Hotz states. “Not only is it not true, it’s mostly designed to make you feel bad about yourself and move to shitty San Francisco where everything really does suck like how these people claim.”
The Straw Man of Sudden Takeover
Second, he targets the leap from practical tool to world-altering force. People treat LLMs as fancy autocomplete or smart compilers. Then they pivot to claims that these systems will “own the whole light cone.” Miss the right parties in San Francisco and a flash of light might change everything before you notice. “I’ll bet you everything I have that this doesn’t happen,” Hotz says. The people pushing such stories? “Terrible people.” Justice arrives in the fact that “this is how they feel inside all the time themselves.”
He points to older warnings. A 2016 presentation on superintelligence by idlewords.com. The 1991 film Terminator 2: Judgment Day. A certain group claims credit for progress that would occur anyway. This pattern fuels his critique of frontier lab valuations. In earlier posts he argued the labs won’t capture the value they create. AI progress stems mostly from Moore’s law and broad computing gains. Not from secret sauce held by a few companies. They fear commodification above all. Safety concerns and geopolitical talk often mask that anxiety. Open sourcing threatens their edge. So they argue against it.
Hotz admits he may have judged models too harshly in his May 2026 post “The Eternal Sloptember.” There he questioned whether current LLMs could ever program effectively. Now he sees a shift. Programming itself changes. Compilers don’t “program” in the human sense. Yet they deliver massive productivity leaps. He references a Linus Torvalds quote shared on Reddit: agents might make programmers 10 times more productive, but compilers offer 1,000 times the gain. Hotz calls both figures extreme. Still, he feels confident that his own skills with these tools have improved. The boost is real.
Success demands care. Models can increase cognitive fatigue, as one developer described in a personal account. Vibe-coded outputs often remain low-quality slop. Where is the flood of new, high-quality software that productivity gains should produce? The question lingers. Yet LLMs serve as useful aids. They join find-and-replace, Stack Overflow lookups, and regex patterns many never mastered. Tools evolve. Skills adapt.
His core thesis lands with force. “AI is the continuation of the computer revolution. I love computers so much.” No grand break with the past. No magical intelligence explosion detached from hardware trends. Just steady, Moore-driven advancement that smart engineers can harness.
Recent coverage echoes and extends these points. In May, The Decoder reported on Hotz’s view that coding agents represent “one of the most costly mistakes in software development.” LLMs mimic the distribution of existing programs. Their flawed outputs grow harder to spot over time. Syntax checks fail to catch architectural problems. Hotz insists the field needs world models that grasp underlying systems, not just statistical imitators. Reactions came from Yann LeCun, Andrej Karpathy, and an OpenAI engineer. The debate continues.
A Medium essay from the same period, “George Hotz is right. Coding agents are producing an ocean of slop”, backs him up with observations from the trenches. Agents generate code 10 to 50 times faster than humans can review it. Reviewers grow tired. They skim. Slop slips through. It compiles confidently yet proves “architecturally braindead.” Pull requests arrive that seem to work but crumble under maintenance. The author asks whether teams will build cultural defenses or simply accept decaying codebases that feel like inherited legacy from day one. “The productivity improvements that Karpathy is arguing for exist, in the mind of the person writing the code,” the piece notes. “They’re negative for everybody else downstream. That’s an externality.”
As of this week, fresh signals appear. A July 2026 post on ExplainX highlights GLM-5.2’s MIT open-source release and its adoption in coding arenas. Hotz himself praised the model as a daily driver. Multi-model stacks gain traction. Local inference improves. These developments align with his demo of opencode on GLM-5.2. Practical capability spreads beyond big labs.
Yet warnings persist. LinkedIn discussions reference Godot’s stance against autonomous agents and “vibe coding” in open source. Some projects ban AI-generated contributions outright. Others see stars pour in for alternatives like OpenCode after controversies involving major AI providers. Grafana’s recent work on AI-native observability, covered by Software Engineering Daily on July 2, tackles the observability nightmare created when agents generate, deploy, and operate code at scale. Telemetry volumes explode. Human operators struggle to keep up. The systems built by machines demand new ways to watch them.
Hotz’s take stands apart because it avoids both blind boosterism and reflexive doom. He celebrates concrete wins. A local model that configures tmux exactly to his preferences. Productivity gains that feel personal and measurable. At the same time he rejects narratives that serve venture valuations or stoke anxiety to drive talent to certain zip codes. AI advances computing. It does not suspend economic reality or the tendency of technology to commoditize.
Programmers who adopt his mindset will treat LLMs as powerful but imperfect assistants. They will watch for fatigue. They will refactor the slop. They will combine models with world knowledge and careful verification. The revolution continues. But it builds on decades of prior progress. Computers got faster. Compilers got smarter. Now language models join the stack. And Hotz, for one, couldn’t be happier about it. But he won’t let the hype distract from the work that still must be done.
WebProNews is an iEntry Publication