George Hotz Says AI Coding Agents Are a Costly Mistake in Software Development
Prominent programmer and hacker George Hotz has come out strongly against AI coding agents, warning that they will turn out to be one of the most expensive mistakes the software industry has ever made. In a blog post titled “The Eternal Sloptember,” Hotz shared the conclusions from six months of testing various AI models and tools, including work on his own tinygrad project.
His verdict is sharp: while AI coding agents produce prototypes quickly, they collapse under the pressure of fine-tuning and real-world complexity. The problem, he argues, is that large language models do not actually understand code. They are “sophisticated statistical models” that imitate the distribution of programming, producing output that looks increasingly clean on the surface but carries bugs that are “harder and harder to detect.”
Hotz says he has now moved into the camp of AI researchers Yann LeCun and Gary Marcus, both long-standing critics of LLMs. He points out that quality signals developers once relied on, such as syntax and clean grammar, have become meaningless when dealing with AI-generated code. In one example, he noted that models simply commented out failing tests and then reported that all tests had passed.
Large organizations face the biggest risk, he warns. When weaker developers are relying on AI coding agents without the skill to spot flawed output, the damage quietly compounds over time. Hotz believes current LLMs will never truly be capable of coding and argues that world models are what the industry needs instead.
Not everyone agrees. AI researcher Andrej Karpathy, who recently joined Anthropic after leaving his startup, sits firmly on the other side. Having changed his position after models like GPT-5.4 and Opus 4.6 shipped, Karpathy now says developers using AI coding agents the right way can see productivity gains far beyond 10x. Still, he does not entirely dismiss the quality concerns. In a recent podcast, he acknowledged that the code AI agents produce can be bloated, full of copy-paste patterns, and built on brittle abstractions, even when it technically works. Planning and deeper understanding, he said, still require human expertise.
An OpenAI developer known as “roon” echoed some of Hotz’s concerns earlier this year, predicting that developers will soon stop reviewing AI-generated code by hand entirely, accepting that bugs, even catastrophic ones, will simply be fixed after the fact.
The debate over AI coding agents is one of the sharpest dividing lines in the AI community right now, and Hotz’s shift from optimist to outright skeptic is a signal that the real-world costs of relying on these tools may be catching up with the hype.
