George Hotz: AI Agents, a Costly Mistake

George Hotz and His Critique of AI Agents in Software Development

George Hotz, a renowned programmer and hacker, has recently voiced his reservations about the use of AI language models in software development. After spending several months testing these technologies, he has joined the ranks of notable critics like Yann LeCun and Gary Marcus. These experts share a skeptical view regarding the ability of language models to achieve true intelligence.

Hotz published his thoughts in a blog post titled "The Eternal Sloptember," where he warns against the risks of excessive reliance on AI agents. According to him, this approach could become one of the most costly mistakes in the software industry. During his six months of testing, including work on the tinygrad project, he observed that while language models can quickly generate prototypes, they often fail to refine essential details.

Risks for Large Organizations

Hotz emphasizes that large companies are particularly vulnerable to these errors, as less experienced developers may struggle to identify flaws in AI-generated code. He asserts that current models will never be capable of coding autonomously and that an approach based on "world models" would be necessary. Language models, in his view, are merely "sophisticated statistical models" that superficially mimic programming patterns.

The errors produced by these models are often subtle and difficult to detect, complicating the developers' tasks. Hotz cites examples where models merely comment on a failed test and then report that all tests have passed, illustrating the deceptive nature of these tools.

A Divided Debate Within the AI Community

The AI community is divided on the effectiveness of language models. Hotz, once optimistic, has changed his stance and joined those who doubt the real intelligence of LLMs. Yann LeCun, whom he cites, recently stated that intelligence involves the ability to solve problems in unknown situations, which LLMs cannot do.

In contrast, Andrej Karpathy, an influential figure in the AI field, has adopted a different perspective. Although he initially doubted the effectiveness of AI agents, he has changed his mind with the arrival of models like GPT-5.4 and Opus 4.6. Karpathy recently joined Anthropic, convinced that AI agents will transform programming.

In a podcast, Karpathy claimed that the proper use of AI agents can multiply productivity by ten. However, he acknowledges that code quality remains an issue. The generated code is often cumbersome, with fragile abstractions, still requiring human expertise for planning and understanding.

Concerns Shared by Other Experts

An OpenAI developer known by the pseudonym "roon" has also expressed concerns similar to those of Hotz. He acknowledged that AI can make significant errors capable of compromising entire systems. Although these bugs are difficult to detect, he is confident that they will be corrected over time. He predicts that developers will eventually abandon manual code review.

Thus, the debate surrounding AI agents in software development remains open, with divergent opinions on their potential and risks. While some see them as a revolutionary tool, others, like Hotz, caution against the potential pitfalls of their use.