George Hotz, a world-renowned hacker, security researcher, and founder of Comma.ai, recently published a blog post stating that over-relying on AI programming agents in modern software development processes may be one of the most costly mistakes for companies. This view has immediately sparked widespread discussion and reflection within the tech community.

Statistical Imitation Is Not True Intelligence, Code Flaws Are More Hidden

Called "the amazing kid," Hotz said that after six months of deep testing, he found that current AI programming tools are very suitable for quickly building project prototypes, but they are not good at dealing with the underlying details that determine software quality. He emphasized that large language models are essentially complex statistical systems, their core task is to imitate the distribution of programming code, not to truly understand and solve problems.

Because of this, AI agents can generate code that looks impressive on the surface, but this makes hidden logical flaws and defects more difficult to identify. Junior developers who lack sufficient ability may fail to detect these subtle hardware and software issues, and blindly integrate flawed code into the official system, eventually leading to extremely high maintenance costs and unpredictable failure risks later on.

Industry Giants Show Divergent Attitudes, Cognitive Conflicts Intensify

Hotz admitted that his technical judgment has shifted from early optimism to skepticism, and his current position is closer to Yann LeCun and Gary Marcus. This pessimistic voice regarding AI programming agents contrasts sharply with the aggressive attitudes of some leaders in the technology sector.

In contrast, Andrej Karpathy, former head of AI at Tesla, has recently taken a more positive stance. Karpathy openly stated that with the release of new generations of large models, AI agents have permanently changed the way humans program. This indicates that there is a significant cognitive divide in the technology industry regarding the future of AI-assisted development.