Geohot warns AI agents will trigger costly software degradation
In a recent blog post, Hotz contends that while AI aids rapid prototyping, its adoption in professional development will degrade output quality, particularly within organisations with slower feedback loops.
George Hotz, known in the tech community as Geohot, has published a critical analysis of artificial intelligence in software engineering, arguing that the current industry push towards AI agents represents a significant regression in code quality. In a blog post titled 'The Eternal Sloptember', Hotz contends that existing models are merely sophisticated statistical mimics rather than true programmers, producing broken output that is increasingly difficult to detect.
Hotz draws on six months of personal experience attempting to utilise AI for his 'tinygrad' project and hardware reverse-engineering. He concludes that manual execution was faster and yielded better results, noting that agents often frontload progress but fail to deliver polished, functional code. He asserts that while AI serves as a useful search tool or for quick prototypes, it falls short of the professional engineering standards required by major companies.
The analysis highlights a divergence in outcomes between high-performing individuals and large organisations. Hotz observes that skilled engineers maintain rigorous error correction and understand when to trust AI tools. In contrast, he warns that large enterprises with slower feedback loops and lower-performing staff are producing vast quantities of low-quality code, or 'slop', which degrades the average output quality of the organisation.
He references Apple’s internal push for AI as a concrete example of potential organisational degradation, questioning whether macOS will improve or worsen over the next two years as a result. Hotz argues that AI-produced artifacts lack the human state of mind behind their creation, making them prone to subtle breaks that old proxies of quality, such as syntax, cannot detect.
Aligning his views with researchers Yann LeCun and Gary Marcus, Hotz asserts that current reinforcement learning methods are insufficient for programming tasks. He concludes that viable programming agents require 'world models' rather than the reinforcement learning techniques currently in use, warning that the 'AI psychosis' driving corporate adoption will ultimately harm large organisations more than high-performing individuals.
