George Hotz—the hacker who first cracked the iPhone at age 17 and reverse-engineered the PlayStation 3 before Sony sued him—published a blog post Sunday arguing that mass adoption of AI coding agents will end in disaster, or something close to it.
"I'm calling it now, the adoption of AI agents into software development will be one of the most costly mistakes in the field's history," Hotz wrote in the post, titled "The Eternal Sloptember."
"Agents cannot program, and it's taking longer and longer to realize that they can't. The output is broken, but in a way that's getting harder and harder to detect. Which is exactly what you'd expect from an increasingly accurate statistical model."
The post arrived five days after Andrej Karpathy, one of AI's most prominent researchers, joined Anthropic's pre-training team with the explicit view that AI agents have already transformed software development. The two men now represent opposite poles of a debate the industry has yet to settle—and both carry genuine credibility to stake a position.
Hotz didn't reach his conclusion from the sidelines. He spent six months using agents on real projects: portions of Tinygrad, his open-source deep learning framework, and a complete firmware reverse-engineering of a USB-PCIe chip. "The agent frontloads all the progress," he writes, then hands you what he describes as a slot machine lever—you pull it and hope the finishing work gets done. It never quite does.
Hotz anticipated the obvious pushback: a programmer who defines part of his identity through his craft would naturally resist tools that threaten to replace him. He took the objection seriously and dismissed it on the merits.
"I thought more about the self worth preservation thing. Google's AFL found more bugs than LLMs and nobody felt that way about it. Chess and Go are more popular than ever," Hotz wrote. Chess AI has dominated human players for decades, and the game only grew more popular—so his concern isn't about being replaced.
Instead, it's about what happens to code quality when everyone uses these tools simultaneously, especially as major corporations and financial institutions push for their widespread adoption. "I almost think this is some kind of psyop to sell agents," Hotz argues. "Fear of loss is one of the only ways to make big companies move. Though I think in that fear they are making a big mistake."
His central argument is organizational. High performers have tight enough feedback loops to catch agent-generated problems before they ship. They read the code, spot the errors, and calibrate when to trust the tool. "The bottom performers won't have that self check," he writes—and they are the ones using agents to produce ten times their previous output. At a large company, that math produces something specific: faster degradation of average code quality, masked by sheer volume.
The outcome, he argues, will be "a golden era for buckets and buckets of slop, and a dark age for gems of quality." As a concrete example, he points to reports that Apple is pushing AI coding tools across its entire engineering organization, then asks simply: "Do you think macOS will get better or worse in the next 2 years?"
Hotz now places himself in what he calls the "LeCun/Marcus camp"—referring to Yann LeCun, Meta's chief AI scientist, and Gary Marcus, a longtime large language model skeptic. Both have argued that language models are fundamentally sophisticated pattern-matchers: they can imitate the distribution of existing code, but cannot reason through genuinely new problems from first principles.
Vibe coding—describing what you want in plain language and letting AI generate the implementation—has exploded over the past year, and the major labs have positioned agent-based coding as a flagship product. Microsoft transformed GitHub Copilot into a full agentic system in 2025, with CEO Satya Nadella describing it as a platform-level shift comparable to the move to cloud.
The pushback to Hotz's position isn't abstract. Karpathy, who had been skeptical of agents earlier in 2025, reversed his position after new model releases and joined Anthropic's pre-training team on May 19—five days before Hotz published. He described the next few years at the frontier as "especially formative." Anthropic CEO Dario Amodei said at Davos that some Anthropic engineers have already stopped writing code themselves, letting models handle it while they review the output.
Hotz, for his part, says he tried to do the same thing and found himself reaching for the manual fix every time.
Why it matters
Hotz's argument is specifically organizational, not technical: his concern is that high-volume, low-oversight use of agents at large institutions will degrade average code quality at scale, even if skilled engineers can manage the tools responsibly.
The Apple example in the article illustrates a concrete test case: a major engineering organization mandating AI coding tools across its workforce is a real-world experiment whose results will be visible in product quality over the next two years.
The Hotz–Karpathy split matters because both figures have hands-on, recent experience with the same tools—not theoretical positions—making this a rare public disagreement grounded in direct practice rather than ideology.
