short
- George Hotz, the hacker behind the first iPhone and PlayStation 3 jailbreaks, published a blog post on Sunday calling the adoption of an AI encryption proxy “one of the most costly mistakes in the history of the field.”
- His basic argument is that high performers can detect bad worker output, but weaker engineers cannot — and weaker engineers produce ten times the volume, degrading average code quality on a massive scale.
- This post arrived five days after Andrei Karpathy joined Anthropic’s pre-training team with the opposite view, representing a clear division among serious engineers about whether AI agents actually work.
George Hotz — the hacker who first hacked an iPhone when he was 17 and reverse-engineered the PlayStation 3 before Sony sued him over it — published a blog post on Sunday arguing that the mass adoption of AI crypto agents will end in disaster, or at least close to it.
“I call it now, adopting AI agents in software development would be one of the most costly mistakes in the history of the field,” Hotz wrote. “Agents can’t code, and it takes longer and longer to realize they can’t.”
“The outputs are broken, but somehow they are becoming more difficult to detect. Which is exactly what you would expect from an increasingly accurate statistical model.”
The post is titled “eternal decline,“Arrives five days after Andrei Karpathy, one of the most prominent researchers in artificial intelligence, Join the pre-training team at Anthropic With a clear view that AI agents have already transformed software development. The two men now represent opposite poles in an unresolved industry debate, and both have the actual credibility to take a stand.
Personal update: I have joined Anthropic. I think the next few years on the frontier of LLMs will be particularly formative. I’m very excited to join the team here and get back into R&D. I remain very passionate about education and plan to resume my work in it in due course.
– Andrei Karpathy (@karpathy) May 19, 2026
Hotz didn’t come to his conclusion from the sidelines. He spent six months using agents on real projects: parts of TinigradAnd its open source deep learning framework and complete reverse engineering of the USB-PCIe chip firmware. “The dealer front-loads all the progress,” he writes, “then hands you what he describes as a slot machine lever—you pull it and hope the final action is done.”
It never happens.
Not about ego
Hotz predicts the obvious reaction: A programmer who defines part of his identity through his craft will naturally resist tools that threaten to replace him. He takes the objection seriously and rejects it on objective grounds.
“I’ve been thinking more about the topic of maintaining self-worth. Google NFL I’ve found more errors than LLM holders and no one feels that way about it. “Chess and the game of Go are more popular than ever,” Hotz wrote. “He is right that chess AI has dominated humans for decades and that the game is becoming more popular.”
So, his concern is not about being replaced. This is about what happens to code quality when everyone is using these tools simultaneously, especially when Big Tech and Wall Street are constantly pushing for universal use of these tools.
“I think it’s kind of the psychological process of selling dealerships,” Hotez says. “Fear of loss is one of the only ways to get big companies to act. Although I think because of this fear they are making a big mistake.”
His central argument is organizational. High performers have feedback loops tight enough to detect dealer-created issues before shipping. They read the code, detect errors, and calibrate when to trust the tool. “Low performers who use agents to produce 10 times their previous output will not do this self-examination,” he writes. In a large company, this mathematics produces something specific: a faster degradation of average code quality, masked by sheer volume.
He believes the result will be “a golden age for buckets and waste buckets, and a dark age for high-quality gemstones.” As a concrete example, he points to reports that Apple is rolling out AI coding tools across its entire engineering organization, then simply asks: “Do you think macOS will get better or better in the next two years?”
Where the camps stand
Hotz now places himself in what he calls the “LeCun/Marcus camp,” referring to Yann LeCun, the chief AI scientist at Meta, and Gary Marcus, a longtime LLM skeptic. They both argued that linguistic models are essentially pattern matchers: they can imitate the distribution of existing code, but they cannot reason through truly new problems from first principles.
Phoebe encoding– Describing what you want in plain language and letting AI generate the execution – has evolved dramatically over the past year, and major labs have established agent-based programming as a mainstream product. Microsoft I turned GitHub Copilot will move to a full proxy system in 2025, with CEO Satya Nadella describing it as a platform-level shift comparable to moving to the cloud.
The response to Hotz’s position is not abstract. Karpathy, who had been skeptical of clients earlier in 2025, reversed his position after the new model releases and joined Anthropic’s pre-training team on May 19 — five days before Hotz went live. He described the next few years on the border as “particularly formative.”
Some Anthropic engineers have already stopped writing code themselves and let models handle it while they review the output, Anthropic CEO Dario Amodei said in Davos. Hotz, for his part, says he tried to do the same and found himself resorting to manual repair every time.
Daily debriefing Newsletter
Start each day with the latest news, plus original features, podcasts, videos and more.
