Famed iPhone, Sony Hacker Says AI Coding Agents Are a Disaster Waiting to Happen

George Hotz—the hacker who first cracked the iPhone at age 17 and reverse-engineered the PlayStation 3 earlier than Sony sued him for it—printed a weblog publish Sunday arguing that mass adoption of AI coding brokers will finish in catastrophe, or a minimum of near it.

“I’m calling it now, the adoption of AI brokers into software program improvement will likely be one of the pricey errors within the discipline’s historical past,” Hotz wrote. “Brokers can’t program, and it’s taking longer and longer to comprehend that they’ll’t.”

“The output is damaged, however in a method that’s getting more durable and more durable to detect. Which is precisely what you’d anticipate from an more and more correct statistical mannequin.”

The publish, titled “The Eternal Sloptember,” arrives 5 days after Andrej Karpathy, considered one of AI’s most outstanding researchers, joined Anthropic’s pre-training team with the specific view that AI brokers have already reworked software program improvement. The 2 males now symbolize reverse poles of a debate the trade hasn’t settled—and each have precise credibility to stake a place.

Hotz did not attain his conclusion from the sidelines. He spent six months utilizing brokers on actual initiatives: components of Tinygrad, his open-source deep studying framework, and a whole firmware reverse-engineering of a USB-PCIe chip. “The agent frontloads all of the progress,” he writes, then arms you what he describes as a slot machine lever—you pull it and hope the ending work will get completed.

It by no means fairly does.

Not about ego

Hotz anticipates the apparent pushback: a programmer who defines a part of his id by his craft would naturally resist instruments that threaten to interchange him. He takes the objection severely and dismisses it on the deserves.

“​​I assumed extra concerning the self value preservation factor. Google’s AFL discovered extra bugs than LLMs and no person felt that method about it. Chess and Go are extra in style than ever,” Hotz wrote. And he’s proper within the sense that Chess AI has dominated people for many years and the sport solely grew extra in style.

So, his concern is not about being changed. It is about what occurs to code high quality when everyone seems to be utilizing these instruments without delay, particularly when Massive Tech and Wall Road are continuously pushing for the mass use of those instruments.

“I virtually assume that is some form of psyop to promote brokers,” Hotz argues. “Worry of loss is likely one of the solely methods to make huge firms transfer. Although I believe in that concern they’re making an enormous mistake.”

His central argument is organizational. Excessive performers have tight sufficient suggestions loops to catch agent-generated issues earlier than they ship. They learn the code, spot the errors, and calibrate when to belief the device. “The underside performers will not have that self test,” he writes—and so they’re those utilizing brokers to supply 10 instances their earlier output. At a big firm, that math produces one thing particular: sooner degradation of common code high quality, masked by sheer quantity.

The result, he argues, will likely be “a golden period for buckets and buckets of slop, and a darkish age for gems of high quality.” As a concrete instance, he factors to reviews that Apple is pushing AI coding instruments throughout its complete engineering group, then asks merely: “Do you assume macOS will get higher or worse within the subsequent 2 years?”

The place the camps stand

Hotz now locations himself in what he calls the “LeCun/Marcus camp”—referring to Yann LeCun, Meta’s chief AI scientist, and Gary Marcus, a longtime LLM skeptic. Each have argued that language fashions are basically refined pattern-matchers: They’ll imitate the distribution of current code, however cannot cause by genuinely new issues from first ideas.

Vibe coding—describing what you need in plain language and letting AI generate the implementation—has exploded over the previous yr, and the key labs have positioned agent-based coding as a flagship product. Microsoft transformed GitHub Copilot right into a full agentic system in 2025, with CEO Satya Nadella describing it as a platform-level shift corresponding to the transfer to cloud.

The pushback to Hotz’s place is not summary. Karpathy, who had been skeptical of brokers earlier in 2025, reversed his place after new mannequin releases and joined Anthropic’s pre-training crew on Might 19—5 days earlier than Hotz printed. He described the subsequent few years on the frontier as “particularly formative.”

Anthropic CEO Dario Amodei stated in Davos that some Anthropic engineers have already stopped writing code themselves, letting fashions deal with it whereas they evaluate the output. Hotz, for his half, says he tried to do the identical factor and located himself reaching for the guide repair each time.