In short
- George Hotz, the hacker behind the first iPhone jailbreak and PlayStation 3 crack, published a blog post on Sunday calling the adoption of AI coding “one of the most expensive mistakes in the history of the field.”
- His main argument: good developers can see the negative effects of agents, but weak engineers can’t – and it’s weak engineers who produce ten times as much, lowering the quality of the code on a large scale.
- The post came five days after Andrej Karpathy joined Anthropic’s training team with a different opinion, reflecting a split among senior engineers over whether AI assistants really work.
George Hotz—the hacker who first hacked an iPhone at age 17 and modified a PlayStation 3 before Sony sued him for it—wrote a post on Sunday arguing that mass adoption of AI will end in crisis, or close to it.
“I’m calling it now, the implementation of AI assistants in software development is one of the most expensive mistakes in the history of the field,” Hotz wrote. “The agent can’t make it, and it takes a long time to realize that he can’t.”
“Its output is degraded, but in a way that’s getting worse and worse. That’s what you’d expect from a very accurate estimate.”
The post, titled “The Eternal Sloptember,“Arriving five days later Andrej Karpathy, one of the most famous AI researchers, joined the Anthropic pre-training group It is a clear assumption that AI assistants will already change software development. The two men now represent the controversial issues that the industry has not resolved – and both have a genuine belief in taking responsibility.
My update: I have joined Anthropic. I think the next few years at the frontier of LLMs will be very productive. I am very excited to join the team here and return to R&D. I am passionate about education and plan to resume my career in due course.
Andrej Karpathy (@karpathy) May 19, 2026
Hotz did not reach his end. He spent six months using agents for real things: parts of Tinygradits open source deep learning system, and the USB-PCIe chip firmware update system. “The assistant fills in all the action,” he writes, and gives you what he describes as an entry-level control panel — you pull it up and hope the final task is done.
It doesn’t at all.
Not about ego
Hotz hopes to push the obvious: a software developer who defines some of his identity through the use of his creativity will naturally resist tools that threaten to replace him. He takes the objections seriously and removes them from his reasons.
“I thought a lot about something to protect myself from. Google’s AFL they found more bugs than LLMs and no one felt that way. “Chess and Go is more popular than ever,” Hotz wrote. And he is right that Chess AI has been dominating people for years and the game has only become more popular.
Therefore, his concern is not to be replaced. It’s about what happens to the code when everyone is using these tools at the same time, especially when Big Tech and Wall Street are constantly pushing to use these tools.
“I think this is some kind of psyop to sell supporters,” Hotz says. “The fear of losing is one of the ways that makes big companies go. Although I think that because of that fear they are making a lot of mistakes.”
His central argument is one of organization. The successful ones have connections strong enough to handle problems created by agents before deployment. They read code, spot errors, and improve the reliability of the tool. “The players on the ground can’t measure themselves,” he writes – and they are the ones who use agents to make the 10 rounds they have already released. At a large company, this math produces something else: a rapid decline in code quality, hidden by its volume.
As a result, he argues, “the golden age of buckets and low buckets, it’s the dark age of high prices.” As a concrete example, he points to reports that Apple is pushing AI writing tools throughout its engineering team, then asks simply: “Do you think macOS will be better or worse in the next 2 years?”
Where are the camps
Hotz now places himself in the “LeCun/Marcus camp” – referring to Yann LeCun, Meta’s chief AI scientist, and Gary Marcus, a long-time skeptic LLM. Both of them said that languages are the best analogies: They can imitate the distribution of existing codes, but they cannot think through new problems from first principles.
Vibe coding-Explaining what you want in simple language and letting AI do the implementation-has exploded in the last year, and major labs have put agent-based scripting as a popular feature. Microsoft to be transformed GitHub Copilot in 2025, and CEO Satya Nadella describes it as a platform change similar to the move to the cloud.
The push to Hotz’s position is not unheard of. Karpathy, who is suspected of being a supporter as early as 2025, changed his position after the release of the new model and joined the training team of Anthropic on May 19 – five days before Hotz’s release. He described the next few years at the border as “mainly formations.”
Anthropic CEO Dario Amodei said in Davos that some Anthropic experts have already stopped writing their notes, allowing the models to run while they analyze the results. Hotz, for his part, says he tried to do the same and found that he was always planning to improve.
Daily Debrief A letter
Start each day with top stories right here, including originals, podcasts, videos and more.
