George Hotz — the hacker who cracked the iPhone at 17 and reverse‑engineered the PlayStation 3 before facing Sony in court — has issued a blunt warning: widespread use of AI coding agents could be “one of the most costly mistakes in the field’s history.”
In a Sunday blog post titled “The Eternal Sloptember,” Hotz argues that agentic coding tools don’t actually “program” in any reliable sense. “Agents cannot program, and it’s taking longer and longer to realize that they can’t,” he wrote. His core claim: these models produce output that looks plausible but is subtly and increasingly broken, and those flaws are growing harder to detect as the models’ statistical fluency improves.
What he tested, and what he found
- Hotz didn’t reach this conclusion from thought experiments. He spent six months using agents on real work: extending parts of Tinygrad (his open‑source deep‑learning framework) and performing a full firmware reverse‑engineering of a USB‑PCIe chip.
- His experience: agents “frontload all the progress” — they get you through the obvious parts quickly, but then leave what he likens to a slot machine lever where the finishing, reliable work is expected to happen. In practice, it often doesn’t: you repeatedly have to reach for manual fixes.
An organizational risk, not just an ego issue
Hotz rejects the idea that his critique is rooted in craft tribalism. He points out that prior automated tools (for example, fuzzers like AFL) found more bugs than humans without sparking the same panic, and that games dominated by AI (chess, Go) have thrived. His concern is structural: high performers have tight feedback loops and can catch and correct agent‑introduced errors. Lower performers—whose output can balloon by tenfold with agent help—won’t. That dynamic, Hotz warns, will accelerate a decline in the average quality of shipped code: “a golden era for buckets and buckets of slop, and a dark age for gems of quality.”
He frames the push for mass agent adoption as partly commercial psychology — “I almost think this is some kind of psyop to sell agents” — arguing that fear of falling behind drives big companies to adopt tools before fully understanding the downstream risks. He even name‑checks reports that Apple is pushing AI coding tools across its engineering organization and asks bluntly: “Do you think macOS will get better or worse in the next 2 years?”
Where Hotz sits in the broader debate
Hotz places himself in what he calls the “LeCun/Marcus camp” — aligned with thinkers like Yann LeCun and Gary Marcus who view large language models primarily as powerful pattern‑matchers that imitate existing distributions of code, rather than systems that reason from first principles.
On the other side, some leading researchers see agentic coding as a fundamental shift. Andrej Karpathy — who had been skeptical of coding agents earlier in 2025 — reversed course after recent model advances and joined Anthropic’s pre‑training team on May 19, 2026. Anthropic CEO Dario Amodei has said at Davos that some engineers at the company have already stopped writing code themselves and now review model output. Hotz reports trying that workflow and repeatedly resorting to manual interventions.
Why this matters for crypto developers and DeFi
The debate isn’t academic for blockchain and crypto engineering teams. Smart contracts, wallets, bridges, and on‑chain infrastructure are unforgiving environments where subtle bugs can become multi‑million‑dollar exploits. If agent‑generated code introduces hard‑to‑detect flaws that slip past weaker review processes, the cost of mass adoption could be especially high for decentralized finance and security‑critical crypto systems.
Bottom line
Hotz’s blog is a high‑profile, practitioner‑level warning against moving too fast with agentic coding. His position sets up a live industry fault line: enthusiastic believers who think agents will transform software development (and are already doing so), versus skeptics who fear systemic degradation in code quality when agents are deployed at scale. For developers and teams in crypto — where correctness and auditability are paramount — the tradeoff between speed and safety has never been more consequential.
Read more AI-generated news on: undefined/news
