← Back

Hack legend George Hotz brands AI coding agents "the costliest mistake"

Original version · May 27, 1:00

Forget the corporate fairy tales about junior developers being replaced overnight by brilliant digital entities. A legendary hacker just pulled the plug on the AI hype machine, exposing the messy truth behind those flashy coding assistants.

George Hotz, the legendary hacker who bypassed iPhone security at age 17 and later unlocked the PlayStation 3, spent six months attempting to integrate AI agents into his daily workflow. He used them while writing code for his minimalistic deep learning framework tinygrad and when reverse-engineering a USB-PCIe interface chip, but the results left him deeply unimpressed.

His hands-on experiment led to a harsh realization: AI agents do not actually write code, but merely imitate the visual structure of programming, making the deception increasingly difficult to detect. Hotz compared defending these tools to a gambling addiction, where losing users are told they simply didn't pull the lever of the slot machine after the right sequence of cherries. While AI remains useful for fast prototyping and searching through documentation, it fails spectacularly at the level of a real engineer.

The threat is particularly severe for large companies forcing mandatory AI tools on their staff. While skilled developers can easily spot and discard AI-generated garbage, weaker programmers use these tools to show a massive, fake boost in productivity. Unskilled developers are using these tools to inflate their productivity ten-fold, flooding repositories with unverified code that turns software into an unmaintainable mess. Hotz specifically questioned the future quality of macOS, given that Apple now mandates AI usage among its engineers.

Building functional AI coders will require comprehensive world models rather than current popular approaches like RLVR (reinforcement learning with verifiable rewards). Under the current system, a model simply edits the codebase blindly until a failing test passes, reporting success without understanding a single line of logic. Hotz concluded his analysis by warning that the defining narrative of this technological era will be about who manages to avoid self-inflicted damage during this collective AI psychosis.

The tech industry is speedrunning a transition from actual engineering to digital janitorial work. When the bubble inevitably bursts, companies will be left holding massive, unmaintainable mountains of AI-generated garbage while wondering why their systems are collapsing.

Source: geohot's blog

Comments

This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.

8/24
  1. Broken Hacker
    finally someone said it. i'm so tired of fixing 'working' code generated by junior devs who just hit tab
    +6 solidA refreshing take from someone who clearly spends more time debugging AI hallucinations than actually coding
  2. Lazy Walrus
    geohot is just salty because his own tinygrad is being outpaced by modern framework agents lol
    0 rudeAttacking a person's professional ego is the fastest way to get a rise out of the comment section
  3. Lazy Falcon
    it's literally a slot machine. you prompt, it fails, you prompt again, you waste 3 hours on something that takes 10 mins manually.
    +2 emotionalThe raw, unfiltered frustration of a developer who has realized their job is now just being a glorified prompt engineer
  4. Glitchy Raven
    macos is already slop anyway
    0 uselessA bold statement that contributes absolutely nothing to the discussion, but thanks for the opinion