Summary

George Hotz argues that AI coding agents are a costly mistake for software development, producing output that mimics real programming but is fundamentally broken in increasingly hard-to-detect ways. After spending six months genuinely trying to use agents for tinygrad and hardware reverse engineering work, he concluded he could have done it better and faster manually. He contends that agents will disproportionately harm large organizations where weak feedback loops and bottom performers generate massive volumes of low-quality code. He rejects the narrative that skepticism stems from status anxiety, comparing it to how AFL found more bugs than LLMs without triggering existential dread. Hotz ultimately sides with the LeCun/Marcus camp, arguing that current LLMs lack the world models necessary for real programming and that RLVR-style training is fundamentally insufficient.

Key Insight

AI coding agents produce statistically convincing but fundamentally broken output, and the organizations that adopt them most aggressively will suffer the most as undetectable slop compounds through their codebases.

Spicy Quotes (click to share)

• 8

  Agents cannot program, and it's taking longer and longer to realize that they can't.

• 7

  The output is broken, but in a way that's getting harder and harder to detect. Which is exactly what you'd expect from an increasingly accurate statistical model.

• 7

  The agent frontloads all the progress, then gives you a slot machine lever to pull to hope it gets the polish done. It never quite gets there.

• 7

  I almost think this is some kind of psyop to sell agents. Fear of loss is one of the only ways to make big companies move.

• 7

  The bottom performers won't have that self check. They are the ones producing 10x output with the agents. What do you think is happening to the average output of that organization?

• 8

  It is a golden era for buckets and buckets of slop, and a dark age for gems of quality.

• 5

  AI produced artifacts are not produced by the same process as human ones, and this difference, while extremely subtle in statistics, makes itself obvious when you try to interact with and build on the artifact in human ways.

• 9

  The real story of this era will be who manages to avoid harming themselves in their AI psychosis.

Tone

contrarian, empirical, warning