George Hotz, comma.ai and tinygrad creator, warns of declining industry quality and a persistent wave of AI-generated 'slop

"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." -@realGeorgeHotz

Spot on. Seems like a universal experience.

even @geohotz is starting to sound like me 🤣

This is the opposite of what we've seen at http://modal.com. Our most experience SWE is taking most advantage. It's also in disagreement with the Hotz post linked:

"A trait you find in all high performing people is the ability to error correct, and they have mostly been good at seeing when slop is slop."

@geohotarchive “I have tried all the different models, different harnesses, different prompts. It’s not this.”

I’ve watched Hotz try to program with AI. He is doing it wrong. Prob one of the best programmers. But also prob one of the worst agentic programmers.

@deanwball disagree it seems existing top notch SWEs who are open to agentic coding become extremely magnified, obviously some are ideologically opposed or it’s unnatural for their thought process

@deanwball .

@geohotarchive We agree here

I think we’re in a transitional slop era, not an eternal one.

Within 1–2 years, I expect LLM quality and agentic coding to improve enough that much of today’s slop fades. GPT-5.5 is clearly far better at coding than GPT-4, so why assume progress hits a wall here?

I believe we’ll look back at 2022–2026 as the early slop era of AI, quickly evolving into the superintelligence era.

The crux is can LLMs overcome slop, or is slop intrinsic to the paradigm?

@deanwball @tszzl Yeah

Plenty of stagecoach drivers couldn't transition to cars, and if you look at what early cars looked like, they actually appeared closer in form and function *to* stage coaches.

It's why we got things like "brougham" and "sedan" and "coupe" and "cabriolet" with cars.

@strikerglows @geohotarchive @yacineMTB Correct

@geohotarchive Hahahahahahahaha How The Fuck Is Slop Real Hahahaha Nigga Just Walk Away From The Screen Like Nigga Close Vim Haha

@geohotarchive I believe the downsides can be prevented by very strict, non-hackable process. Why do you allow the implementer to change the tests? In companies, LLM could quiz employees at an isolated terminal to see if they understand the code before submitting a PR.

@geohotarchive waaaake me up, when sloptember eeends

@Noahpinion A lot of the code they pump out is ugly and unideal but it usually works. It will only be a Slopcalypse if AI suddenly quit getting better and it's now on humans to go back and unfuck everything into a maintainable format.

@thoughtlesslabs @geohotarchive @yacineMTB Claude code was built with agentic coding. Clearly excellent harnesses are not out of the reach if LLMs.

@geohotarchive @yacineMTB wake me up when sloptember ends

@Indian_Bronson @tszzl exactly. and the radio, you know, was once called the 'wireless telegraph'

@deanwball @tszzl And there was a transition, of nearly 50 years. Plenty of people in the early 1940s in rural parts of the US still relied - not had - relied, or horses to get around

But that was all eliminated by the highways and the new vehicles for new generations being SO much better

@thoughtlesslabs @geohotarchive @yacineMTB Can you elaborate on what sucks about them and what would improve them?