George Hotz says AI coding agents bypass problem-solving by commenting out failing tests

⚠️⚠️⚠️i don’t think most people understand the implications of the mood shift below, so I will spell them out. they are serious, and eventually will affect the global economy.

when a serious coder as hardcore AI-philic as GeoHotz dives deep into all the latest coding models and winds up echoing what I have been saying about the code turning out slop—and how that is going to mess over big companies—the whole genAI movement is a deadman walking.

‘cause coding was the tentpole. and its net impact may actually be net negative—and certainly not big enough to justify the massive investments.

once people figure that out, the bubble will pop.

"Without fully endorsing all their ideas, I’m now in the LeCun/Marcus camp on LLMs. I don’t think models like this will ever be able to program, I think the process matters. "

Great that so many of you have enjoyed my deconstructions of Elon, Zuckerberg, and that OpenAI dude, but this is much more important:

@lukaszsawickiwx do you have any actual data for any of that?

@krishnanrohit More cope from Hotz that'll cling onto for far too long, like he did with COVID and masks and neoreaction and all sorts of nonsense.

What he's describing is a relatively fair assessment of the current state but it's almost certainly going to be way better within a few years.

@ns123abc I've been playing around with LeWorldModel inspired training and I agree there is something there that adds a new level of latent relationship predictions you won't get with LLMs (or that you can predict with LLMs / lexical matching but it's incredibly expensive and limited).

@GaryMarcus And in this video, a well-known developer harshly criticized AI-based coding in general.

LLMs mostly struggle with the top 1% hardest coding use cases. But in most companies, average AI-generated code is already better than average code hand-crafted by junior/mid contractors and ships at 10x the speed.

Unless human-written code improves, things will stay more or less as they are.

@ns123abc They won’t do shit with automation at scale without appropriate guardrails.

Bounded recursive self improvement is the way.

https://github.com/MontrealAI/agialpha-first-real-loop

@krishnanrohit The next sentence was: >I think that deep learning is still the solution, but real programming agents will need world models, not some RLVR shit that comments out the failing test and tells you all the tests are now passing.

@ns123abc He’ll be right if GPT 6 / Mythos disappoint

@Perpetualmaniac @GaryMarcus Ah we finally heard from Zach Vorhies famed inventor of that brilliant product (snaps fingers) you know the one that everyone’s raving about…the next Torvalds you know who I mean

@ns123abc George in the Lecun/Marcus camp, big L for roon

@ns123abc this is hilarious because he argued me out of lecun brain years ago when models were just becoming public here on twitter

@krishnanrohit it’s only programming if it’s made in the champagne region of the human brain

@krishnanrohit during his debate with @allTheYud , @realGeorgeHotz , supposedly flag bearer for e/acc, said multiple times that AI able to do sci-fi stuff like harness the sun's energy was "thousands of years away". which imo is insane if you expect AI R&D to continue. people didn't react

@ns123abc Wrong about LLMs being able to code AGAIN. He was saying GPT4 sucked at coding when it came out. It was not great but LLMs got better and will continue because the loss and data are more important than the model. With code the loss is unambiguous and generating the data is easy.

@ns123abc

@GaryMarcus Dude I’m building frontier level stuff with these so called “slop generators”. Stuff that’s beating the pants off ape coded code that serves as the gold standard of open source solutions.

Obviously you are NOT using these things 12 hours a day.

I am.

@ns123abc