Yes, that's why good datasets are important. But usually the difference between model generations is architectural improvements

late training is where a lot of capability gains come from too

Interesting, how does that work?

the art of reinforcement learning

How much do we assume human written code for training the original models is free of human made slop? We all know we all take shortcuts and have code we are not proud of. If they are I deed averaging machines, is it possible their output is the average of human output?

I'd argue that the way attention works plays into the slop too

30 years of looking at human written code, not great :) in the early years for me (late 90’s) code was just solid. dotcom boom happened and “overnight” everyone was a fucking coder. there are too many of us and average coder is so bad that you would never hire them. not only that, even at 90 percentile is fairly terrible coder. AI has absolutely been training on this garbage human code so it was terrible at coding initially… but it is evolving now on its own

They were trained on a lot of garbage human-written slop originally, they have to post-train the models so the responses are good. Luckily if you’re an AI company you probably have the whole chat, so it’s easy to categorise it: the user might respond “Good job, that fixed it” or “WHAT THE FUCK IS THIS??? WORST CODE I’VE EVER SEEN”.

What tests did you run on code from the older models? What kind of security review did it go through? When you say it was "not so good," do you mean it failed to compile or couldn't even pass basic tests?

Model collapse theory was complete nonsense pushed in the early days of LLMs by people who had no idea what they were talking about and thought that LLMs were literally just stochastic parrots that were simply outputting the median token in a probability distribution containing increasing amounts of junk.

In reality there are lots of ways to bias towards quality and all decent labs will be feeding quality signals via techniques like RLHF and supervised training so that you don't get the median human-slop output, but the borderline super-human modern LLM output.