This is, once again, about diversity and the lack thereof (and I don't mean diversity in a political sense).

LLMs seem fundamentally incapable of producing truly diverse outputs, truly creative and different responses to the same prompts in different runs. Because you and me use the same Claude, if you want a website and I want a website, we'll get (almost) the same website. This is not some BS about "the average of its training data", most of the LLM style (both in design and in text) comes from reinforcement learning. You could RL Claude to produce a very different style, but you couldn't RL it to produce a different style for me than it does for you.

I think this is also where a lot of the complaints about "Claude writing" come from.

RL causes distributional collapse, it's how the models get consistent. Anyone who generated images with early gen (SD1.5-2) models will remember the wild variance between seeds, which newer models have mostly lost, and similarly GPT3.5/4 could produce weirder, more original outputs even if they were less consistently "good" in some sense.

It's worth mentioning that they do RL for aesthetics to some degree based on human expert feedback, but whatever the model tends to produce quickly becomes debased by its ubiquity. They could RL for output diversity, but it's less well studied and likely to cause minor regressions in coding performance, at least until the algorithms are dialed in.