(because it was trained on real people who write with those quirks)
Yes and no. Generally speaking, ML-Models are pulling towards the average and away from the extremes, meanwhile most people have weird quirks when they write. (For example my overuse of (), too many , instead of . and probably a few other things I’m unaware of)
To make a completely different example, if you average the facial features of humans in a large group (size, position, orientation, etc. of everything) you get a conventionally very attractive person. But very, very few people are actually close to that ideal. This is because the average person, meaning a random person, has a few features that stray far from this ideal. Just by the sheer number of features, there’s a high chance some will end up out of bounds.
A ML-Model will generally be punished during training for creating anything that contains such extremes, so the very human thing of being eccentric in any regards is trained away. If you’ve ever seen people generate anime-waifus with modern generative models you know exactly what I mean. Some methods can and are being deployed to try and keep/bring back those eccentricities, at least when asked for.
On top of that, modern LLM chatbots have reinforcement learning part, where they learn how to write so that readers will enjoy reading it, which is no longer copying but instead “inventing” in a more trial-and-error style. Think of the videos on youtube you’ve seen of “AI learns to play x game”, where no training material of someone actually playing the game was used and the model still learned. I’m assuming that’s where the overuse of em-dash and quippy one liners come from. They were probably liked by either the human testers or the automated judges trained on the human feedback used in that process.
Yes and no. Generally speaking, ML-Models are pulling towards the average and away from the extremes, meanwhile most people have weird quirks when they write. (For example my overuse of (), too many , instead of . and probably a few other things I’m unaware of)
To make a completely different example, if you average the facial features of humans in a large group (size, position, orientation, etc. of everything) you get a conventionally very attractive person. But very, very few people are actually close to that ideal. This is because the average person, meaning a random person, has a few features that stray far from this ideal. Just by the sheer number of features, there’s a high chance some will end up out of bounds.
A ML-Model will generally be punished during training for creating anything that contains such extremes, so the very human thing of being eccentric in any regards is trained away. If you’ve ever seen people generate anime-waifus with modern generative models you know exactly what I mean. Some methods can and are being deployed to try and keep/bring back those eccentricities, at least when asked for.
On top of that, modern LLM chatbots have reinforcement learning part, where they learn how to write so that readers will enjoy reading it, which is no longer copying but instead “inventing” in a more trial-and-error style. Think of the videos on youtube you’ve seen of “AI learns to play x game”, where no training material of someone actually playing the game was used and the model still learned. I’m assuming that’s where the overuse of em-dash and quippy one liners come from. They were probably liked by either the human testers or the automated judges trained on the human feedback used in that process.