Hacker News Reader

Comment by Retric | Hacker News Reader

People use language in their own distinct fashion. To the point where you can often figure out the writer from a long enough sample. Newspapers and internet comments end up a blend of different voices but in visually distinct boxes.

LLM intentionally randomize output to cover up all the plagiarism, but this largely destroys the consistency that cause writers to feel unique. Larger patterns still emerge but it just feels stale.

View on HN

That's a misconception. LLMs are trained to suppress regurgitation, not because of plagiarism concerns, but because quoting training data verbatim means the model didn't learn from it, just memorized it.

Models absolutely do have voices. Spend some time using them, and you'll see. I can easily tell Claude-written prose apart - a consequence of using it daily.

You misunderstood, “Larger patterns still emerge” means yes you notice LLM’s but it’s for different kinds of patterns than you get from people.

> not because of plagiarism concerns

I’ve second hand accounts of meetings focused on exactly the problem of how much verbatim text was being spat out and what to do about it. Some versions of specific models are far worse at this but it’s a huge legal liability not to address it.

Models are of course different from humans, because the process by which they were created is different. But what’s undeniable is that humans also love to quote or repeat what others have said — we learn from our own kind, that’s how we gain experience. Yet rarely does anyone hold a person to a word-for-word standard over the course of their entire life, checking how many of their sentences were picked up from other people’s conversations — and then call that infringement. Just as no one right now would say I’m plagiarizing a model’s speaking style because I use em dashes, and no one would consider it misconduct when I reference a famous person’s ideas in everyday conversation. I don’t think models intend to infringe; it’s simply how they learned language. But perhaps, just like humans, they need to learn to provide more citations for what they reference.

Even individual words can be an issue. Humans understand when you want a new story about a magic school there’s issues using Hogwarts as the name.

LLM’s will happily regurgitate multiple paragraphs practically verbatim if you don’t specifically design them not to behave in this way.

Because ultimately when you design a system to best predict the next output based on a given input perfect replication of prior training is the correct response mathematically.

Reply on news.ycombinator.com