Hacker News Reader

Comment by BobbyJo | Hacker News Reader

I use Claude and I find that it speaks in a very obfuscated manner when explaining things. It seems to make up jargon as it goes on top of spending a lot of tokens dancing around a point. I often find myself having to ask it to rephrase things, or speak directly about mechanism or consequence, in order to understand the point.

View on HN

Im starting to believe it is exactly that...obfuscation. it started after it was known Chinese were generating models that were much smaller and more accurate from the llms response. The more generally articulated the answer is along with its accuracy in what it is and not what it isn't along with many strange never written before English euphemisms that dont make sense is definitely semantical obfuscation probably added in as a post processing transform to prevent these types of actors from being able to get easy to digest answers, and leave watermarks in their models to detect.

Using Claude for any kind of technical writing makes me feel like it was trained on snarky Huffington Post articles written by a 23 year old mixed media arts graduate and then was told to intentionally obfuscate the most important elements of any text by extensively rambling about what was not done and for what reason.

GPT is less bad for this, which is why I've mostly shifted to using it.

Not to accuse them of doing this, but AI vendors have an incentive to generate verbose responses, given that you pay per token

Reply on news.ycombinator.com