Comment by polycaster | Hacker News Reader

Comment by polycaster | Hacker News Reader

While the post is specific to Claude, interestingly I found the same terms and language patterns appear noticeably more often in recent Open AI conversations as well. Also in the same timeframe as for the Claude models. And I can see two reasons for that. One being that one company is training on the models of the other company, but I highly doubt that because of the apparent similar appearance without noticeable delay. And the other being that there has been some fundamental realization on how to improve and shape conversations.

I see people around here giving advice on how to reduce this vocabulary in the model output. I’m not certain this is a clever thing to do, as it appears to me, the emphasis on certain words is deliberate, and an important step towards overall quality as the words are being played back to the model in the next turn and thus supporting the model shape its own thoughts and stay on course.

> words are being played back to the model in the next turn and thus supporting the model shape its own thoughts and stay on course.

Why is that exactly?

It’s obviously more complex than this, but my intuition is that the previous messages of the conversation define the playing field for the next response in the same way as thought blocks do. consequently, when the model arrives in round n and “notices” a lot of words like “plain”, “genuine” and so on in its own previous responses, then, in its pursuit of continuity, it takes this as a strong trigger that it should not be talking fiction here. If that picture is right, then trying to remove those terms would be a bad idea. At least if you intend to continue the conversation.

Don’t you feel your conclusion is preceding the reasoning, forming a confirmation bias?

Have quick test. In a codebase or PR (assuming you’re a dev) with claude’s signature wordings unmodified, command claude (/code-review) with same “thinking” effort and model. You may find (as I have many times) that claude points out the comments it made in earlier sessions, as factually wrong or misleading. Repeat this in a fresh session a few times and on different code.

Reply on news.ycombinator.com