Why is it that so many AI models have this tendency to add usefless fluff like "Sure, here's an answer to your question: " to their output?

If it was a Chat model I'd understand, but this happens with Instruct models all the time too. I'm looking at you Gemma.

My arguments against it:

1. It costs extra tokens, and thus time.

2. It costs extra tokens, and thus electricity.

3. Removing this stuff from the output is a chore, while for people who want this fluff, it's relativy easy to use a simple bit of code to generate random versions. That's a silly imbalance, especially in relation to point 2.

4. A societal issue is that people tend to antropomorphize LLM's, which can cause them to more uncritically consume their output. Making LLM output seem 'like they're talking to a human' helps to obscure that it's all just 'statistics on steroids'.

5. It causes rant posts on Reddit.

Here's an example where Gemma 2 2B is summarizing Alice in Wonderland:

```

Sure, here's the summary:

Alice was feeling bored sitting by her sister on the bank. She wanted to do something but didn't know what.

Sure, here's a concise summary of the text:

The Mouse was shy and timid, and was afraid of the French language. It was adamant about not discussing its family's dislike for cats, refusing to be called by its name.

Sure, here is the summary you requested:

Alice was apologetic after an argument with a Mouse. etc

```

Please keep this stuff contained to Chat versions of models.

// End of rant