Hacker News Reader

Comment by xwowsersx | Hacker News Reader

Isn't it wild how seemingly overnight we've become comfortable speaking to AI as if it were human? Writing things like "Actually, one more thing." to ChatGPT for example, are strictly unnecessary, but I write like this also even though it's more verbose than a simple "Yes" or even just rudely submitting the next thing I need help with. Fascinating.

View on HN

I've wondered about this. If it's trained on human responses, and humans are more likely to be helpful to a polite question, then shouldn't ChatGPT also be more inclined to give helpful replies if I say please and thank you?

It sounds strange, but I actually ran into a prompt problem a few weeks ago for a GPT project I'm working on. If I don't include "please" in the prompt, the response is the typical "I'm just an AI and can't do that". I had to leave a comment in the code to be aware of changing the prompt in the future, because if we aren't polite to the LLM it may not respond correctly.

Obviously too much INTERCAL in the training set.

edit: someone else made the same joke 5 mins ago :D

Can you pre-prompt politeness?

For example, ‘assume every question I am about to ask is phrased in an incredibly polite way.’

Unfortunately the data it’s trained on (internet text - commoncrawl) has an antisocial data bias, and there’s no alternative “pro-social” data corpuses that exist.[1]

Every one of these has to be “taught” not to be toxic - usually by some version of exploited labor paid pennies to look at the worst stuff on the planet [2]

It’s true of every larger model trained on internet corpus - Microsoft notoriously had this issue with Taybot

[1] https://hci.stanford.edu/publications/2022/Park_ContentModAu...

[2] https://time.com/6247678/openai-chatgpt-kenya-workers/

The data it was trained on and how it is weighted is mostly proprietary.

While I'm sure the publicly available and already moderated Reddit comments OpenAI pays workers to classify are truly disturbing, nobody should be forced to read Reddit after all, I think the moderators at Facebook, Google, porn websites, etc, that have to deal with moderating graphic videos of extreme violence, torture, gore, abuse, sexual abuse, child sexual abuse, etc, are the ones that have to look at the worst stuff on the planet.

They both use the same type of outsourced labeling services:

https://www.sama.com/

I appreciate your point, and it's horrifying.

I still wonder how many coding questions, stack overflow posts, comments on tech blogs etc just naturally have better responses to polite and clear discourse? SO is heavily moderated, and I'm yet to see unmoderated vitriol in the comments of a useful blog post. And then the horrifying reality of the human assisted fine tuning process would also push it towards polite discourse.

I mean it's definitely possible that using polite language gets you into better latent space in the model for productive answers. Personally (and very subjectively) I get a lot of good use out of ChatGPT and I wonder why others seemingly find it useless. Could it be because I'm polite and say "thank you" before moving to the next ask (within the same conversation)?

> "Actually, one more thing." to ChatGPT for example, are strictly unnecessary

I won't run a conclusive test about it but treating ChatGPT well gives me much better results.

Someone should. I've noticed this phenomenon as well.

It's not "strictly" unnecessary since the LLM is responding with respect to its training, this type of phrasing could actually produce more desirable output if the training data favors conversational exchanges with this type of phrasing.

Yeah if ChatGPT helps me solve a hard problem I always tell it thanks.

I’ve read multiple places that being polite and professional actually results in better output quality. I think recently there was even a paper published quantifying it.

It is interpreted as positive/affirmative against the correctness or relevance of the previous output. Telling it 'That is correct' will achieve a very similar effect (as will saying 'No'/'That isn't relevant'/etc. would in the converse.)

This cracks me up. For decades only INTERCAL demanded an appropriate level of politeness. It was a great joke, and we all laughed.

Now it’s an actual real thing that the computer performs the task better when you ask nicely! (I am also polite and professional whenever I talk to ChatGPT)

I believe the rationale is that in the LLMs’ training data, higher-quality information is available in places where people are polite. The LLM is just copying that statistical fact.

People have been doing that for decades. People were treating ELIZA (from 1964) like a real human. Human treat their household appliances like other humans.

It's trained on humans, so it makes sense that you get the best out of it by treating it in the way that gets the best out of humans.

It's hardly overnight, we've been talking to smart assistants for years.

It's not like it has no effect.

https://arxiv.org/abs/2307.11760

Reply on news.ycombinator.com