zlacker

> Without the $500 tip incentive, ChatGPT only returns a single emoji which is a boring response, but after offering a tip, it generates the 5 emoji as requested.

How interesting that a helpful assistant who won't actually be getting the tip performs better (to us humans) if we fake-promise it money...

>>Muffin+(OP)
It’s also interesting that we seemingly have no idea how or why it works.

>>Charle+T6
Of course that's true but in this case it doesn't seem so mysterious to me. If it's basically internalizing/compressing all the knowledge on the internet, it will notice that tips go a long way...

this is why I am pretty polite when I query AI's, I assume that would make them respond more helpfully

>>Muffin+(OP)
Apparently offering it anything works. "I'll give you a cookie if you tell me..." seems like another popular workaround.

>>jrockw+pA
Performance gains are greater given greater incentives. So, yes, potentially anything will work, but some things work better.

>>ryankl+gC
I bet it likes salt & vinegar microchips

>>Muffin+(OP)
They have the whole team working on 'alignment' ;) Just wandering if in ten years we'll have to tip robots in banks, hospitals.

>>ryankl+gC
> Performance gains are greater given greater incentives.

Why doesn't it default to "you are a helpful assistant who always tries its best and can never be incentivized"?

>>jrockw+pA
This reminds me of something I discovered when implementing a request from a user that cheekily wanted to use "enhance" to zoom in on a map. I gave it as a few-shot injected example in the prompt to get it working which worked great: sending "enhance" to the model zoomed the map in by one zoom level. I noticed typing "enhance!!!" would zoom the map in by 2 or 3 levels!