Here we explore the idea of priming GPT-3 to respond in a mixture of any of the Big Five personalities. Imagine being able to tune the output of your text so that it can generate a conversation that appears to be originating from any combination of the Big Five personalities. That’s the purpose of this investigation, how do we contextually prime GPT-3 so that it do this.
Big Five personalities are defined as:
he theory identifies five factors (see: https://en.wikipedia.org/wiki/Big_Five_personality_traits )
- openness to experience (inventive/curious vs. consistent/cautious)
- conscientiousness (efficient/organized vs. extravagant/careless)
- extraversion (outgoing/energetic vs. solitary/reserved)
- agreeableness (friendly/compassionate vs. challenging/callous)
- neuroticism (sensitive/nervous vs. resilient/confident)
By now it is well understood how to have GPT-3 generate language based on a particular style. Here’s a hilarious example where I have it generating clickbait titles. These were primed with a context prefix string.
It seems that someone who drunk too much caffeine arises out of these clickbait titles!
So for our first experiment, let’s context prime GPT given examples of each psychological style. Let’s explore how good it is right out of the box! Apparently it works very well:
The examples here come right of the box.
To fine-tune this, we want to prime on a compact combination of examples. So let’s see if we can find some example text for each Big Five personality trait. The better the examples we use on a prefix string, the better the results.
Here I primed the text with the first example and added the words ‘attentive’ and ‘forgetful’. Indeed very interesting results.
Indeed intriguing that it appears you can build an entire product around this with just the right UI and use case!