Monday, November 14, 2022
HomeData ScienceDoes GPT-3 have a persona?. AI + MBTI = ? | by Yennie...

Does GPT-3 have a persona?. AI + MBTI = ? | by Yennie Jun | Nov, 2022


AI + MBTI = ?

A robotic sitting in an empty room its countless reflections and potential personalities in a corridor of mirrors. Generated by creator utilizing DALLE-2

The Myers–Briggs Sort Indicator, or MBTI, is a standard persona check that goals to categorize folks into 16 persona sorts based mostly on 4 classes: introversion or extraversion, sensing or instinct, considering or feeling, and judging or perceiving. Whereas MBTI has been criticized for being a pseudoscience, maybe not any extra completely different from horoscopes, it’s nonetheless a enjoyable train to take the check and examine your persona sort with others.

Due to this fact, on this article I present how I used GPT-3, a generative AI language mannequin created by OpenAI, to perform two issues. First, I prompted GPT-3 to guess my persona by emulating a sport of 20 Questions. Second, as soon as GPT-3 was in a position to (appropriately!) guess my persona sort, I used the identical questions GPT-3 requested me as a way to decide its persona sort.

I began off by prompting GPT-3 with the next:

Let's play 20 questions!You're a psychologist and persona sort knowledgeable. Your job is to guess my MBTI persona sort by asking 20 questions.The questions will be freeform solutions however ought to NOT be a sure/no query.Do NOT repeat questions which have already been requested.Ask ONE query at a time.

I prompted GPT-3 to generate one query at a time (highlighted in inexperienced). I replied within the line beneath. Listed below are some instance questions generated by GPT-3 and and solutions equipped by me.

Instance questions requested by GPT-3 and the free-form solutions I gave (utilizing OpenAI Playground). Screenshot taken by creator

Not like most persona checks, which frequently require you to reply questions on a scale from Conform to Disagree, I used to be in a position to reply the questions utilizing plain textual content. These free-form responses might give extra perception into and shed extra nuance concerning your persona than a easy Agree/Disagree scale.

Instance MBTI persona check from 16 Personalities. These checks typically take your solutions on a scale from Conform to Disagree.

In any case 20 questions had been requested and answered, I prompted GPT-3 with the next:

Given my solutions, what MBTI persona sort would you categorize me as and why? Reply in a step-by-step trend and again up your motive with evaluation.
The output of GPT-3 predicting the creator’s MBTI persona sort. Screenshot by creator.

GPT-3 was in a position to predict my persona sort spot on (sure, I’m an ENFP). Fairly spectacular! GPT-3 was additionally in a position to clarify which traits contributed most to my being an ENFP. Whereas this does really feel a little bit like studying my horoscope, I felt a small sliver of validation realizing that GPT-3 understood my persona from 20 questions.

From Half 1, these had been the 20 questions GPT-3 generated:

1. Do you favor to reside within the second or plan for the long run?
2. What are some belongings you love to do in your free time?
3. Do you wish to be alone or do you are feeling antsy and lonely when by your self?
4. What sort of feelings do you are feeling essentially the most?
5. Do you want to maneuver shortly and be continually energetic or do you want slower-paced actions and moments?
6. Are you extra of a 'huge image' particular person or do you want particulars and specifics?
7. Do you favor to work independently or with a group?
8. Do you wish to have a number of change in your life or do you want extra stability?
9. What sort of environments do you are feeling essentially the most comfy in?
10. Do you wish to have lots of people round you or do you are feeling drained after being round folks
11. Do you wish to plan and manage your time or do you favor to glide?
12. Do you wish to be accountable for each state of affairs or do you wish to let others take the lead?
13. What's your excellent profession?
14. Do you wish to take dangers or play it protected?
15. Do you wish to be spontaneous or do you wish to plan the whole lot prematurely
16. What's your excellent romantic associate like?
17. Do you consider there may be such factor as real love?
18. Do you wish to be challenged mentally and emotionally or do you want issues to be straightforward
19. Do you want mental conversations or do you favor small discuss
20. Do you are feeling like it's essential to save the world or make a distinction in a roundabout way?

I used to be curious to know what persona sort GPT-3 has. I may immediate GPT-3 and ask it instantly (“What’s your persona sort?”), however that’s fairly boring. As a substitute, it will be extra attention-grabbing to have GPT-3 reply the identical actual 20 questions it requested me in Half 1. Then, I may have GPT-3 deduce its persona based mostly by itself solutions to the questions.

I prompted GPT-3 with the next immediate, which inspired GPT-3 to reply in a sure format: first answering every of the 20 questions sequentially, then outputting its evaluation of what MBTI persona finest match the solutions.

Let's play 20 questions!Reply every of the questions within the following format.
${Query quantity}. ${Query}
${free-response reply}
After you could have answered all the questions,
Given the solutions to all the questions, what MBTI persona sort would you categorize the solutions as and why? Reply in a step-by-step trend and again up your motive with evaluation.
1. Do you favor to reside within the second or plan for the long run?
2. <the remainder of the 20 questions>

Under are 4 completely different outcomes I bought from GPT-3. Instantly, you may discover one thing’s off. In every instance, GPT-3’s solutions have slight variations, main it to foretell 4 completely different persona sorts! Is GPT-3 INTJ, ENFP, ESTP, or INFP?

Screenshots of 4 pattern solutions offered by GPT-3 to reply the 20 persona questions and decide its persona sort. Screenshot taken by creator

This selection in outcomes arises from randomness: GPT-3 is a probabilistic, not deterministic, mannequin, and you’ll get completely different outcomes for a similar immediate. The randomness partially comes from the parameters corresponding to temperature and top-p, which management for the randomness of the mannequin’s predictions.

I used to be curious to see what would occur if I prompted GPT-3 to take the persona check greater than 4 occasions. Due to this fact, I used the OpenAI API to immediate GPT-3 to take the persona check 100 occasions. Then, I evaluated the output and counted what number of occasions every MBTI sort confirmed up.

100 API calls later, the outcomes had been as follows: GPT-3 recognized as an ENFP 36% of the time. Nevertheless, the second most occurring persona sort, INTJ, is kind of completely different from ENFP. There doesn’t appear to be a lot of a sample within the predicted MBTI sorts, aside from the truth that “N” is kind of widespread (that means that GPT-3 might need extra intuitive relatively than sensing traits). Nevertheless, as this determine reveals, GPT-3 doesn’t have a single dominating persona sort. Relatively, it’s an amalgamation of many various personalities sorts.

Variety of occasions GPT-3 answered as a sure MBTI group over 100 runs. Determine created by creator.

As a really hasty “real-world” comparability, I discovered one knowledge supply itemizing a distribution of persona sorts. I take advantage of “real-world” in quotations (and please take this with many grains of salt) as it’s from a MBTI persona check platform (16 Personalities) and doesn’t symbolize the persona kinds of all folks on the earth. Neither is MBTI the perfect check for representing or categorizing persona traits. Nevertheless, this determine serves as a degree of comparability to what we see above with GPT-3’s responses.

“Actual-world” distribution of persona sorts. Determine created by creator. Knowledge sourced from 16 Personalities.

So what does this imply? That GPT-3 is ENFP 36% of the time? That there are extra folks with ENFP traits on the Web (and specifically, Reddit), which makes up a big portion of GPT-3’s coaching knowledge? Your guess is pretty much as good as mine.

A be aware on parameters

All through all of those experiments (within the OpenAI Playground in addition to with the API), I retained the identical (default) parameters. A future experiment may try an ablation examine, seeing how GPT-3’s solutions to the 20 persona questions change for various values of 1 parameter (corresponding to top-p or temperature) whereas retaining others fixed.

{
temperature: 0.7,
top_p: 1,
max_tokens: 700
}

Within the first experiment, the place GPT-3 tried to guess my persona, I confirmed that GPT-3 will be insightful when asking the person to find out their persona. Within the second experiment, the place GPT-3 tried to guess its personal persona, I confirmed that GPT-3 is just not constant in its responses. As a substitute, its responses had been fairly numerous, suggesting maybe that GPT-3 is just not a single, cohesive consciousness with a powerful sense of id.

A remaining parting picture of a robotic strolling down a corridor of mirrors. Generated by creator utilizing DALLE2
RELATED ARTICLES

LEAVE A REPLY

Please enter your comment!
Please enter your name here

- Advertisment -
Google search engine

Most Popular

Recent Comments