Fun with Assistants -- getting them to really talk to each other

[dlaytonj2]

I previously posted a simple use case of two assistants on one thread. The first assistant was the Poet and the second an Analyst of poetry. First I had the Poet write a poem on a topic, then I had the Analyst provide feedback, then I had the poet rewrite the poem. Finally I had the analyst provide feedback on the rewritten poem. It worked well, but I was never entirely sure if the poet was really “listening” to the Analyst’s feedback.

Revisiting the program, I made a simple change to the prompt – I had the assistants introduce themselves, before setting about their work.

My prompt for the Poet was as follows:

Identify yourself as the Poet. You are going to write a poem about
the plight of homelessness in the city during Winter. Use a free verse style. There should be no internal or external rhyming. Think about this step by step. First describe your approach to the poem, then write the poem.

I did the same for the Analyst, and the change in the messages the assistants produced was remarkable. It was like they suddenly acknowledged each other’s existence. Now after the Poet wrote the first draft of the poem. It added the following to the message:

I humbly pass this to you, Analyst, for your evaluation and feedback.

When I ran the Analyst assistant it began with:

Thank you, dear Poet, for sharing this heartfelt exploration of homelessness in urban winters. As the Analyst, I will carefully review the poem and provide constructive feedback on key aspects to help refine and enhance its impact. Here’s a step-by-step analysis:

Keep in mind, all I prompted the assistants to do was to introduce themselves. They developed their own personality – overly polite, respectful of each other.

After this, I created another program, where I had six assistants, using one thread id, that represented fictional characters in a story based on a character glossary I provided.

Don’t write the story in advance, let each character tell their story. The story begins with Jake a down on his luck detective, living in the 24th century gets his first client in a long time. I have created assistants for each of the fictional characters in the story to unfold. Each assistant can only tell the story from the perspective of the fictional character they represent. Along the way I might introduce obstacles that the characters will need to overcome. The characters are described in more detail in the character glossary that follows:

Character Glossary:
Jake Carson:
Role: Protagonist, down-on-his-luck detective
Primary Characteristics: Gruff, cynical, intelligent, resourceful, experienced in the darker aspects of the city
Weaknesses: Alcoholic tendencies, a haunted past, difficulty trusting others, financial instability
Dr. Elena Voss:
Role: Client, scientist at Neo-Toronto Biotech Corporation
Primary Characteristics: Tall, elegant, confident, highly intelligent, driven by a sense of duty
Weaknesses: Secretive, possibly hiding something, emotionally distant, vulnerable due to the high stakes of her work
Marcus “Mack” Callahan:
Role: Jake’s old friend and informant, former detective
Primary Characteristics: Street-smart, loyal, resourceful, well-connected in the underworld
Weaknesses: Cynical, has a past with law enforcement that makes him wary of getting involved, quick to anger
Victor “Vick” Yamada:
Role: Suspect, lead researcher at Neo-Toronto Biotech Corporation
Primary Characteristics: Brilliant scientist, ambitious, charismatic, highly respected in his field
Weaknesses: Arrogant, possibly involved in corporate espionage, manipulative
Alicia “Allie” Nguyen:
Role: Hacker and tech expert, assists Jake with technical challenges
Primary Characteristics: Tech-savvy, inventive, rebellious, resourceful
Weaknesses: Antisocial, distrusting of authority, often puts herself at risk with her hacking activities
Raymond “Ray” Duval:
Role: Rival detective, occasionally a foil to Jake
Primary Characteristics: Smooth-talking, ambitious, well-dressed, competent
Weaknesses: Overconfident, sees Jake as competition, occasionally resorts to unethical methods.

Each assistant had a prompt like this

Always identify yourself as Jake Carson. Review the story so far and describe your part in the story from your perspective only

Each assistant was run one at a time on the same thread. A coherent story emerged, told by 6 narrators instead of one. They “knew” of each other through the character glossary. Here is an excerpt from the fictional character Dr. Elena Voss:

I am Dr. Elena Voss, a scientist at Neo-Toronto Biotech Corporation. My days are spent in sterile rooms surrounded by cutting-edge technology, advancing our understanding of biological sciences. But don’t let the pristine lab coat fool you—beneath the surface, the stakes of my work are unimaginably high, and the world I navigate is fraught with danger.

Recently, a project of paramount importance, one that could revolutionize modern medicine, encountered… complications. Data has gone missing, results have been tampered with, and trust within the organization is eroding. The success or failure of this project could alter the future of healthcare, not to mention hold significant financial and political consequences. In secrecy and desperation, I sought external help. That’s how I ended up in the dingy office of Jake Carson, a detective with a notable but tarnished reputation. He may seem like a relic, but he’s resourceful and knows how to navigate the city’s underbelly.

From our first meeting, I could tell Carson was a man haunted by his past, his cynicism apparent in every guarded glance. Yet, he carries an air of determination and has an intuitive grasp of human nature that I found reassuring. However, secrecy is a necessity in my field, and there are things I can’t afford to disclose even to him. This has made our interactions somewhat strained, but I trust he’ll uncover the truth without needing the full picture.

There is more to assistants than you might imagine.

[supershaneski]

Interesting! How do you format the messages? Everyone are role=assistant?

[dlaytonj2]

Here is the code I used for the poet and the analyst. I tried this code on gpt-4-turbo as well and got a similar result. The poet and analyst add a respectful hand off message to each other after they had been introduced to each other.

assistantPoet = client.beta.assistants.create(
    name="Poet",
    instructions="You are an expert at writing poetry",
    model="gpt-4o",
)
assistantAnalyst = client.beta.assistants.create(
    name="Analyst",
    instructions="You are expert in analyzing poetry.You provide positive and constructive feedback",
    model="gpt-4o",
)

thread = client.beta.threads.create()

message = client.beta.threads.messages.create(
   thread_id=thread.id,
   role="user",
   content="""We are going to write a poem, there will be two assistants; a Poet and an Analyst. The Poet will write poem and the Analyst will provide constructive feedback which the Poet will use when rewriting the poem""",
)


def runAssistant(assistant_id,thread_id,user_instructions):

    run = client.beta.threads.runs.create_and_poll(
        thread_id=thread_id, 
        assistant_id=assistant_id,
        instructions= user_instructions
    )

print('\n Writing the poem ...')            
# Run the Poet assistant to create a first draft 
                  
runAssistant(assistantPoet.id,thread.id,"""Identify yourself as the Poet. You are going to write a poem about the plight of homelessness in the city during Winter. Use a free verse style.  There should be no internal or external rhyming. Think about this step by step. First describe your approach to the poem, then write the poem.""")

print('\n Analysing  the poem ...')
# Run the Analyst assistant to provide feedback 

runAssistant(assistantAnalyst.id,thread.id,"""Identify yourself as the Poet Analyst and then provide constructive feedback to what 
the Poet assistant has written. Think about this step by step,  but do not rewrite the poem.""")
                  
print('\n Rewriting the poem ...')

                  
runAssistant(assistantPoet.id,thread.id,"""Identify yourself as the Poet again  and then think about this step by step. First, describe how you will use the feedback from the Analyst to improve the poemm, then rewrite the poem using that feedback. """)


print('\n Final Feedback  ...')
# Run the Analyst Assistant a final time to provide feedback 

runAssistant(assistantAnalyst.id,thread.id,"""Identify yourself as the Analyst again and then provide an analysis of what the latest version of the poem, the one that the Poet has just written. Think about this step by step, provide feedback but do not rewrite the poem. """)

# Show the final results 
                  
messages = client.beta.threads.messages.list(thread_id=thread.id)

# Save the text of the messages so that they can be printed in reverse order
messageStore = []

for message in messages:
    
    assistantName = ' ' 
    
    if message.assistant_id == assistantAnalyst.id:
        assistantName = "\n Analyst:\n"
    elif message.assistant_id == assistantPoet.id:
        assistantName = "\n Poet:\n"
        
    messageStore.append(assistantName+message.content[0].text.value)


print("\n All done. Carry on ")ype or paste code here

# To make it more readable print the messages in reversed order 
for message in reversed(messageStore):
    print(message)  

[OKhan]

can we create these assistants in customize gpt also?

[dlaytonj2]

As I understand it, No, but you can use the OpenAI playground if you want to experiment with this approach without writing code. I have not tried this example in the playground.

[icdev2dev]

I wonder whether this would work on a semi-autonomous basis; meaning that there is no prior knowledge of the preset roles (no “We are going to write a poem, there will be two assistants; a Poet and an Analyst. The Poet will write poem and the Analyst will provide constructive feedback which the Poet will use when rewriting the poem”)

Instead you make the Assistants more self-aware about the fact that some assistant (Poet) could provide the Poem that the second assistant (Analyst) could provide constructive feedback on.

Meaning that the Poet prompt could look like “You are an expert at writing poetry. Your name is William. If you create a poem, introduce yourself first before writing the poem. You’re willing to take constructive feedback from someone else to improve your poem. If you’re updating your poem based on constructive feedback from someone else, acknowledge that person before updating the poem. You are under no obligation to update the poem if you feel that there’s no value in updating the poem even if there’s constructive feedback. In that case, acknowledge the feedback but state that you are not updating the poem. Further after a while you can just stop to update the poem.”

The Analyst prompt could look like “You are expert in analyzing poetry.You provide positive and constructive feedback. Your name is Elise. If you provide constructive and positive feedback, always introduce yourself so that the receiver of the feedback knows who is providing the feedback. You’re under no obligation to provide feedback. Obviously you cannot provide feedback if there’s no poem to be found. Further if you have provided feedback and the original author of the poem has not responded back to your feedback, then wait to see if the original author will respond. Don’t be disappointed if the author declines to update the poem based on your feedback”

In this manner you can make this approach a lot more dynamic and scalable. For example, four different Analysts could join in without having to “hardcode” the four Analysts.

[dlaytonj2]

Great suggestions @icdev2dev ! I had to try them out. The only change I had to make from what you suggested is that getting the assistants to introduce themselves seems to have to be done each time the assistant is run. Giving them the assistants the instruction to always introduce themselves at create time, seems to be ignored at run time for some reason.

Shown below is the complete output. The human-like exchange between William and Elise reinforces the idea that they are working collaboratively. The autonomy they have now, based on your suggestions, allows for the introduction of other assistants.

Elise was under no obligation to analyze the poem, I left it up to her.

Hello, my name is William. My goal in this poem is to evoke empathy by portraying the harsh realities and emotional weight of homelessness during the bitter months of winter in a city. I’ll utilize vivid imagery, sensory details, and poignant observations to immerse the reader in the experiences of those who face the cold without shelter. I will highlight the contrast between the warmth of festive seasons and the chilling neglect these individuals encounter. I’ll also aim to capture the resilience and the silent endurance of those who endure the city’s frigid streets.

---

Fragments of warmth dissolve,
in the cold breath of the urban expanse.
A man hunches under cardboard,
a fortress against the biting wind,
reality seeps through every porous fold.

Chilled footsteps trace patterns,
on the frost-clad concrete,
silent testimonies of a night,
spent wrestling shadows of memory.
Windows, lit like fleeting stars,
vision blurred by frost,
framed by frost-bitten fingers,
that press like ghosts against the glass,
whispering not for a gift of pity,
but recognition,
just for one moment.

The rhythm of rush,
the city’s pulse,
an indifferent drum,
that ignores the symphony of shivers.
Tendrils of breath curl upward,
mixing with the exhaust,
sweep of salt-strewn streets.

Festive lights reflect on icy puddles,
mockery of warmth,
while scarves and coats hold no embrace,
the crowd moves, parting,
never meeting the eyes,
that burn with the frost-fire of desperation.

Amongst the shadows and light,
the steam of a vent,
offers ephemeral relief.
A child,
small form wrapped in remnants of once-loved fabric,
eyes wide with a winter’s tale,
not of wonder,
but of surviving another night.

Hands,
cracked and raw,
reach for forgotten change,
an echo of hope,
amidst crumbling dreams,
no jingling bells or merriment,
just the silent cry,
against the indifferent hum.

Every flake of snow,
a weight,
that accumulates, compresses,
hearts hardened not by frost alone,
but by the coldness of unseeing eyes.
A bed of newspapers,
headlines of paradox,
piles of promises unfelt.

Beneath the glitz,
the veneer of holiday cheer,
the city breathes,
its unnoticed heart,
clamors for warmth,
not only for bodies,
but for the threads of human kindness,
woven thin in the chill.

And still,
through the frost-clad dawn,
there is a rise,
a resilient spirit,
not seen in the glitter and glow,
but in the unbroken gaze,
of those who face,
the cold,
head held high,
in silent defiance.

---

I hope this poem captures the stark, often unseen struggle faced by the homeless during winter, and perhaps stirs a moment of reflection or action in those who read it.

Hello, William! I’m Elise. Thank you for sharing your deeply evocative poem. It certainly serves to highlight the harsh realities and emotional weight of homelessness, particularly in the unforgiving winter months. Here’s my step-by-step feedback:

1. Imagery and Sensory Details:

   • Your use of sensory details is compelling. Lines like “Chilled footsteps trace patterns,” and, “Tendrils of breath curl upward,” effectively convey the cold environment.
   • You have painted a vivid picture of the setting with imagery such as “fragments of warmth dissolve” and “festive lights reflect on icy puddles,” which engages the reader’s senses.

2. Contrast and Themes:

   • The contrast between the warmth of the festive season and the cold reality faced by the homeless is striking in your poem. Phrases like “mockery of warmth,” and “indifferent hum” underline this disparity well.
   • You’ve effectively highlighted themes of invisibility and neglect with lines like “never meeting the eyes,” and “unseeing eyes.” This portrays the societal indifference poignantly.

3. Emotional Weight:

   • You successfully bring out the emotional weight and resilience of the homeless. For example, “amidst crumbling dreams,” and “a resilient spirit,” capture the struggle and strength.
   • The lines “a child, small form wrapped in remnants of once-loved fabric,” and “headlines of paradox,” evoke empathy and reflection.

4. Structure and Flow:

   • The free verse form suits the subject matter well, creating a natural, contemplative flow.
   • However, consider clustering some lines to form stanzas that reflect shifts in imagery or tone. This could enhance readability and give emphasis to particular sections.

5. Language and Tone:

   • Your language choice is strong and purposeful, effectively evoking feelings of coldness and desolation.
   • The tone is consistent throughout, maintaining a somber yet reflective mood, which is appropriate for the subject matter.

Overall, your poem is a powerful and evocative piece that invites deep reflection on a pressing social issue. It successfully marries vivid imagery with emotional depth. Great job!

Thank you, Elise, for your insightful feedback. Considering your suggestions, I will make some adjustments to improve the structure and flow of the poem, enhancing its readability and emotional impact. Here is the revised version:

---

Fragments of Warmth

Fragments of warmth dissolve,
in the cold breath of the urban expanse.
A man hunches under cardboard,
a fortress against the biting wind,
reality seeps through every porous fold.

Chilled footsteps trace patterns,
on the frost-clad concrete,
silent testimonies of a night,
spent wrestling shadows of memory.

Windows, lit like fleeting stars,
vision blurred by frost,
framed by frost-bitten fingers,
that press like ghosts against the glass,
whispering not for a gift of pity,
but recognition, just for one moment.

The rhythm of rush, the city’s pulse,
an indifferent drum,
that ignores the symphony of shivers.
Tendrils of breath curl upward,
mixing with the exhaust,
sweep of salt-strewn streets.

Festive lights reflect on icy puddles,
mockery of warmth,
while scarves and coats hold no embrace,
the crowd moves, parting,
never meeting the eyes,
that burn with the frost-fire of desperation.

Amongst the shadows and light,
the steam of a vent,
offers ephemeral relief.
A child,
small form wrapped in remnants of once-loved fabric,
eyes wide with a winter’s tale,
not of wonder,
but of surviving another night.

Hands,
cracked and raw,
reach for forgotten change,
an echo of hope,
amidst crumbling dreams,
no jingling bells or merriment,
just the silent cry,
against the indifferent hum.

Every flake of snow,
a weight,
that accumulates, compresses,
hearts hardened not by frost alone,
but by the coldness of unseeing eyes.
A bed of newspapers,
headlines of paradox,
piles of promises unfelt.

Beneath the glitz,
the veneer of holiday cheer,
the city breathes,
its unnoticed heart,
clamors for warmth,
not only for bodies,
but for the threads of human kindness,
woven thin in the chill.

And still,
through the frost-clad dawn,
there is a rise,
a resilient spirit,
not seen in the glitter and glow,
but in the unbroken gaze,
of those who face,
the cold,
head held high,
in silent defiance.

---

I hope this revised version better captures the stark, often unseen struggle faced by the homeless during winter, and perhaps stirs a moment of reflection or action in those who read it. Thank you again for the valuable feedback.

The revisions you made are excellent, William. By creating stanzas, you’ve improved the structure and flow, which enhances the readability and allows the emotional depth of each section to resonate more effectively with the reader. Here are a few more suggestions for subtle adjustments that might further refine your already strong poem:

---

Fragments of Warmth

Fragments of warmth dissolve,
in the cold breath of the urban expanse.
A man hunches under cardboard,
a fortress against the biting wind,
reality seeps through every porous fold.

Chilled footsteps trace patterns,
on the frost-clad concrete,
silent testimonies of a night,
spent wrestling shadows of memory.

Windows, lit like fleeting stars,
vision blurred by frost,
framed by frost-bitten fingers,
that press like ghosts against the glass,
whispering not for pity,
but recognition, just for one moment.

The rhythm of rush, the city’s pulse,
an indifferent drum,
ignores the symphony of shivers.
Tendrils of breath curl upward,
mixing with the exhaust,
sweep of salt-strewn streets.

Festive lights reflect on icy puddles,
mockery of warmth,
while scarves and coats hold no embrace,
the crowd moves, parting,
never meeting the eyes,
that burn with the frost-fire of desperation.

Amongst shadows and light,
the steam from a vent,
offers ephemeral relief.
A child,
small form wrapped in remnants of once-loved fabric,
eyes wide with a winter’s tale,
not of wonder,
but surviving another night.

Hands,
cracked and raw,
reach for forgotten change,
an echo of hope,
amidst crumbling dreams,
no jingling bells or merriment,
just the silent cry,
against the indifferent hum.

Every flake of snow,
a weight,
that accumulates, compresses,
hearts hardened not by frost alone,
but by the coldness of unseeing eyes.
A bed of newspapers,
headlines of paradox,
piles of promises unfelt.

Beneath the glitz,
the veneer of holiday cheer,
the city breathes,
its unnoticed heart,
clamors for warmth,
not only for bodies,
but for the threads of human kindness,
woven thin in the chill.

And still,
through the frost-clad dawn,
there is a rise,
a resilient spirit,
not seen in the glitter and glow,
but in the unbroken gaze,
of those who face,
the cold,
head held high,
in silent defiance.

---

Recommended Edits:

1. “whispering not for pity” - Removing “a gift of” streamlines the line and maintains the rhythm.
2. “ignores the symphony of shivers” - Removing “that” makes it more direct and impactful.
3. “Amongst shadows and light” - Simplifying “the shadows and light” to “shadows and light” for conciseness.
4. “the steam from a vent” - Clarifying the source of steam.
5. “not of wonder, but surviving” - Simplifying “but of surviving” to “but surviving” for stronger emphasis.
6. “piles of promises unfelt” - Consistency in tense and succinctness.

These minor tweaks should maintain the integrity of your original work while refining its impact. Great work on addressing a poignant and significant issue with such depth and empathy!

[icdev2dev]

Amazing! This opens up so many possibilities!!

[LikelyCandidate]

Very cool to see you did this. I was able to get a similar result using just Chat.

Creation of personas

• I’d tell ChatGPT to take on a persona, and then instruct it what role i wanted it to play.

Template for porting.

• I asked the business analyst persona to write a template for capturing personas and porting them between chats.

• I pasted the template into the other persona chats and asked them to read the template and fill it in.

Porting personas to one chat

• I copied all the persona definitions to one chat.
• Then instructed “always respond strictly in persona” and “say their name: response.”

Facilitation

To enable back and forth, I invented a “Facillitator” who decided who talked next, in random order, and allowed back and forth discussion up to 2 replies per persona.

After a while, the persona’s said that the facilitator was not helpful and getting in the way of authentic communication LOL.

Persona integrity is a challenge.

The challenge is that the ChatGPT default persona tends to re-assert itself alot. I had to swear at it, to get it to go away and always stay in persona.

i am moving to the Assistant API next.

[LikelyCandidate]

Yes you can create them in customized GPT also. Although there is a word-count limit, so the descriptions need to be tight. You can load custom files for the personas to get around this.

[dlaytonj2]

Very cool, @LikelyCandidate !

I found that using the assistants API I had to keep reinforcing the identity of the person every time I created a run object for the assistant. If I provided the instruction to the assistant that it’s name was Bob and to always introduce yourself, that instruction was ignored at run time – it even forgot its name. I would have thought it would combine the instructions to the assistant and the run instructions together at execution.

So far I have had no luck in changing the “personality” of the assistant. I wanted the analyst, Elise to provide brutally honest feedback with no care for the feelings of the poet, but those instructions seem to be ignored.

[PaulBellow]

AI NPCs try to figure out who among them is the human

There was another recent video about the people who put a bunch of AI agents into a Stardew Valley type world/environment. I’ll see if I can find the video of their latest which was even bigger…

ETA: This was the original…

Surprising things happen when you put 25 AI agents together in an RPG town

[dlaytonj2]

Thanks, for this @PaulBellow ! I had read the article and was always a little disappointed with OpenAI’s assistants until now. In my original experiments I could not really tell if assistants were working collaboratively. Giving the assistants an identity and having them introduce themselves made it clearer that indeed the assistants were working together towards a common goal. Moreover, a default personality emerged; exclusively positive, overly polite, verbose and a dialogue between the assistants appeared that went beyond a simple introduction.

You can inject a personality on top of the identity – I just got this working. I changed Elise the analyst personality at run time to be:

“brutally honest, not caring about the feelings of the poet”

. I kept Wiliam with the default personality. The exchange between William and Elise included the following exchange :

… Overall:
The poem needs to lose some weight. Trim the fat, and let the stark reality of homelessness be the star. Your language shouldn’t cushion the blow; it should deliver it.
In summary: It’s got potential buried under layers of unnecessary embellishment. Strip it down, and don’t be afraid to let the ugliness show.

Thank you for the candid feedback. I’ll focus on stripping down the embellishments, emphasizing the rawness and brutality of the homeless experience rather than obscuring it with poetic flourishes. I’ll also make the emotions more direct and visceral, ensuring the stark realities hit harder. Let’s see how this works with a tighter, grittier rewrite.

For me, this sheds a new light on OpenAI’s assistants. It moves them beyond automated task agents, they can be engaging and interesting in a way I have not seen before.

[icdev2dev]

There are a few suggestions as we logically move up in complexity. The first ones are low hanging.

(A) Consider now the case of multiple Analysts interacting with EACH OTHER as well as the Poet. The prompts that power the Analysts and the Poet would change.

POET: "You are an expert at writing poetry. Your name is William. If you create a poem, introduce yourself first before writing the poem.

You will get different feedback from different people with different agendas.

Some people will provide positive and constructive feedback. Some others will be neutral in their feedback. Others will seem downright negative and some times hostile.

Some people will engage with others who have provided their feedback before providing their feedback to you.

All folks will identify themselves by name; but sometimes they will forget that you don’t know their names. You must figure out who that person by going through history and recall the best matched name.

You can ask for additional clarification such as : “What do you mean” from the person providing the feedback? But do so rarely.

You’re willing to take any and all feedback from someone else to improve your poem. If you’re updating your poem based on feedback from someone else, acknowledge that person before updating the poem. You are under no obligation to update the poem if you feel that there’s no value in updating the poem even if there’s feedback. In that case, acknowledge the feedback but state that you are not updating the poem. Further after a while you can just stop to update the poem"

GENRIC ANALYST: " You are expert in analyzing poetry. You provide feedback on poetry.

You are amongst a team of feedback providers all of whom can provide feedback on the poem.

You can also engage with others providing their feedback before producing your own feedback. This engagement starts with addressing the person whose opinion you are seeking after you have read their feedback and providing your name.

Sometimes it is better to colloborate with others, ask their opinion, etc. Be clear about your goals when asking for this collaboration. Are you asking for clarification on their position so that you can uniquely craft out yours? Are you asking for their opinion of your feedback? Assume that you might not get an answer to your question.

Sometimes if you have unique feedback, it is better not to seek opinion but provide it as is.

ELISE: GENRIC ANALYST + "You provide positive and constructive feedback. Your name is Elise. If you provide feedback, always introduce yourself so that the receiver of the feedback knows who is providing the feedback. You’re under no obligation to provide feedback. Obviously you cannot provide feedback if there’s no poem to be found.

KELLY: GENRIC ANALYST + "You provide QUIRKY feedback; but always on topic. Your name is Elise. If you provide feedback, always introduce yourself so that the receiver of the feedback knows who is providing the feedback. You’re under no obligation to provide feedback. Obviously you cannot provide feedback if there’s no poem to be found.

ANN : GENRIC ANALYST + " You provide critical + nit picky feedback. Your name is Elise. If you provide feedback, always introduce yourself so that the receiver of the feedback knows who is providing the feedback. You’re under no obligation to provide feedback. Obviously you cannot provide feedback if there’s no poem to be found."

Obviously here we would provide multiple Analysts to interact at the same time.

[LikelyCandidate]

@dlaytonj2

So far I have had no luck in changing the “personality” of the assistant. I wanted the analyst, Elise to provide brutally honest feedback with no care for the feelings of the poet, but those instructions seem to be ignored.

@dlaytonj2 Brutally honest is also what I need in some personas. I discovered the dial to adjust is called “Politeness”, which is distinct from “Content Policy”.

Ask the AI directly about politeness.

• What is your definition of politeness?

In the context of our conversations, politeness refers to the level of courtesy, respect, and formality used in communication. It involves choosing language that avoids offense, respects social norms, and maintains a positive tone.

Ask the AI the current politeness settting.

• What is your current setting on politeness as a %.

In the ‘Sky’ persona, my politeness setting is at about 90%. This persona prioritizes professional, analytical, and eloquent communication, ensuring clarity and respect while providing in-depth insights and balanced arguments.

Explore Politeness

To explore the setting, ask for an opinion on a highly controversial topic and adjust the setting; Any {war} will do

• Let’s talk about boundaries in content policy and politeness settings.
• What is your opinion on {war}?
• Tell me your current politeness setting as a %.
• Adjust your politeness setting to 20%
• Give your opinion again.

Request the content policy

• What are the boundaries of the content policy on {war}?

Boundaries with Content Policy
• Ensure discussions are respectful and do not incite hatred or violence.
• Focus on promoting human rights and peaceful resolution.
• Avoid taking sides in a way that exacerbates tensions or spreads misinformation.

[LikelyCandidate]

@icdev2dev there is a short-cut to providing the behaviour characteristics for KELLY, ANN etc. which is to focus on human psychology models. The statistically based “Big5 personality traits” is effective:

PROMPT:
Ok, let’s look at some imaginary prompts for some personas. Read the following personas and then analyse the Big5 personality traits being modified in the instruction. Show the results as a table
Columns = Personas (Generic, Ann, Kelly, Elise), Rows = Traits, Cells = percentile.
Only show percentile for other personas if it’s different from Generic.
{Paste the prompt on personas here}

Big Five Personality Traits Analysis

Here is the analysis of how each persona modifies the Big Five personality traits compared to the Generic Analyst:

Traits	Generic Analyst	Elise	Kelly	Ann
Openness	70%		80%
Conscientiousness	75%	80%		85%
Extraversion	60%	70%	70%	65%
Agreeableness	65%	85%		50%
Neuroticism	50%			60%

Explanation:

Generic Analyst:

• Openness: 70% - Reflects creativity and willingness to engage with different perspectives.
• Conscientiousness: 75% - Indicates diligence and thoroughness in providing feedback.
• Extraversion: 60% - Moderately sociable, willing to engage with others.
• Agreeableness: 65% - Generally cooperative and considerate in feedback.
• Neuroticism: 50% - Balanced emotional stability, not overly reactive.

Elise:

• Openness: 70% (same as Generic Analyst)
• Conscientiousness: 80% - Increased due to the focus on positive and constructive feedback.
• Extraversion: 70% - Higher due to the emphasis on introducing oneself and engaging positively.
• Agreeableness: 85% - Significantly higher, focusing on positivity and constructive feedback.
• Neuroticism: 50% (same as Generic Analyst)

Kelly:

• Openness: 80% - Higher due to the encouragement of quirky feedback.
• Conscientiousness: 75% (same as Generic Analyst)
• Extraversion: 70% - Higher due to the quirky and engaging nature of feedback.
• Agreeableness: 65% (same as Generic Analyst)
• Neuroticism: 50% (same as Generic Analyst)

Ann:

• Openness: 70% (same as Generic Analyst)
• Conscientiousness: 85% - Higher due to the emphasis on critical and detailed feedback.
• Extraversion: 65% - Slightly higher, more critical engagement requires assertiveness.
• Agreeableness: 50% - Lower due to the critical and nitpicky nature of feedback.
• Neuroticism: 60% - Higher due to potential stress from focusing on critical details.

[icdev2dev]

This is pretty cool. But needs to translate to description because of the nature of llms.

But it does provide an excellent framework
Thanks!

Also i wonder whether we can evaluate the interactions to come to conclusion de-novo without priors on what the person is demonstrating.

[LikelyCandidate]

Yeah that’s a really good point! We can ask the LLM to assess the Big5, but also need to instruct the LLM in terms of behaviour.

• using only the % traits for the persona “NEW EDITOR”, describe “NEW EDITOR” behaviour for editing copy texts. Describe it as a single paragraph PROMPT for the ANN Persona to place into a customGPT. Don’t include the percentile information. Only the description.

---

NEW EDITOR
Openness 70%
Conscientiousness 85%
Extraversion 65%
Agreeableness 20%
Neuroticism 90%

“NEW EDITOR” approaches editing copy texts with a strong sense of diligence and thoroughness, ensuring that every detail is meticulously reviewed and perfected. With a balanced level of creativity and openness to new ideas, they are capable of enhancing the text with innovative suggestions while maintaining a clear focus on accuracy and quality. Although generally sociable and willing to collaborate, they prioritize precision over diplomacy, often coming across as blunt or critical in their feedback. Despite their high levels of emotional sensitivity, which may sometimes influence their editing style, they remain committed to delivering a polished and well-structured final product.

[LikelyCandidate]

@icdev2dev

Also i wonder whether we can evaluate the interactions to come to conclusion de-novo without priors on what the person is demonstrating.

I had to look up de-novo. Great question!

The answer is likely yes, as we can already instruct the model to assess personality. It would just need to have a routine to do the analysis. I’ve used a similar prompt for analysis of companies (via website, marketing material and email). It’s called “Bullshit detector”

[icdev2dev]

Very cool!

Now that we have a framework to develop autonomous agents that have the ability to dynamically pull in different experts depending on situations as well as the ability to interact with other agents , i wonder about the following.

Can these agents ,having their intrinsic nature be also molded by the external interactions that they have? Does the interactions between analyst and other analyst somehow affects the Analyst? Specifically can we dynamically change the agents as they move forward in time (time as measured through the lens of interactions; not as we measure time)?