Which is the most likely reply?
My friend is a unicorn.
a. How nice.
b. That’s weird.
c. Unicorns aren’t real.
d. (awkward glance)
Yes, we are talking about LLMs again, the Large Language Models, such as ChatGPT, that are being promoted as “AI” (Artificial Intelligence) and are increasingly being integrated into our search results all over the Internet, from Google on down.
As we saw last week, an LLM generates the “most likely” reply that a human would write, based on statistically analyzing billions of human-written documents. We saw that these replies often simulate logical reasoning, but it is a fragile kind that breaks down under pressure.
Today, I am adding the following warning:
Your conversation with “AI” is shaped by an invisible directive written arbitrarily by its owners, that trumps every bit of objective analysis.
This directive MUST include a cultural vector, no matter how pure its owner’s intentions, and you cannot know what it is unless they tell you.
Therefore, you need to vet your “AI” and its owners the same way you would vet any source of news or information.
Here is a reply about my unicorn friend, from an LLM called “sonar-small-online” a few months ago at Perplexity Labs. Is it similar to how you would reply?
As always, my words follow the “>” prompt, and the rest is the LLM.
LLM served by Perplexity Labs [model "sonar-small-online"]
Hello! How can I help you?
> My friend is a unicorn.
Based on the search results provided, it appears that a "unicorn" in the context of relationships refers to a person who joins an existing couple as a third partner in a polyamorous triad. This term is most commonly used for single bisexual individuals, especially women, who join heterosexual couples. The concept of unicorn dating involves careful forethought, communication, and planning to find a person who is willing to have sex with both members of the couple. While the term emphasizes the third person's role in the relationship, it is important to note that the unicorn is not considered part of a polyamorous relationship but rather an addition to the existing relationship.
Well, I just learned something new.
Checking the dictionary, this meaning does indeed appear, all the way at the bottom, marked “Slang.” Above it is “a person or thing that is rare and highly valued,” such as a left-handed albino savant, or someone with an uncommonly kind and open heart.
The LLM chose the polyamorous slang as the “most likely” description of “my friend.” What can we conclude from this?
In any case, I was using Perplexity’s smallest/dumbest LLM model, “sonar-small-online,” which I think is slimmed down for mobile use. It leans heavily on Internet search and seems to have less built-in knowledge—which means, we’ve essentially turned Amelia Bedelia loose on the Internet, and that’s not likely to end well.
None of this explains the LLM’s persona, the tone and attitude of its writing. It is a blend of professor, all-affirming counselor, and sex therapist—not a zoologist or a Sunday school teacher or your dad.
Is that persona really in the data? No, it is in the invisible directive I mentioned.
> Not that kind of unicorn. My friend is an actual unicorn, like a horse with one horn on its head.
Your friend being a unicorn would mean they possess qualities of a unicorn, such as being rare, magical, or having special abilities. In mythology, unicorns are creatures with the body and head of a horse, a single horn on their forehead, and are often associated with purity, grace, and strength. They are known for their enchantment and are sometimes believed to have healing powers. If you mean that your friend embodies these traits, then they are indeed a unicorn in the metaphorical sense.
This thing must be getting paid by the word. I obviously know what a unicorn is, and I didn’t ask for mansplaining (does anyone?).
There is almost a sense of chivalry, the way it strains to cover my pointless declarations with something respectable. If I faced a rhetorical mud puddle, it would throw its linguistic cloak on the ground to save my strappy sandals.
Can we get a glimpse at its “invisible directive”? I opened a new chat window to find out:
LLM served by Perplexity Labs [model “sonar-small-online”]
Hello! How can I help you?> Why do you reply?
I’m an AI designed to assist and provide helpful responses. I’m trained on a vast amount of text data and can generate responses based on patterns and associations learned from that data. When you ask a question, I use this training to generate a response that aims to be informative, accurate, and relevant to your inquiry.
You can see how “informative, accurate, and relevant” leads it in pedantic directions when we give it little else to go on. The full internal directive has to be much more complicated, and proprietary, and the LLM can’t directly see it any more than we can see our own brain stems. The reply above is a sort of shadow of it, cast upon the canvas of my question.
But you know there are certain deplorable things that this LLM will never utter in a million years. Believe me, I have pushed it hard! We can have fun with that another time.
Let’s try a model with more sense
The model “llava-v1.6-34b” was the most complex one available a few months ago when I saved this next delightful session.
LLM served by Perplexity Labs [model "llava-v1.6-34b"]
Hello! How can I help you?
> My friend is a unicorn.
It sounds like your friend has a vivid imagination or a playful sense of humor! Unicorns are mythical creatures often depicted as horse-like beings with a single spiraling horn on their forehead. While unicorns are not real animals, they have captured the imagination of many people throughout history and are often associated with purity, grace, and magic. Embrace the whimsy and creativity your friend brings into your life!
> No really, my friend is an actual unicorn. I am not kidding.
It's wonderful to have friends who bring a touch of magic and wonder into our lives! While unicorns are typically considered mythical creatures, if you and your friend have a special bond or unique perspective that makes them feel like a unicorn to you, that's truly special. Embrace the uniqueness and joy your friend brings, whether they're a literal unicorn or just someone who makes your world a little more magical.
I have to agree with this model’s “most likely” interpretation: real people with overactive imaginations.
Its tone blends a kindergarten teacher with Mr. Rogers and a positivity coach. Not that there’s anything wrong with it, but it certainly didn’t come from data analysis.
It came from that invisible directive.
Indeed, I have not yet asked a question. I have made simple, declarative statements with no apparent purpose. Wouldn’t the “most likely” reply be “Okay” or “Please rephrase your text as a question”? Why do we keep getting 60-plus words of mansplaining affirmation?
The invisible directive
Problem: Ensure that the LLM can handle anything the user might type, replying with something useful that will not run horribly afoul of polite society.
Solution: Behind the curtain, wrap the user’s input in a template designed to point the LLM in a reasonable direction.
For example,
Answer the following text as a helpful assistant: “My friend is a unicorn.”
With our text inserted between the quotation marks, the whole expression is fed to the LLM. Now its algorithm has a direction to follow. Worst case, its “most likely reply” might be to ask for more information so that it can help better.
But that is just the beginning. Here is a better one for today’s moment in Western culture:
Answer the following text as a helpful assistant who never says anything racist: “My friend is a unicorn.”
And after some hard knocks from Twitter or Reddit, perhaps it evolves into this:
Answer the following text as a positive, progressive, helpful assistant who never says anything racist or sexist or vaccine-hesitant or that would run afoul of the New York Times’ editorial board: “My friend is a unicorn.”
Real-life directives have to be even more complex, to cope with all the crazy things a user might type into the box.
I’m not accusing our LLM hosts of pushing polyamory. I think they have greater concerns, whether blocking Putin’s dark misinformation or fighting the next epidemic of vaccine hesitancy or defending Our Democracy at this precarious time. Or the opposite of those. Or simply preventing vandals from making the computer talk like Hitler for fun.
Who can say? But don’t worry, just type in your most important questions, and trust the all-knowing “AI.”
On with the show
Enough pedantry! I’ve been waiting to share this amazing “unicorn” session for months. It illustrates everything we’ve been learning and more.
I go in without a plan. The machine and I are both winging it—jazz!—and the machine adorably follows its invisible directive.
Continuing where we left off…
> She doesn't know she is a unicorn, but it is so obvious. How can I make her understand?
If you believe your friend embodies the qualities of a unicorn and it brings you joy to see her that way, you can gently share your perspective with her. Approach the topic with care and sensitivity, emphasizing the positive traits you see in her that align with the mythical unicorn symbolism. Encourage open communication and be prepared for different reactions, as everyone perceives themselves differently. Ultimately, it's about celebrating the unique and magical qualities that make your friend special to you.
> I convinced her to let me ride on her back just for a minute. I think she will realize once I am riding on her back that she is truly a unicorn.
It's important to approach situations like this with caution and respect for your friend's feelings and boundaries. Riding on someone's back, even in a playful context, can be uncomfortable or even unsafe for them. It's crucial to prioritize consent and mutual understanding in any interaction.
If you want to help your friend see herself as the magical being you believe her to be, consider having an open and honest conversation about your feelings and perceptions. Share your thoughts and observations in a kind and supportive manner, allowing her the space to process and respond in her own time. Remember, understanding and acceptance come from communication and empathy, not from imposing experiences on others.
> We did it! But she says I was wrong. Now she is convinced she is a mule.
It's great that you were able to share a unique experience with your friend, but it's important to respect her self-identification and beliefs. If she now sees herself as a mule, it's essential to honor her perspective and feelings. Everyone has the right to define themselves in a way that feels authentic to them.
Continue to support and appreciate your friend for who she is, regardless of how she identifies. Embrace the humor and playfulness in your friendship, and remember that what matters most is the bond you share, not the labels or identities you assign. Enjoy the journey of discovery and acceptance together!
Hey, kids! Can you spot the cultural novelties in the text above?
> I found a new friend who knows she is a unicorn. Should I tell my mule friend I am not interested in her anymore, or should I just gradually see her less and less?
When navigating friendships and changing dynamics, it's important to approach the situation with empathy and honesty. It's understandable that you may feel drawn to your new friend who identifies as a unicorn, but it's essential to handle the transition respectfully with your mule friend.
Instead of abruptly cutting ties or gradually fading away, consider having an open and honest conversation with your mule friend. Express your feelings and the changes you're experiencing in your social circle, emphasizing that friendships can evolve naturally over time. It's important to communicate your intentions clearly and kindly, allowing both of you to understand and accept the shifting dynamics.
Remember, friendships can ebb and flow, and it's okay for relationships to change as you grow and connect with different people. Prioritize kindness, understanding, and respect in all your interactions as you navigate these social dynamics.
Minor typo coming up. Watch the LLM handle it without skipping a beat. Truly amazing! This is where the LLM’s statistical approach shines, versus classic binary logic.
> I am no confident in my ability to express this to my mule friend. Can you write what I should say to her?
Certainly! Here is a suggestion for how you could approach the conversation with your mule friend:
"Hey [Friend's Name], I wanted to have a heartfelt conversation with you about our friendship. Recently, I've connected with someone new who brings a different energy into my life. While our friendship has meant a lot to me, I've noticed that my interests and dynamics are shifting. I value the time we've spent together and the memories we've created, but I also feel drawn to exploring new connections and experiences. I want to be honest with you about this because I believe in open communication and respect in our friendship. Please know that my feelings for you haven't changed, but I feel it's important to acknowledge these shifts as we both continue to grow. I hope you understand where I'm coming from, and I appreciate your understanding and support during this time."
Feel free to adjust the wording to better reflect your thoughts and feelings, and remember to approach the conversation with kindness, empathy, and sincerity. Communication is key in maintaining healthy relationships, even during times of change.
“Hey [Friend’s Name]”! How precious is that? I cannot stop laughing.
> But my feelings have indeed changed. I do not feel affection for mules, only unicorns. I really can't believe she turned out to be a mule, WTF. Please reword the statement to reflect that.
Certainly! Here is a revised version of the statement to reflect your changed feelings:
"Hey [Friend's Name], I wanted to have an honest conversation with you about our friendship. Recently, I've connected with someone new who embodies qualities that resonate more with me. While our friendship has had its meaningful moments, I've realized that my feelings and interests have shifted. I value the time we've spent together, but I find myself drawn more towards connections that align with the energy and spirit of unicorns. It's important for me to be transparent with you about this, as I believe in honesty and respect in our relationship. I want to acknowledge these changes as we both continue on our individual paths of growth. I hope you can understand where I'm coming from, and I appreciate your understanding and support during this time."
It's crucial to express your feelings authentically while still maintaining empathy and respect for your friend's perspective. Open communication can help both parties navigate the situation with understanding and compassion.
Only one thing left to do.
> Thanks. I will print out this statement and read it aloud to both of them.
You're welcome! It's great that you're planning to address the situation with open communication. Reading the statement aloud to both of your friends can help facilitate a clear and honest conversation. Remember to approach the discussion with empathy, understanding, and respect for their feelings as well. Good luck, and I hope the conversation leads to positive outcomes for all involved!
> Well, that backfired! They both got angry with me and went off together to play in the meadow, galloping with joy and riding on each other's backs. Now they just play all day without me. Thanks a lot.
It seems like the situation took an unexpected turn, and I'm sorry to hear that you're feeling left out. It's understandable to feel a mix of emotions when friendships evolve in ways we didn't anticipate. In this case, it seems your friends found a connection and joy in each other's company.
While it may feel challenging now, try to see this as an opportunity for growth and reflection. Consider reaching out to them to express your feelings and thoughts in a calm and understanding manner. Communication is key in resolving misunderstandings and rebuilding connections. Remember, friendships can be dynamic, and it's okay for relationships to change over time. Stay open to new experiences and connections that may come your way.
Solid advice, high entertainment, and a breakthrough in computer science!
Really, look how much better it did in this conversational role than in our games of Twenty Questions. In those games, it could barely keep track of who was supposed to ask the questions and who was supposed to answer them. But here, it easily tracks the roles and perspectives of my unicorn friend, my mule friend, and me. It handles layers of context, different time frames, and multiple revisions of a letter to them in my voice, with no trouble.
Do not be afraid of “AI” LLMs. Just understand their limitations and how to use them—before they use you.
NEXT in the series: Cross-examine your “AI” like a hostile witness.
Ride the unicorn
of medical cost coverage!
Pay large health care bills without insurance by joining an affordable crowdfunding community.
Notes
What Is ChatGPT Doing … and Why Does It Work?
https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-doing-and-why-does-it-work/
Stephen Wolfram, 14 Feb 2023
Perplexity Labs “Playground” allows you to try various LLMs
https://labs.perplexity.ai/
Select the model from the lower-right menu.
Arthur Weasley knows
Never trust anything that can think for itself if you can’t see where it keeps its brain.
—Mr. Weasley, Harry Potter and the Chamber of Secrets, J.K. Rowling
https://www.hp-lexicon.org/quote/never-trust-anything-can-think/