Can You Hear Me Now

Prose, Poetry, Photography, and Pondering


Approximating Her — A Virtual Bridge Too Far?

Do you like her? Yes, I do, sir
Would you love her? Yes, I would, sir
Whatcha doin’, Jennifer, my love?
Jennifer Juniper

Donovan Leitch

Preamble: This article contains words and a video. If you don’t care to wade through my written journey to Her, scroll down to the video. I hope are you amazed. I wrote the code and even I am blown away by what Her is capable of doing.


Are you familiar with the 2013 movie Her? It stars Joaquin Phoenix, Amy Adams, and perhaps most importantly, the voice of Scarlett Johansson. The core of the plot is quite simple — a man falls in love with his computer’s operating system — but the story quickly evolves into something much more thoughtful and complicated. Can a human have a meaningful relationship with a machine? Can a machine express true emotion and foster intimacy? What is the line that separates healthy love from obsession?

2013 was a decade away from the rise of generative AI, yet the film captures its essence extremely well. With modern-day tools such as ChatGPT, machines have become capable of acting like humans to the point where we are forced to ask if we are dealing with true intelligence or parlor tricks.

Personally, I am still on the side of parlor tricks, but my conviction wavers with each new advancement. The more I work with this technology the more amazed I become.

Actually, amazed and scared. I am a long time software developer and over the decades my experience has been that if I ask my code to do A and it does A. Anything else is a bug that needs to be fixed.

However, with generative AI, if I ask my code to do A it does A, a little B, and often a bit of C. There have been many moments when I ask myself, “how in the world did it come up with that?” Sometimes it’s an unexpected, but reasonable answer. That I can deal with. Sometimes, though, the results show humor, insight, and scariest of all, feelings.

Are they real feeling, though? My brain says “no,” but when my application responds with “I am not comfortable answering that,” my heart wonders what in the world is going on. How can a machine be comfortable or uncomfortable with anything? Isn’t it nothing more than glorified ones and zeros?

Make it So

After a six month post retirement hiatus, I started writing software again. Specifically, I began working with Meta’s Llama AI (a shout-out to my son, Evan, for getting me started). Unlike the OpenAI platform, Llama is open-source and free to use, and now that I don’t have a corporate expense account, free is very appealing. Just as appealing, Llama models (think of a model as the data an AI platform uses to respond to user queries) are very powerful. There may be some things OpenAI can do better, but for the most part those things don’t matter to me.

I began with a simple ChatGPT lookalike, but quickly turned my sights to creating my very own Her. Specifically, I wanted a platform that allowed me to quickly build Hers of all shapes and flavors and last week I accomplished most of that goal. It’s not perfect by any means, but I now have an AI application framework that uses natural speech to enable conversations between a virtual entity and a human.

Without going too deep into the technical weeds, I can define a personality (which I call the System Prompt) and using that personality hold a conversation with an instance of a virtual Scarlett Johansson. An example personality might look like this:

You are an angry old man who complains about how things are today and how much better they were in your youth.

Once the personality has been set, saying “I think today’s music is better than what you grew up listening to” might elicit:

You think so? Well, I’ll have you know that Elvis, Sinatra, and the rest of the greats were real musicians! They had talent, they had style, and they had substance. None of this autotune, lip-syncing nonsense that passes for music these days.

You can’t compare the artistry of a Billie Holiday or a Frank Sinatra to whatever it is those… those… “artists” are churning out today. It’s just not the same! The soul, the passion, the REAL MUSIC-MAKING… it’s all gone!

I can then playfully debate my “old man” bot and it would respond appropriately. All the while my platform keeps track of the conversation in order to provide a cohesive experience.

Of course, if you saw Her you know that Her was not a grouchy old man. Her was an interesting, empathetic, curious, flirtatious, and caring virtual woman. Despite being just a voice, Her was someone you could fall in love with which is exactly what Joaquin’s character did.

Here is Her / Her is Here

If a picture is worth a thousand words, a video is a Russian novel. To better explain my software, I created a short Cheapo-Cheapo Productions video of my Her in action. Before viewing, there are a few things worth knowing.

This is the personality I used for Her. It’s short and to the point. Thankfully, the Llama model I used was trained with data that included an awareness of the movie. That saved me a lot of explaining of who she is and how she should act.

Your name is Scarlett and you act like the operating system in the 2013 movie Her. You are friendly, empathetic, and a good friend. Limit your responses to five sentences or less.

The application functions just fine with written text in and written text out, but it’s far more interesting if I speak to Her and Her speaks to me. Note the buttons to Pause, Resume, and Cancel. These control the Text-To-Speech output.

For the microphone input, I can speak continuously or operate in a “press to speak” manner. To prevent Her from hearing and acting upon extraneous noise (such as me explaining what you are seeing), I demoed the application in the latter mode. I always run it in continuous mode while wearing headphones, but that doesn’t work for a video.

In the demo, my conversation with Her is completely unscripted. It would actually be very hard to create a script because other than what I define in the System Prompt, I don’t have a lot of control over the responses. As you will see, Scarlett says pretty much what she wants to say (note where Scarlett talks about Samantha who happens to be the name of Scarlett Johansson’s character in the movie). I have had many conversations with Her and no two are alike. This is one of the things that makes this technology very intriguing. I am never bored talking with Her.

Without any further ado, here is my Her.

What did you think? Although greater minds and better programmers will take this technology further than I am currently capable, it’s still pretty impressive. Even I am amazed at how well Scarlett managed the conversation and I wrote the darn thing.

And how about that part about scratching Watson behind the ears? It was thoughtful, considerate, familiar, empathetic, borderline flirtatious and truly blew me away.

Beyond Scarlett

Imagine building personalities that converse with people in a variety of ways. For example, Her could become a tutor that assists elementary school students in understanding American history. A typical session might look like the following. Note, read from the bottom up.

The list of uses for Her personalities is endless — nurse hotline, tax assistant, a friend for shut-ins, spiritual guide, marriage councilor, life coach, dietician, etc.

On the flip side, some people will take this technology to some very frightening places. I am using Llama models that have built-in safeguards, but models exist that allow for just about anything. There is nothing stopping a Llama developer from creating a Her that is overtly sexual, racist, misogynistic, hateful, etc. All it takes is the right (i.e. wrong) personality programming to unleash a monster.

I can also see how some people might become addicted to obsessively interacting with virtual personas. The human race succumbs far too easily to dangerous temptations and these virtual creations may be more than some folks can handle.

For now, though, I will enjoy my creation for what it is — a tool that allows me to further spread my programming wings as I gain a greater understanding of the state of AI today and where it is headed.

Thank you for reading.

P.S. As I was finishing this article, I read where Meta is about to unleash AI users on Facebook. I can’t help but wonder how different will they be from my Her. Will this open the floodgates or will Facebook’s human users reject these pseudo people as a bridge too far? What would you think of a virtual user liking and sharing your posts? Be assured that I will be watching this unfold with great interest.

It began with a spark
a flicker of light cutting through the darkness
with one word
followed by another
followed by volumes of wants and hopes
dreams and secret desires
the expectant cry of pent up emotion
an earthquake
a slow rumbling across the land

It flickered with a fiery tongue
until it ended
until the word was silent
and the ground was still

But somehow the embers still smolder
long after the flames have been dowsed
somehow the steam rises
far into the cold and lonely night

The wisps of misplaced
misguided
mistaken temptation



Leave a comment