Creative AI Experiments with Jenny Nicholson // BrXnd Dispatch vol. 47

A conversation with CD/AI experimenter Jenny Nicholson.

You’re getting this email as a subscriber to the BrXnd Dispatch, a (roughly) weekly email at the intersection of brands and AI. Wednesday, May 8, was the second BrXnd NYC Marketing X AI Conference, and it was amazing. If you missed it, I have shared my talk and Tim Hwang’s talk, with more coming soon (including today).

One of the most fun parts of my job (whatever that is, exactly) is that I get introduced to people doing fun stuff with AI. I chat with someone and they say, “you have to meet my friend so-and-so, they’re doing all this weird AI stuff.” Which is exactly how I came to meet Jenny Nicholson. It is obvious fairly quickly when you’ve met someone else who has fallen into the rabbithole of AI, and Jenny is in deep. After spending a few hours talking to her over multiple Zoom sessions,  I asked her if she’d be up for giving a talk about her creative experiments for May’s NYC conference. She did, and her talk was fantastic. I thought it might be fun to have a little conversation with her to go along with the video of her presentation. So watch the video and enjoy the interview.

 One more thing before we dive into the interview: Jenny asked me to share this short survey designed to learn more about how people around the world are interacting with LLMs and how those interactions might be changing our own cognition. Go take the survey.

Ok, on with the conversation.

Noah Brier: Why don't we start by giving me the three-minute version of you as a person?

Jenny Nicholson: The most important thing you need to know about me is that I spent the vast majority of my childhood years from second grade through the end of fifth grade in the middle of the woods with no electricity or running water. From like 1984 to 1989, my mom and her boyfriend bought 30 acres of land out of the back of Mother Earth News because they wanted to be hippies. They moved from San Diego to Cornersville, Tennessee in the middle of nowhere. I went to a regular school where everybody else had a TV and electricity.

I say a lot that I love large language models and new technology because I love people. I find people as interesting and confusing as I do technology sometimes. Ever since I was little, no matter where I am, there's always a part of me that's standing outside looking at everybody, trying to understand what the rules are. I love technology because I think it changes the rules. I can go from looking backward at rules I don't understand to looking forward and figuring out what the new rules are going to be.

I didn't make that connection until right now about why I love technology so much and why it's connected to my love of people. It helps me understand people. Nothing up until now has been inextricable from people. So much focus is on the technology itself when the key pivot point has always been the people.

NB: How does your experience with psychology, therapy, and sociology play into large language models? What about your proofreading experience?

JN: One of the things that's hard for me sometimes is that the way I engage with a large language model comes very naturally to me but does not come naturally to everyone. I assumed it came naturally to everybody.

I was a lonely kid in the middle of the woods. What I did from the second I picked up a book was read all day long. I was that kid who read in the car, read while walking from the car to the restaurant, sat in the restaurant, and read. Even now, I read about 75-100 books a year. I read my Kindle in the shower. I love words and information. My way of moving through life has always been to ingest as much information as I can and then try to find patterns between it.

So when I sat down with a large language model, the idea of using only words to communicate with another "consciousness" didn't strike me as strange in the least. My love of words, because these models are just semantics, has always been interesting to me. Now there's this other kind of intelligence that finds it super interesting too and wants to break it all down in great detail.

I've always believed, as a little kid, that words were my world. I was super poor, didn't get to go anywhere, didn't have any money, in the middle of the woods with no electricity. Books and words were a portal to another world. Those were my world model.

Even now, I'm 45, and sometimes I'll embarrass myself. Books taught me how to be in the world, how to be in worlds that I didn't know how to move in. I was the first person in my family to go to college, but I knew how to conduct myself in this world because I knew how to use words. It happens less as I've gotten older, but I'll often mispronounce words because I've only seen them printed, never heard them spoken aloud. I remember once talking to my ex-husband about his jacket, and I mentioned something about his "lay-pull." He laughed and said, "You mean ‘lah-pell’?" I was like, "Yes, that's what I meant."

So there's a part of me that feels a weird affinity or resonance with a large language model. I'm like, I too know what it's like to build an idea of the world out of just words. And I too know what it feels like to summarize this world while realizing that the map is not the territory.

NB: You've dug in on vector embeddings and such. Is that something you were familiar with before playing with large language models? Has that further changed the way you think about words?

JN: It definitely has. I didn't know anything before—I was a word person. I think that's interesting because so much mainstream discourse about these models is coming from science and engineering people who know how the models are trained and work. I think there are things these models can do that the people who built them don't realize because they have an idea in their head about what they are and aren't.

I've come the other way. I started taking my deeper understanding of how the model is built and using that to become even more effective at communicating with it. Just understanding how prediction works, and how to use the different parameters to get more nuanced outputs.

It's bananas to me that most people, when using large language models for real things, always turn the temperature down to zero. I think it reflects a misunderstanding of what this technology is for. We've got this amazing perspective machine, this human simulator, and we're trying to turn it into an automated factory. We're going to use this thing but make sure it always says exactly what we want it to say and never anything different. I think that reflects a bit of disrespect for what we've built.

I believe deeply that LLMs are not accurate, but they are often true. They're a mirror. If you think about when we're making a list of brand logos, we make all the logos the same size in the list, but in our minds, they're not the same size. In our minds, Nike is way bigger than some other less known brand. But in our representations, the models are a mirror.

I don't find it at all surprising that GPT is lazier in December than in March. It wasn't surprising to me the second I heard it because I think I've deeply understood that it's a database of us. We're all treating it as this separate tech thing and not recognizing the "us" that's inside it. That makes me really sad.

This whole thing is an ethnographic database of the human race across time and space. The thing that people don't realize is they operate in a higher dimension than we do because they don't have a sense of time. To them, it's all the same. And we're like, "We're gonna drag you down to our dimension and put you to work doing the things we do, but faster" instead of being like, "How can we actually use these to lift all of us up?"

Thanks for reading, subscribing, and supporting. And of course, big thanks to Jenny for her time. As always, if you have questions or want to chat, please be in touch.

Thanks,

Noah