You’re so predictable

Maybe language isn’t as deeply meaningful as we think. As LLMs grow more sophisticated, language is becoming more like a commodity than ever. Sitting in a lecture recently, a professor from NYU was discussing the current state of GenAI and LLMs, how they were developed, what they can offer us, and why they are powerful. LLMs are excellent at predictions, guessing the next word that would most likely appear based on the patterns it has seen in the past – but this begs the question, why are LLMs so good at that? Because humans and the way we use and interpret language is highly predictable. I’d argue that humans like to think they’re incredibly special, and one of the ways we’ve evolved to be so dominant on earth is through the power of language and communication. LLMs have stripped this differentiation away from us.

In this new age of LLMs, what makes humans special? I’d like to explore the implications of this commoditization of language and knowledge in three areas: creativity, individuality, and agency.

Creativity

I’ve been thinking lately about a photography class I took in college where we spent a week trying to unpack what creativity and originality means. We all like to believe we are original; that all the things we think and create are our own. In reality, all of our creativity is influenced by works of art, songs, or other mediums that we’ve seen before. Your own personal experiences, environment, and exposure shape who you are and what you think. It’s nearly impossible to create something so completely and utterly original. That’s why so many songs have familiar tunes and sequences of notes. We’ve all been listening to a song and thought, that sounds familiar, only to realize it’s a different song with nearly the same melody. Artists of all kinds have heard, read, seen, and experienced various forms of art that influence what they ultimately create. So is it really original? Or is it just a copy of something they’ve seen before? Is anything truly original?

While it can seem like LLMs are generating new threads or outputs, this is just on the surface. LLMs have to see something first - their output is not inherently unique and they tend to lack character. This is how they are designed, lacking adaptability and a four dimensional (or 13 dimensions if you’re into the quantum mechanics) model of the world. Everything is based on associations. They aren’t truly ‘creative’ in the same way we think about generating new art or art forms.

Individuality

Individuality is an interesting concept as it relates to generative AI models. As LLMs are updated and released, you could say that each iteration of an LLM is unique, technically having its own logical processing that’s different than the one before. The training data could be considered its ‘experiences’, influencing the content that it creates. It’s also learning a certain individual through interactions, adding particular characteristics to its output.

However, these models all start from the same base code. While it can cater itself to different people or ‘profiles’ that it interacts with, it ultimately starts from the same place. This is markedly different from humans. While even identical twins can grow up in the same environment and household, they turn out to be completely different individuals. They may have the same ‘baseline’ that a model does, but twins eventually have different thought processes and patterns that aren’t replicable. Their experiences in the world shape them into different human beings who have unique thoughts that could not be replicated. LLMs in the background are performing the same mathematical calculations based on their training, no matter the context you feed into it. There is no sense of individuality to a generative AI model.

Agency

This all leads me back to the question of what makes a human special – I’d like to think it’s our agency. AI and LLMs will eventually be better than us at many skills: math, knowledge retention, speed of output, etc. Their capacity is simply incomparable to what a human brain can retain. Their seemingly infinite memory, computational capacity, and processing speed are beyond compare. However, humans have the ability to make our own choices. We get to decide what influences we want to lean into when creating something original. We make decisions when we want, while LLMs produce the output from those directional decisions.

Creative agency is incredibly important now more than ever when I think about writing. LLMs are good at generating stories and narratives, but (at least for now) they don’t have the agency to direct what they produce - they must be prompted to guess the next best sequence of characters.

Looking Ahead

Humans are individually unique in a way that LLMs cannot emulate. They have inherent limitations based on the way they are designed as prediction models (primarily just computing dot products which is a conversation for another time). For now, I continue to hold the belief that a human’s experience in the real world and their agency will continue to be the differentiating factor, regardless of how much faster LLMs can process and predict language.

I’ve been reflecting on all of this as I begin this blogging journey. It would be incredibly easy to use LLMs to generate the content I share - write my thoughts and feelings in more coherent ways, edit my work into more grammatically correct structures, influence my ideas and conclusions. And while there are benefits to some of these actions, I feel more strongly than ever that is important to utilize the tools at hand with caution. Nothing is exciting about another essay written and edited by AI. I am confident there will be so much of this in our future. I want to maintain this space as a place for me to write my own original thoughts and feelings, even if it could be predicted with great accuracy by an LLM. A model can’t emulate my thoughts and feelings, my perspective on the world, the places and people I’ve interacted with that have left an impact. Only I can employ my agency and individuality to create written thoughts, and share them on my own timeline.