Why do LLMs make you more creative?
Why is it that access to large language models like Claude 3.5 Sonnet and GPT-4o improves creativity? What is the mechanism by which LLMs allow you to dream bigger about your ideas? Maybe I’ve sneaked in an extraordinary claim—that having access to a large language model can make a scientist more creative, and more productive—but I’d like to back up the claim with a concrete mechanism. So, why do LLMs make you more creative?
Here is why. It’s because when you are engaged in a creative endeavor—such as scientific discovery, or creating music, or writing code—your ability to imagine a new idea, a new riff, or a new function is fundamentally limited by what you are capable of.
In the case of jazz, the sounds you’re physically able to produce on your horn are the only sounds you’ll reach for in a session. If you are a musician, your imagination, the songs you hear in your head, the things you think you make up when you’re improvising on the bandstand, the music that you make is fundamentally limited by your ability to execute that music. If you can’t play like Coltrane, you will not imagine music like Coltrane's music in your head, and you will not create music like that. Coltrane spent endless time working on his chops. This is why we first must be trained on the fundamentals of our instrument before we can use it to make music. Your internal creativity is fundamentally limited by your ability to execute the things you imagine in the real world.
Another way of thinking about that, is every time you learn a new trick on your instrument, or learn a new way of coding, or a new mental model, it expands your internal ability to imagine. Your internal world now includes that ability, and you can imagine using it on problems inside your brain. Which is what creativity is. So by learning new abilities, you enlarge your creativity.
So how does access to a large language model have a similar effect on creativity? As an example, consider writing software. You know the model is an extremely capable coder. You can ask it to implement things that you may not be able to implement, or things that may "take you a long time", or things that will likely take you a lot of annoying debugging which you don’t wanna do, or things that you just straight up don't know how to build. All you have to do is be very precise about what you want. This process creates a change in your own imagination. Because, internally, you become a better coder. It doesn't matter that your own hands aren't typing. What matters is that your creative ideas are transformed into reality.
So I think the mechanism by which large language models enhance creativity and effectiveness, and will advance science and creative arts, is by expanding the internal worlds of scientists, engineers, and artists. LLMs let you dream bigger.