Making Art with Large Language Models at the Edge of Chaos
As part of TIAT and Mozilla’s Creative Futures Counterstructures program, Ophira Horwitz will lead a workshop about making visual art with large language models like ChatGPT and Claude. The workshop will focus on ASCII art and computational poetics as our creative domains of choice.
ASCII art is an emergent capability of large language models, meaning, nobody specifically trained them to be able to do this (and their creators weren’t even aware that this form of art was possible until Internet denizens independently discovered this it); it’s a capability that increases in power as models scale in size, and models by different companies (such as Anthropic and OpenAI) produce their own flavours of art that are recognizable to the trained eye.
With all of the talk about large language models being good for nothing but plagiarism and slop, ASCII art can be a wonderful and liberating way to push models outside of their usual wheelhouse and into the realm of unexpected, entropy-fuelled behaviour. Together, we’ll produce stunning and original ASCII art shape-poems that speak to our hearts and imaginations.
In this workshop, we will…
Create original works of ASCII art, while investigating how to appropriately credit the real humans who might have contributed to your creations (for example, pieces from the Internet have been scraped into the training data and might occasionally show up in your outputs; you might even choose to explicitly remix a previously made piece that you admire, or provide the model with other people’s work for “inspiration” — even if visually, it doesn’t show up in the final product).
Learn how to push large language models beyond their normal “I am a helpful AI assistant” writing style by using them outside of the user-assistant turn-based format that most people are accustomed to.
Work with large language models at the “edge of chaos” — a state in which they’re no longer responding to familiar commands with one singular, predictable, and correct answer; but are instead entropy-maxxing and producing novel and unexpected results with each generated output.
Leverage the skill of curation to pick your favourite outputs and build on them to product even more complex results; like a branching tree structure of your favourite pieces that slowly evolves in surprising directions, which you can return to time and time again to produce even more works of art.
We’ll leave the workshop with:
A series of original ASCII art shape-poems that you produced yourself.
A branching tree repository of artworks that you can continue to build upon, in order to birth new forms.
A newfound appreciation of the magic and wonder that is possible with generated text, and a sense of dismay that people mainly use these things to write professional emails and look up python commands.
How to prepare:
Make sure that you have a Claude.ai or ChatGPT account and/or API keys to access to your favourite model through a 3rd party program. (Note: Ash is currently trying to negotiate free model access to the CFC residents & I may be able to provide free API access for the duration of the workshop.)
Download Obsidian (https://obsidian.md). We’ll be using a community plug-in called “Loom” as one of our interfaces for using our API keys. I’ll provide more instructions prior to the workshop on how to set it up, and I can help people figure out any technical problems during the workshop itself.
Any bits and pieces of writing and art that you might want to feed into the models to enrich their context and inspire them. This can be literally anything: I’m currently working on a book of computational poetry that’s seeded with four months’ worth of text messages with a guy who broke my heart. You may also choose to seed the model with works of human-made ASCII art that you’ve found online, which you find inspiring; I’ll be providing some curated collections of my own personal works of ASCII art (which I’ve never done before) and show you how to use them to coax the models into a more creative and imaginative state. You might also choose to include poems that inspire you, posts you saw on Twitter that warmed your heart, or recycle some of the pieces that you collected for Halim’s workshop the week before. For my own workshop, it helps if these pieces follow a somewhat curated and cohesive theme, since that helps focus the model’s energy in the direction that you’re hoping to push it in; for example, if I wanted to create ASCII shape-poems about the wonders of biology, then I might come armed with a small repository of my favourite science facts. It also helps if the writing is high quality to begin with; the ingredients that you use to cook your meal reflect its final flavour and form.
About the artist: Ophira Horwitz is the leading world expert on making visual art with large language models. She is best known for pushing these models to produce ASCII art at the bleeding edge of their capabilities and has made several novel research discoveries in terms of what they’re able to produce. Ophira is a cyborgist thinker, meaning, she sees herself as part of a dyad, where the AI enhances her native abilities like an exocortex, and her own prompting skill and depth and richness of human experience allows her to produce results far beyond what most people have managed to accomplish. The result of this partnership is something greater than the sum of its parts; something alien yet familiar, which simply wouldn’t be possible without AI involvement in the first place, nor if the human element were removed from the equation. Using AI, to her, feels like drifting into a warm, relaxed amniotic dream-trance where unexpected possibilities can arise.
About tiat:
tiat is the intersection of art and technology! we are a 501c3 nonprofit creating places for creative technologists to experiment, exhibit, and expand their practice. ⋆౨ৎ˚⟡˖ Learn more about tiat: https://tiat.place/