Computational Intuition: Insights, References, and Learnings from Our Experiments with AI
At Imagination of Things, we have spent the last few months experimenting with a potential role for AI in the creative process. It started with a collaboration with Baltan Laboratories, who prompted us to create a digital tool for ideation and wordsmithing using semantic similarity and machine learning. Called Fabricating Alternatives, the research project allowed us to get lost and explore the realm of language, computational thinking, and weird algorithms, and we would like to share our unintended outputs, and what we have learned so far.
The initial premise was to develop narrative devices to inquiry about our reality, especially in the process of inventing, designing, developing long-term visions and near-future scenarios. Our working prototype uses machine learning as a provocateur, a multiplier of shared intuition about a certain topic.
An inspiring figure in this realm for us is Allison Parrish, a pioneer and educator spreading the possibilities around electronic creative writing — even before the fancy language models available right now. In the book ARTICULATIONS, for example, Allison crafts poems that are the output of a computer program that extract linguistic features from over two million lines of public domain poetry then trace fluid paths between lines based on their similarities. It demonstrates the intuitive coherence found outside the bounds of intentional semantic constraints.
One of the directions we are taking with verbal language in our practice is about designing systems that can recycle and upcycle key elements of language, foster or mimic agency with room to grow, adapt when provoked. What can emerge from this choreography? It feels like dancing around computational intuition.
It’s like a playground for cognitive glitches — when we face computational processes that do not produce predetermined results. Sometimes they feel like chaos-machines, but most importantly they are open, often generative, systems that offer a new infinite — sometimes messy, sometimes unreasonable, but always playful. Messiness is smarter than newness as Keller Easterling said in her book titled Medium Design.
Here we share some projects, experiments, and thoughts that have been populating our minds about this topic.
Fabricating Alternatives
The tool remixes submitted online texts from a workshop leader (e.g. a Wikipedia page or news article), generating words, concepts, and phrases around the topic that the workshop is investigating. Participants enter into the tool and react to prompts, questions, and suggestions which spark dialogue and ideas between group members. As they play, the machine learns more about the interests of the group, generating more specific sentences.
Disjointed Research Machine
Approached by the JRC/EU Commission, we engaged in an attempt to shift the original reason d’être of their database of science reports. We designed a playful experience infused by machine intelligence that unapologetically disjoints the data into pieces for more creative exploration and recombination. We called disjointed research the investigation and experimentation that embraces chaos, machine randomness, and intuition.
As a player, you set up a challenge for yourself, a mix of our preset (a tabloid headline, a slogan for EU 2030, conspiracy theory) with room for you to apply a sentiment, tweak & customize. The machine would then pick a series of quotes from the database using semantic similarity, and then a classic word game unfolds where you have to complete the challenge using those words.
In the end, using the same process underneath, the machine gives you a score (a combination of time spent, and comparison with classic examples that each preset challenge required).
GPT-2, our funky friend
Most of our current experiments and projects are using a language model called GPT-2, released by OpenAI. It was created scrapping massive content from the internet, only using pages curated/filtered by humans, mostly outbound links from Reddit.
The results are quite impressive, and with the right prompt, it has the ability to deliver coherent sentences. It’s also interesting that in the beginning, the full model wasn’t openly available, exactly due to its power and the researchers’ concerns about “malicious” applications.
A key inspiration was the work of Max Woolf who wrote a user-friendly python library wrapping GPT-2 and scripts for retraining GPT2 on custom datasets. This ability to retrain or fine-tune the algorithm to a specific domain was very useful to get GPT2 to follow specific types of text, from film scripts to poetry. We even used this ability to capture the essence of our own writing.
For example, by fine-tuning GPT2 on all of the text we had ever written, we could ask the algorithm to describe us as a company by supplying the prompt, ‘Imagination of Things is a …’ Some of the best include, “Imagination of Things is a creative studio based in Amsterdam which uses design as an ally.” All of the completions were taken from what it knows about our writing.
Italian Neoavanguardia
What we are ultimately interested in is the language system that can run with pseudo-autonomy, and our new role interacting with it. Me, You, and the Machine, almost like an improv game.
Language is extremely important to us. We are in no way attempting to automate creativity in written language through AI. Rather, we are more inspired by the Italian poet and novelist Nanni Balestrini. In the account by Bifo Berardi, “He is the first poet in the history of humankind who never wrote a single word. He refused the dirty work of writing words. He asked: Why should I do that? Why should I spend my time writing words? I’m a poet. I don’t write words. I take signs from the infosphere, from the daily conversations of people in the subway, from newspapers, from advertisements. His activity, he said, was to recombine. Recombination is also our task, and we should take a cue from him.”
Radical recombination infused by computational thinking: how can we actively (poking and customizing the system) and reactively (assigning new meaning, reacting, the improv) interact with it? Berardi’s follow up account has a precious definition: “Poetry is the consistent and intentional recombination of what exists, with the aim of creating what does not yet exist.”
We just expanded our own definition of poetry to embrace this hybrid possibility of language as a cultural trojan horse.
Just to conclude
Computational intuition is the bedrock of modernity, the idea that the only thing bigger than ourselves is hatred of human intelligence. Evolution is over, and humans are just another part of it. In a world in which intelligence is so common that society is so monolithic that even the most intelligent people we know are convinced they will never be able to fully master the machine, and the market will go bust.
Humankind is a weirdly recurrent mixture of all those things that are weirdly easy to weirdly delete.
This last paragraph was created by a machine trained on some of the influences & references of this article.
“Knowledge is often produced at the edges or in the gaps of not-knowing.”
Markus Miessen, Crossbenching (2016)
