AI: Mind the Gap

At the first brainstorm with the museum, a colleague and I wrote related ideas on sticky notes. One imagined an AI-written poem that visitors could print and take home. Mine imagined an AI with a creative will of its own, writing alongside you, like a typewriter that writes like Shakespeare. The two merged into Collaborative Poetry, which the design team developed into a quiet writing station where visitor and AI share one page.

In 2020 the most capable text model was GPT-2. I built an interactive prototype with a version fine-tuned on poetry: the visitor writes a line, the model answers, and they take turns. It was rough, but it convinced the museum the idea could work.

By 2021, early GPT-3 access through the museum made the writing far more natural. I used new prototypes to map what the model could reliably do: suggest a subject and mood, continue a poem, and title it when it was done.

Early flows assumed a limited AI, so they wrapped writing in a guided dialogue: questions, branches, and fallbacks. Once the prototypes showed what GPT-3 could do, the team cut the interaction down to a blank sheet of paper and a single button: “AI, write.” The simpler experience was also the more powerful one.

I built React prototypes for user tests to settle details that decide how collaboration feels, such as whether the AI should offer a line on its own and how long it should wait first.

A public writing tool also has to stay safe. I built moderation that screens both visitor and AI text, plus a review tool that lets museum staff remove anything that slips through.

The design team wanted the intimacy of real paper, a quiet contrast to the rest of the gallery. I built it with Unity’s cloth simulation. When a poem is finished, the AI suggests a title, signs it with the visitor, and sends it up to the curved display to join the stream.

Handwritten digits are the classic first dataset for neural networks, but they leave people out. Asking visitors to draw a face instead makes the network’s task a human one: reading emotion. It also shows that AI inherits the limits of the human-made data it learns from.

As the team’s neural network specialist, I worked out what was feasible: what the network should compute, how it could be visualized, and how it would be built.

No face-drawing dataset existed, so I built one. A mobile web app asked people to draw a face with a given expression; the museum shared it with the MIT community, and more than 7,000 labeled drawings came back.

I trained a series of networks in TensorFlow and compared their size, accuracy, and look. The chosen model is over 80% accurate, yet small enough to visualize every neuron and connection, and imperfect enough that visitors can try to fool it.

The team explored many forms, from physical builds to laser sculpture, and I tested what each could show. Even a small network has too many connections to build physically. The answer was to project each layer onto its own semi-transparent scrim, stacking them into a volume. A Three.js prototype on the real hardware gave the team the confidence to commit.

The museum’s top priority was showing real computation, not an animation. I ran the neural network inside Unity itself: compute shaders update every neuron and connection each frame, and a camera per scrim renders its slice of the network.