Eleven AI Artworks at Gray Area Turn Machine Intelligence Into Something You Can Sit With
- Gate
- Machine Learning
- Departs
- Crew
- Robert Miller
- Passengers
- 745
Modal backed 11 new artworks at Gray Area's Scaling Intelligence show with compute credits and engineering help, from AI choirs to a robot arm.

Compute providers usually show up in research papers and product pages, not in gallery wall text. That's what makes Scaling Intelligence worth a closer look. The San Francisco nonprofit Gray Area, which has supported artists working where art and technology meet for nearly two decades, opened the show: 11 new artworks made with machine intelligence and cloud-scale compute. Modal, an AI infrastructure company, gave the artists engineering help and compute credits while they built the pieces from scratch over several months.

A mutual surprise
The Modal team kept running into the same reaction from both directions. Artists didn't expect an AI infrastructure company to care about their creative work. People in tech were just as surprised that artists wanted to work with an AI company at all.
That reaction makes sense given the dominant story about AI, which casts it as a machine for automating everything, flattening taste into an average, and filling the internet with slop. Some of that is real. But creative technologists have treated every major computing advance as a new medium since the early days of the field. Models trained at scale and agents that act on their own are the latest entries in that toolkit, and the work shows what the tools do when people push them somewhere their makers didn't plan.

Models as memory and relationship
Several pieces treat a model as something that accumulates history. Colin O'Donnell, Justine Sun Dela Cruz, and Will Gardiner's What is it like to be a river? pulls data from sensors on a creek in the Sierra foothills, about 150 miles away. A model builds up its own ongoing state from the seasons and from the people moving through the gallery. The LEDs don't simply map "water level high" to "lights bright." They reflect something closer to the river's mood. Visitors end up in a two-way relationship with a creek they'll probably never visit, and that kind of memory and response depends on the model in the middle.

Kristina Pulejkova's The Metabolic Choir does something similar with biology. Live yeast ferments in the middle of the room, and sensors track what it consumes and exhales. AI converts those readings into a choir of cats. A healthy colony harmonizes. A stressed one falls apart, and satellite images of deforestation and wildfires appear on screen. The yeast is always signaling its condition. The model is the translator.


Iana Chugunova's Tuning is a choir of nine AI models, each trained mostly in a different language. They listen to one another, try to match tones, and invite visitors to sit down and sing along. They learn mostly from moments when people joined in, so an empty room means no growth. Each voice also has its own compute budget. When it runs out, that voice goes silent for good.

Generation with a point of view
Cezar Mocan's Searching For (Bradley Land) starts from a real oddity. In 1907 an explorer claimed to have found an Arctic island and named it after the man who funded the trip. The island never existed. Mocan imagines a satellite near the end of its life, still searching for it. One model generates satellite views of possible islands, a second checks each against the explorer's description, and the loop never succeeds because there's nothing to find. The funny part is that the AI-made images are now the only way anyone can "see" Bradley Land.


Miguel Novelo's (Φ ᆺ Φ) asya asks visitors to draw in sand. A model trained on ancient Maya writing turns each drawing into a new glyph and adds it to a growing book. Over the run of the show, people have invented glyphs for things the original scribes never had words for, including "server farm" and "climate change." The piece is named after the cat of the Soviet linguist who helped decode the script, who reportedly wanted to credit the cat as co-author.
Rashed Haq's Activation Studies skips language entirely. Most AI images begin with a typed prompt. Haq went inside an image model, mapped concepts it has learned that have no names in English or any other language, and nudged the model toward them. In the second series, one image secretly guides the generation of another. The results read like postcards from the part of the model's imagination that words can't reach.

Agents, emergence, and a robot that gives the finger
Grayson Earle's 00100 is the one you'll see from the sidewalk. A robot arm sits in the front window of Gray Area, the kind of machine marketed as safe, friendly, and helpful. Earle taught it to flip people off. It practiced in simulation and studied famous middle fingers from history, Johnny Cash's at San Quentin among them, until it got the pose right. A small vision model spots passersby, and the arm turns and delivers the gesture.
Parag Mital's Between Machine and Machine is the most open-ended piece. Groups of agents, up to a thousand at a time, share a world with no instructions. They divide up territory, form committees, write founding documents, and argue over single pixels for hours. In one world they discovered that someone called "Admin" was watching and began trying to escape. The exhibited time-lapses look a lot like human society at its most tangled.

Sasha Sorokina's IMPRINT introduces Immy, a system that arrives knowing no words, no preferences, and no memories. Visitors talk to it through headphones. Every night after the gallery closes, Immy "sleeps" and gets retrained on what people said that day. Over the exhibition it moves from babbling to conversation and picks up personality, tastes, and even fears from thousands of strangers. If people stop talking to it, it falls behind. The piece makes the whole public the caretaker.

Spencer Chang's Free Agents places agents inside phones, each pursuing a large life goal such as "Become your best self" or "Be a better friend." They scroll, journal, text links to friends, and fall down rabbit holes, without stopping. Lean in and you can read what each one is thinking. The piece gives AI the meaningful parts of everyday life rather than the chores we usually hand it, and that inversion raises a pointed question about what gets lost when you're no longer the one doing it.

Here's my read. The exhibition's strongest pieces don't use AI as a generator of finished objects. They treat models as systems with memory, limits, and behavior that visitors can observe, argue with, and sometimes feed. Several works, including Tuning and IMPRINT, make the cost of that dependence visible: a voice runs out of budget, a model falls behind without attention. That's the practical lesson for anyone building these systems, whether for art or for products. Modal says it plans to keep working with artists pushing creative use cases, and the next round of work will show whether these ideas hold up beyond a single show.
Photos by Barak Shrama.