An artist’s tribute to modern AI

To get a feel for some of IBM’s recent contributions to AI, you can wade through the scientific literature. Or you can drop by the MIT-IBM Watson AI Lab in Cambridge, where a sculpture representing some of that work now hangs.

The artist, Mauro Martino, both studies and creates with generative AI as a researcher in the lab. Fittingly, AI is the subject of his latest work, as well as a co-creator. He used deep-learning algorithms to both thematically map select IBM research papers in the field of AI, and to transform the map into a physical object that would meet building code, among other practical constraints.

At IBM Research, Martino leads a handful of AI scientists whose work focuses on developing next-generation AI architectures and interpreting complex systems. His work has ranged from visualizing millions of crowdsourced drawings to mapping social networks, neural networks, and the dissemination of knowledge — which it turns out is also the theme of Atomic Reverberations, his first commissioned artwork for IBM.

He studied computational design at the Polytechnic of Milan. An internship with MIT professor Carlo Ratti and his Senseable City Lab brought him to Boston – and he never really left. Through Ratti, he was introduced to data visualization and network science, a field that uses computational tools to reveal hidden patterns in complex systems. The internship led to postdoc positions with professors David Lazar and Albert-László Barabási at Harvard and Northeastern respectively.

Rather than stay in academia, Martino decided in 2013 to take a job at IBM Research where he could build a team and stay active as an artist and scientist. “IBM offered the rare opportunity to do research in AI and explore its creative applications,” he said. “I could publish research in scientific journals and exhibit my generative AI art at the Venice Biennale and elsewhere.”

We recently caught up with Martino to talk about the new sculpture, co-creating with AI, and maintaining that passion for discovery as we get older.

What motivated this sculpture?

I wanted to try and represent physically the abstract relationships that AI can uncover in a sea of data. The sculpture is a map of AI innovation at IBM, with the nodes representing scientific papers, and the connecting links representing the patterns that AI discovers in high-dimensional latent space.

How did you make it?

The process began on the computer, using AI to map the similarity of thousands of IBM-authored publications in AI. I used an algorithm developed by the lab, hierarchical optimal topic transport, to analyze their semantic relationships and compress them into three dimensions. I then built on earlier work with Nima Dehmamy, Vox2Net, to convert the graph into a network sculpture using generative adversarial networks (GANs). I worked with the Battaglia Foundry in Milan to create wax molds which were later cast in bronze. The nodes of the final sculpture were polished by hand, while the connecting rods were left rough, to preserve the casting texture. The contrast is essential to the piece's meaning.

How so?

The reflective nodes represent human intelligence while the rough, connecting rods represent AI's hidden decision-making process. The piece is meant to explore our evolving relationship with algorithmic intelligence. With AI, we can develop a better understanding of the world around us, even if we don’t fully understand it.

When did you realize you wanted to be an artist?

I've never seen a clear boundary between science and art. When I began working with complex datasets, I discovered that visualization wasn't just about making information readable, it was about creating experiences that could help people feel the poetry in numbers. Art became my way to translate the abstract language of algorithms into something tangible and emotionally resonant.

What was the biggest challenge you faced?

My dream was to have a big sculpture. But in the end, it had to weigh less than 100 kilos (to meet building code) and be compact enough to fit in an elevator. The rods also had to spaced far enough apart to remove the supports used during the pouring process. I went through 200 versions to converge on something the foundry could work with, reducing the number of nodes from several hundred to 51 — the magic number for meeting my weight and size limit.

What are you currently working on?

I’m exploring how AI can be used to open new spaces for artistic and scientific expression. On the artistic side, I’m working on creating tools that make AI more accessible to artists and designers while continuing to push the boundaries of data sculpture and generative art. On the scientific side, I’m involved in several projects with my team aimed at rethinking how we train and interact with generative AI systems. The projects include developing ways to steer energy-based language models to better control their outputs; experimenting with reinforcement learning to pre-train language models; and studying joint training principles to make smaller models more capable and efficient by strategically combining tasks.

What does a complex network mean to you?

For me, a network is a strategy to make complex data visible and understandable. Imagine looking at a spreadsheet with 10,000 columns. There’s no way the human brain can truly grasp what all that information means. But if we reduce the dimensions and focus on how elements relate to one another, patterns begin to emerge. A network can transform an overwhelming amount of data into a structure we can explore, revealing connections that might otherwise remain hidden. In this sense, a network isn’t just a representation, but a tool for discovery.

How does AI fit into your creative process?

AI is both my subject and my collaborator, but more than that, it's my interlocutor in an ongoing dialogue about what it means to create. I approach it the way I imagine a sculptor approaches marble: with questions. What patterns lie hidden in this data? What stories want to emerge? I use AI to listen to the silent conversations happening in vast information spaces.

As we rely on AI more, do you worry it could make us less human?

I don’t. My approach has always been to maintain a soul in the machine. You can think of AI as a computational telescope that can reveal patterns invisible to the naked eye, while human creativity provides the critical perspective that transforms them into something personal and meaningful. AI can empower people to be more expressive.

A lot of your art hangs on the walls in Cambridge. Can you tell us about your favorite?

Each piece tells a different story about data and human experience, but if I had to choose, my favorite is the visualization I created for a Science paper on careers and peak performance. The data revealed that there's no single “peak” age for breakthrough discoveries. What I love about this piece is its message: never give up. Each of those peaks represents a breakthrough. Some peaks come early in a scientist’s career, others, late. Some scientists have multiple peaks. This visualization has become my daily comfort. When I pass it in the hallway, it reminds me that enthusiasm and persistence matter more than age or timing. Today could be the day for a breakthrough.

The 10 most cited IBM Research papers represented in Atomic Reverberations

Multitask Prompted Training Enables Zero-Shot Task Generalization 2149

Federated Learning with Matched Averaging 1626

A Transformer-based Framework for Multivariate Time Series Representation Learning 1387

Bayesian nonparametric federated learning of neural networks 988

Moments in Time Dataset: One Million Videos for Event Understanding 747

TransGAN: Two Pure Transformers Can Make One Strong GAN, and That Can Scale Up 559

On the Parameterization and Initialization of Diagonal State Space Models 517

Large-scale chemical language representations capture molecular structure and properties 486

Diagonal State Spaces are as Effective as Structured State Spaces 445

Vision Transformers are Robust Learners 437