Mind of Swarat Chaudhuri

 

 

An out-of-the-bots approach

·         Hindustan Times (Bathinda)

·         25 May 2025

·         Swarat Chaudhuri [ swarat@cs.utexas.edu ]

====================================================

  ]

Chaudhuri just won a Guggenheim Fellowship for his unusual work with AI. The computer science professor at University of Texas, Austin, is trying to get artificial intelligence to propose its own math problems. 

The larger aim:

to explore whether such programs can develop a sense of ‘curiosity’. The even bigger question: could AI actually advance the scale of human knowledge?

Swarat Chaudhuri was obsessed with puzzles as a child, he says. Growing up in Kolkata, he spent the afternoons solving every number pyramid and word jumble he could find in local publications. He then graduated to Bengali translations of the American mathematician Martin Gardner’s popular-science books on math, logic and puzzles.

Chaudhuri dreamed of a world in which he could solve puzzles for a living.

That is, more or less, what he now does, as a researcher and professor of computer science at the University of Texas, Austin.

One of the puzzles he’s working on is particularly crucial. It is the question of whether artificial intelligence can actually expand the scale of human knowledge.

Here’s how Chaudhuri, 46, is going about it: At his Trishul lab at the university, he has developed an AI program called COPRA (short for In-context Prover Agent) that works with large language models (in this case, GPT-4), to prove mathematical theorems.

That may not sound like much fun, but here’s what waits down the line. As the two systems work together, Chaudhuri’s eventual aim is for Copra to propose new math problems of its own, and then work to solve them.

This would be a crucial step towards determining whether AI can emulate the curiosity-driven explorative nature of the human mind. It would go some way towards answering questions such as: Could AI eventually co-author a scientific paper?

In other words: Can an AI program reach beyond what it knows, in order to not just connect dots in new ways (they are already doing this) but to reach out and gather more dots to add to the matrix? (Dots that we may not have factored in at all.)

“This would mean a big leap, from the AI engines we see around us, which deal in available data and perform somewhat repetitive tasks, to a system that uses a lot more logic and can perform ‘superhuman’ tasks,” he says.

That “superhuman” bit is what interests him, because it would mean newer and faster solutions to some of our most pressing problems. Such programs could potentially alter how we view our world, and navigate it.

New maths problems would be just the start. Down the line, he believes, these programs could be collaborators working alongside researchers.

“They could be like curious children stepping out to findthings on their own and figure out what works and what doesn’t,” he says.

With the help of such a program, small start-ups and lone tinkerers could take on giant puzzles such as cleaner energy and urban planning.

New answers could emerge to questions such as: How do we move large numbers of people from Point A to Point B and back every day? How do we address the issue of solar cells having such a short lifespan? Or, how can we better manage indoor temperature control, amid the climate crisis?

Chaudhuri has been at this for a while. In 2017, he and his students created Bayou, an early Ai-led tool that could build code based only on text prompts. That early win set the stage for his current Copra system of AI agents.

Chaudhuri has now been awarded a prestigious Guggenheim Fellowship, for his body of work and for the projects he is leading on “open-ended mathematical discovery”.

He knew early on, he says, that the path to the world’s greatest and gravest puzzles lay through the world of computing.

After school, he studied computer science and enrolled at the Indian Institute of Technology (Iit)-kharagpur. After college, he began studying neural networks — which are types of programs that seek to mimic the dot-connecting capabilities of the human brain, rather than relying on the linear (this-therefore-that) progressions that guide traditional software.

At this time, 20 years ago, the idea that a computer could one day sift through options and pick the right one, rather than spit out a ready answer that had been fed to it, was considered outlandish. Today, of course, all LLMS do it. It’s how CHATGPT converses; how Midjourney and Sora create their eerily realistic images and videos.

Five years from now (if not sooner), Chaudhuri believes it will be as common for AI programs to “think” in ways that more closely mimic the human brain, and to have superhuman abilities in many areas of human activity.

A switch to renewable resources will be vital, to lessen the environmental impact of the server farms and data centres that support such systems, he adds.

What’s something he believes even this advanced version of AI would struggle to do? Create profound art, Chaudhuri says.

For the simple reason that the arts are driven, perhaps more than any other human endeavour, by the artist’s lived experience of the world.

“It is unlikely that AI will start writing like Rabindranath Tagore or churning out innovative movie scripts because, to produce something like literature, the constantly shifting inputs from the world and the interaction of the artist with the world are vital,” Chaudhuri says.

“That level of input is a long way away for AI.”

Article Name:An out-of-the-bots approach

Publication:Hindustan Times (Bathinda)

Author:Swarat Chaudhuri  /   dipanjansinha2000@gmail.com

===========================================