“Technology doesn’t evolve on its own. We shape it — or we let it shape us,” said Dhrumil Mehta, Associate Professor in Data Journalism and Deputy Director of the Tow Center for Digital Journalism at Columbia Journalism School, opening the first panel of Data Science Day 2025 with a challenge.

Titled The Future of Data Science and AI, the conversation quickly turned toward a deeper set of questions: What kind of systems are we building? What problems are they designed to solve? And who decides?

Read 5 Insights from Data Science Day 2025

Key Takeaways

1. Why “Next Word” Prediction Feels So Smart

Vishal Misra, Vice Dean of Computing and AI at Columbia Engineering, broke down the core mechanism behind large language models like GPT-4.

“All they’re doing,” he said, “is predicting the next word.” But that simple objective, when applied at scale, creates surprising flexibility. These systems adapt mid-conversation, infer patterns, and take on new tasks without retraining.

“The brilliance isn’t built in — it emerges,” he said. “They learn to pay attention. That’s the shift.”

Misra’s takeaway: the next phase of progress may not come from building bigger models, but from designing smaller, faster, more purposeful ones.

2. We Don’t Just Need Explainability. We Need Meaning.

Smaranda Muresan, Associate Professor of Computer Science at Barnard College, works at the intersection of Natural Language Processing (NLP) and public interest domains—building models that support writing, detect misinformation, and inform public health.

Her focus: context, not just correctness. “If we’re using AI to support student writing or public health decisions,” she said, “we need more than accuracy. We need cultural nuance. We need reasoning.”

She emphasized the importance of knowledge-aware models informed by theoretical insights from linguistics and social science—not just scale.

3. Making Intelligence Work

“We won’t see productivity gains until AI touches the physical world,” said Matei Ciocarlie, Associate Professor of Mechanical Engineering at Columbia Engineering. His lab builds robotic hands and exoskeletons that bring AI into embodied interaction with people and objects.

One application: wearable robots trained on muscle signals (EMG) to support stroke rehabilitation.

“AGI is real,” he said. “But there’s no guarantee it will change the world unless we give it a body.”

His team’s work is a glimpse of what’s next—not just systems that think, but systems that feel and act.

4. AI Could Help Fight Climate Change — Or Make It Worse

David Sandalow, Inaugural Fellow at Columbia’s Center on Global Energy Policy, spoke about AI’s role in accelerating climate solutions. His team recently produced a comprehensive roadmap  for how AI can support emissions reduction—through better forecasting, smarter energy infrastructure, and faster materials discovery.

But, he warned, AI’s own energy use is rising fast.

“Right now, most of the public conversation is focused on energy consumption,” he said. “That matters—but the bigger question is whether we’re using AI to reduce emissions faster than we’re adding them.”

Sandalow offered a clear framework: AI can detect, predict, optimize, and simulate. Doing all four—at scale and with purpose—is the opportunity.

Final Word

Mehta closed the session with a reminder that agency matters.

“Whether you’re a developer, a journalist, a policymaker, or a student, the decisions you make now shape the future of these tools—and the world they build.”

At the Columbia University Data Science Institute, shaping that future begins with asking the right questions—early, and across disciplines, and with urgency.  

More News From Data Science Day 2025

Five Insights from Data Science Day 2025

Panel 2 Recap: Designing for Trust: AI Security, Risk, and Accountability

Panel 3 Recap: When the System Is the Patient: AI in Health Care

Keynote Highlights: Building Systems that Scale