In the final session of Panoramai's Swiss Generative AI Summit, Raphael Briner welcomed Mara Graziani, Research Scientist at IBM Research Europe, for an intimate exploration of cutting-edge AI research and its practical applications. As the day's last speaker, Graziani brought a unique perspective on how fundamental research translates into real-world innovation.

The Evolution of AI Research at IBM

Mara Graziani positioned IBM Research as « the incubators of new ideas, of innovation, of what's coming next. » She emphasized their role in thinking about the future and helping different industries adopt AI revolution technologies. « We really want to empower everybody, all the people that work with us, all the collaborators that we have and the partnerships, to help them, to empower them with the generative AI tools, » she explained.

Graziani outlined IBM Research's three-pronged approach to advancing AI innovation: developing new architectures that integrate different modalities beyond text, exploring synthetic data generation capabilities, and implementing generative workflow systems that facilitate AI applications across diverse fields.

Digital Twins: Revolutionizing Predictive Modeling

One of the most compelling applications Graziani discussed was AI-powered digital twins. She described a breakthrough project where their team successfully predicted sensor behavior and lifecycle evolution with remarkable accuracy. « The revolution we are seeing with digital twins that are powered by AI is that you can predict that and model it actually quite accurately, » she noted, explaining how this eliminates the traditional wait-and-see approach to system monitoring.

The transformation represents a fundamental shift in how digital twins operate. Previously, creating effective digital twins required deep understanding of physical equations and system mechanics—knowledge that took years to develop. « Before we needed to know the physical equations of a system, we needed to understand very deeply the system to be able to model it, which is something that now we can bypass because we can have data-driven approaches, » Graziani explained.

This data-driven approach has enabled IBM Research to tackle complex systems like battery lifespan prediction, where multiple interacting factors make traditional physical modeling extremely challenging. The breakthrough allows for much faster discovery and accelerated innovation cycles.

Synthetic Data Generation and Model Customization

Graziani highlighted IBM's InstructLab project as a revolutionary approach to synthetic data generation. The initiative enables users to become « creators of value because they can use their own taxonomy and their own interpretation of the world and their own domain knowledge to build better models. »

The InstructLab system creates domain-specific taxonomies, generates synthetic data following those taxonomies, and fine-tunes language models accordingly. This approach operates at high throughput levels while using quantized models that don't require massive infrastructure investments. « Data is where a lot of work needs to be done, right? A lot of companies have a lot of data, but it requires a lot of cleaning and preparation for all of the machine learning pipelines, »Graziani observed.

Overcoming Data Limitations Through Innovation

Even when companies possess machine learning-ready data, Graziani noted that traditional approaches sometimes fail to extract value. IBM Research addresses this challenge by developing new models and technologies, including reinforcement learning techniques that make previously « unexploitable » data finally exploitable for business purposes.

Personal Journey and Industry Evolution

When asked about her decade-long journey in AI research, Graziani reflected on the dramatic transformation she's witnessed. Starting her master's degree in 2015, she experienced the field when it still required extensive hand-crafting of features and domain knowledge. « There was machine learning, but it was a lot of hand crafting of features, a lot of domain knowledge needed and it was obviously very expensive because you had to take all of those data sets and annotate them, » she recalled.

The revolution beginning around 2016-2018 introduced self-supervision and foundation model paradigms that dramatically reduced costs and annotation requirements. When discussing the economic impact, she noted how this transformation allowed organizations to take greater risks, explaining that « you also increase the return of investment if you want because whereas before you had to put all of your investments in taking these data sets and labeling them and annotating them, now you can... focus only on having very small, very high quality data set that you can tune on. »

Maintaining Research Focus in a Rapidly Evolving Field

Despite the acceleration of AI development tools and support systems, Graziani emphasized that IBM Research maintains disciplined focus. « We still keep quite focused. It didn't make us lose the focus, » she explained. The team continues investing significant time in strategic thinking about development directions, using conferences and community engagement to gather feedback and understand field evolution.

The Physical World Integration

Looking toward the future, Graziani expressed excitement about AI's expanding understanding of the physical world. « That's really where you see the revolution happening, » she said, noting that while digital twins have existed for decades, AI-powered versions can now model systems for which traditional physical understanding was previously impossible or too complex to develop.

This expansion enables digital twin applications for highly complex systems like weather forecasting (through IBM's NASA collaboration) and intricate manufacturing processes. « When you have a lot of data, you can infer those from the data. And that's how you learn much faster and you accelerate the pace of discovery, » Graziani concluded.

The fireside chat revealed IBM Research's strategic positioning at the intersection of fundamental AI advancement and practical business application, demonstrating how cutting-edge research translates into tools that empower organizations across industries to leverage the generative AI revolution.