‘We’re Interested in Huge Leaps’: Behind the Scenes with Salesforce’s AI Research Leader

When Silvio Savarese left full-time academia for Salesforce in April of 2021, he had one goal in mind: making an impact. 

In two decades as a scholar, most recently as a tenured professor of computer science at Stanford University, Savarese made major contributions to the fields of robotics, machine learning, and language models. By joining Salesforce AI Research as Executive Vice President and Chief Scientist, Savarese knew he could spearhead even bigger leaps in artificial intelligence. 

We sat down with Savarese to learn more about his team’s pioneering work. 

Q: How does your computer science background influence your research scope and interests?

Initially, my research was focused on modeling how human vision works in very granular detail — such as how a person can look at a photograph and identify the various elements within the picture. This is at the foundation of modern machine vision, which focuses on large-scale image classification and scene understanding problems.

My research evolved into understanding and predicting human behavior and intent from observations. For example, how can machines see images of people eating, and know from data and cultural context that they’re having Christmas dinner? This naturally led me to consider the role of language, which is a central part of the human experience and collaboration — how we ask questions, share information, and express our needs. 

I’m excited by the potential of machines that can do the same — from customer service chatbots that truly understand our questions to support robots that respond to verbal commands. These are extremely complex challenges, of course, but I’m optimistic about what the coming years will bring.  

Silvio Savarese

Q: What inspired you to leave full-time academia and take the helm at Salesforce AI Research?

Over the years, I became more committed to spending my energy and time on things that can change people’s lives. At Salesforce, we can deploy our research into real products that immediately help customers. 

Salesforce also has an incredible array of resources that give us the opportunity to have an impact at scale. Given how AI has evolved, you need a lot of computing resources and enterprise-level data to make major progress in research.

Q: What is the Salesforce AI Research team’s primary mission?

We’re a growing group of researchers working to develop new AI technology that can help our company innovate and transform society at large. Overall, our goal is to empower people by making innovative AI tools that are also intuitive and simple to use. 

We don’t just want to make incremental advances though; we’re interested in making huge leaps in AI research. At the same time, we want to pay attention to real-world use cases and customer pain points. 

Our job is to provide a glimpse into the future, highlight capabilities that don’t yet exist in the market, and create building blocks that can be deployed into products. 

Q: Can you share some of the contributions your team has made to Salesforce?

Internally, we’ve been helping our engineering teams by using AI to improve system availability, and detect potential failures and anomalies with our Merlion project. 

CodeGen acts as a coding assistant and leverages large language models to translate user specifications, written in plain English, into executable code. I’ve seen the excitement it inspires in beta testers across Salesforce and it makes me optimistic that a revolution in the way we code is on the horizon.

Finally, our latest work on conversational AI allows us to produce summaries that answer particular questions of interest, enabling greater user control and personalization. 

Q: How can today’s AI professionals stay ahead of the curve?

First, it’s essential to invest in security processes for data — that’s still a critically important aspect of AI research. 

I also think it’s crucial to enable cross-disciplinary work. The boundaries between different fields of AI are now fuzzy and fluid, so you need collaboration across teams. 

Companies should take an AI-first approach to products in order to be competitive in this complex space — companies need to lower the bar for customers to use AI capabilities. For example, conversational AI means replacing the learning curves of traditional interfaces with a system the user can simply talk to, to get work done. 

Finally, ethical principles for AI systems should be baked in at the design stage, not added later. We need to make sure these tools are fair, safe, inclusive, transparent, and accessible.

Learn more about Savarese’s team’s work on Salesforce’s AI and Research blog.