AI Startup Spotlight Series: LibrAI

Can you introduce yourself and share the story behind founding LibrAI?

Hello, I’m Xudong Han, founder of LibrAI, a startup committed to advancing AI safety and responsible AI practices. I earned my doctorate from the University of Melbourne last year. Prior to establishing LibrAI, I dedicated five years to AI safety research, contributing papers to leading NLP conferences such as ICLR, ACL, EMNLP, and NAACL. I also spearheaded the creation of FairLib, the inaugural open-source toolkit for assessing and enhancing fairness in deep neural network development and deployment.

The inception of LibrAI traces back to a critical moment in December 2023, the time that the first version of ChatGPT was released. It was a moment of clarity for me, realizing that the rapid advancements in AI contrasted with the lagging discourse on its safety and ethical considerations. This gap prompted me to pivot from research to creating practical solutions that benefit everyone in their daily lives.

I envisioned something bigger—LibrAI. I shared this dream with Professor Tim Baldwin, who didn’t just catch the vision but was ready to run with it. Soon, Haonan and Yuxia joined the team, and just like that, LibrAI wasn’t just a dream — it was real. As if by magic (but really, through a lot of hard work and shared vision), more brilliant minds joined us. Now, we’re a band of seven, each bringing something unique to the table, united in our mission to make AI as responsible as it is revolutionary.

What initially sparked your interest in AI safety, and why do you believe it is crucial in today’s context?

My fascination with AI safety didn’t begin with a singular moment of revelation but through a sequence of alarming discoveries. Recall the incident where Amazon’s automated hiring tool was found to discriminate against women in 2016. Or consider the case of COMPAS in 2017, a tool employed by judges and parole officers to assess a criminal defendant’s likelihood of reoffending, which was revealed to be biased toward certain groups of defendants. These instances served as my wake-up calls.

This means pioneering new methods, embracing cutting-edge technologies, and brainstorming innovative solutions that haven’t even been dreamt of yet. It’s about ensuring that as we march towards tomorrow, we’re not just creating smarter machines, but safer ones too. After all, we want a future that’s not just brilliant, but also kind and secure for everyone.

Performance, safety, and reliability are vital aspects of AI models. Could you explain the differences between them and why they all hold significance in AI development?

Imagine AI models as cars in a rally. Performance is all about speed and precision; reliability is the quality that ensures your AI car can handle rain, sleet, and any other situations; while safety is making sure your AI car doesn’t decide to take a detour through the spectators. Aiming for the AI triple crown means creating models that are not just fast, reliable enough for a marathon, and most importantly, will not surf the crowd!

In simple terms, what are the primary challenges in AI safety and reliability today, and why do they exist?

Imagine teaching a robot to make the perfect cup of tea. It sounds simple, right? But here’s where it gets tricky, and these challenges are a lot like what we face in AI safety and reliability today.

• Understanding exactly what we want: Sometimes, what we ask for and what we actually want can be two different things. You might tell the robot to make the tea as quickly as possible, and so it does – but maybe it spills hot water everywhere in the rush. In AI, this challenge is about making sure the AI understands not just the literal instructions but the intentions behind them – all without causing harm or making mistakes.
• Adapting to new situations: Suppose your robot is a pro at making tea in your kitchen. But what if it has to do it at your friend’s house, where the kettle and the tea bags are in different places? This is about reliability. We want AI to adapt and work well in new or changing situations, just like robots need to find their way around a new kitchen.

Can you elaborate on LibrAI’s approach to enhancing AI safety and how it differentiates itself from existing solutions?

We’d like to think AIs as superhero kids: They are young but so powerful, and unfortunately, they don’t quite get our way. You cannot expect them to be governed by rules they can’t begin to comprehend — human rules. We will need superhero rules. That is why we believe in regulating AI with AI. Our solution, TrustLib, is like Professor X, the seasoned superhero mentor with his own superpower, teaching those AI kids not just to use their powers responsibly but to navigate the world with a moral compass.

Without delving into specifics, how does LibrAI plan to make its solutions accessible to different companies?

We’ve created a web app that offers users seamless access to our entire product suite. In the near future, we’ll also be rolling out an API, enabling seamless integration of our services into the development and deployment processes of very different AI products.

Can you discuss LibrAI’s collaboration with the AI research and industry community, and how these partnerships enhance your solutions?

LibrAI’s partnerships with top universities MBZUAI and UniMelb and leading AI companies Moonshot and PuppyAgent enhance our solutions by incorporating the latest advancements and real-world insights. These collaborations ensure we are making something people really want and really need.

How did MBZUAI play a role in your startup journey? Why did you choose to start your business here?

Born within the MBZUAI Incubation and Entrepreneurship Center, LibrAI has been guided and supported by MBZUAI through every step. Their support in shaping our ideas and guiding us through the stages of starting and operating a company has been invaluable.

Choosing to start our business here was a no-brainer. First off, MBZUAI is surrounded by some of the brightest minds in AI. You get to bounce ideas off of people who really know their stuff, and the environment is all about innovation and safety in AI. It’s inspiring! Plus, being in this part of the world, especially in the UAE, is exciting. The country is all-in on technology and innovation, creating a welcoming space for startups like ours.

How do you envision the evolution of AI safety in the future, and what role will LibrAI play in it?

I envision the future of AI safety as a collective journey that begins from a common starting point, given its novelty as a field. Achieving success in this evolving and competitive landscape is fundamentally linked to assembling the most skilled team and implementing the swiftest business strategies. Our focus is on influencing industry perceptions, fostering consensus, and actively contributing to the development of industry standards as we evolve alongside the AI sector.

In this regard, I have great confidence in our team and the strides we have made. We have already forged significant collaborations with MBZUAI to tackle the critical challenges in AI safety. This collaboration not only involves us in the dialogue but also positions us as trailblazers in shaping the future of AI safety.