“It took 3 and a half years for Netflix to reach 1 million users, Airbnb 2 and a half years, Facebook 10 months, and if you look at ChatGPT, it took 5 days”, Raquel Urtasun, CEO and founder of Waabi, a startup leading the autonomous truck industry, breaks down the preternatural numbers. As a part of the Department of Computer Science’s Distinguished Lecture Series, Urtasun spoke about AI transitioning from the digital to the physical world.

AI is transforming our institutions and societies, but what enabled this revolution? Uratsun explained the three major components: Data, Compute, and Algorithms.

Data: Large Language Models (LLMs) need copious amounts of data, and the internet was the perfect source for mining it.

Compute: NVIDIA is a 3.3 trillion dollar company, for good reason. Models need powerful computing technology to sustain AI solutions; the infrastructure is vital.

Algorithms: Improvements to algorithms lead to increased scalability of systems.

But surely, everyone’s wondering, “What’s next?” as we stand on the precipice of change. Uratsun admits while you won’t see any robots in Toronto outside your window today, in 10 years, they’ll be everywhere. There are three main challenges to deploying models in the physical world: Generalization, Efficiency, and Provable Safety.

Generalization: There’s insufficient real-world data to address every nuance, complexity, and unforeseen possibility. However, one wrong decision can lead to catastrophic consequences and put lives at risk. Uratsun emphasizes that the system must always make the “right” decision and generalize to unfamiliar situations.

Efficiency: Safety-critical systems must have efficient, low-energy architectures to make swift decisions. Moreover, sustainability should be a priority. As the models grow, the cost and power to train them will increase exponentially.

Provable Safety: We must demonstrate the safety of physical systems but also prove it beyond a reasonable doubt.

Over the last two decades, Uratsun has been exploring the idea of using foundational models and single AI systems that can reason like humans. In 2021, she decided to focus on the trucking industry and founded Waabi, foreseeing the potential for mass deployment and scale in the near future. In hindsight, she was onto something. The trucking industry is a 900 billion-dollar market in the United States. Autonomous trucks could address several pain points, such as the shortage of drivers, safety concerns on the road, and environmental impact.

However, this unique opportunity presents many challenges. The average American spends 1.5 hours daily in their cars, yielding 1.5 hours of uncertainty and unpredictability. In practice, it’s increasingly difficult for a machine to anticipate and generalize to new situations. The industry is currently built around AV 1.0, which requires incorporating knowledge for the system to reason about it. But Uratsun wanted to leverage AI to disrupt the status quo and autonomous driving. She explains that at Waabi, like how LLMs can predict the next word or token in a sentence, they’re working to predict what the truck will encounter next. Further, to address the data problem, Waabi architects a new world, a simulation.

Uratsun explains that simulations can model safety-critical situations instead of driving aimlessly to collect data and risking accidents. Generative AI models mimic reality through a 4-dimensional neural world, simulating real-time sensors and dynamics, testing the entire system, and measuring system latency. Humans, vehicles, and animals are reactive agents in the simulation, effectively cloning reality.

Uratsun concludes that self-driving trucks will be hitting the roads next year, and robots will be deeply integrated into our routines- with the unexpected possibility of making our lives more human.

Written for Neural Notes, U of T AI’s Newsletter