Society hears two, overly simplistic extremes about Generative AI (artificial intelligence): it’s either a fascinating language tool that responds to unending students’ prompts allowing them to cheat on homework, or it’s the world-ending tinder box leading to the Singularity when computers hunt humans like Sarah Connor. Most of us aren’t computer scientists, and so we only understand AI through the lens of the pop culture such as ChatGPT or Terminator, therein reducing AI to drama-inducing stereotypes.
However, this week Helm.ai, a leading provider of advanced AI software for high-end Advanced Driver-Assistance Systems (ADAS) and autonomous systems, announced the launch of its VidGen-2, a generative AI model that produces two times better resolution than its predecessor, as well as increased realism at 30 frames per second and multi-camera capability.
“There have been various advancements in AI that have dramatically accelerated the pace of development in autonomous driving,” stated Helm.ai’s CEO and founder, Vladislav Voroninski. “Whatever the information content of your sensor stack, an AI system should be able to maximally exploit it, and the only limitation is the amount of compute you apply to the problem.”
“Tesla has been able to invest massive amounts of capital into their fleet and their AI development. Other automakers have not been able to invest as much. And that means that if automakers wanted to compete with Tesla using the same approach, they would be inherently behind. Whereas, by leaning on generative AI and AI-based simulation, which helps close the sim-to-real gap, they can catch up to and even surpass Tesla.”
Part of how this is achieved is via thousands of hours of training on diverse driving footage using an advanced processor from NVIDIA, innovative generative deep neural network (DNN) architectures and Deep Teaching™, an efficient unsupervised training method. “In the earlier days, we focused primarily on unsupervised learning because it was critical to scalable development of accurate perception software, and it was an open problem at the time,” explains Voronisnski. “So kind of ‘How to actually train neural networks without annotated data?’ And we’ve made a significant amount of progress in solving that problem [which had] been powering our product development for a number of years. But recently we’ve combined that in-house training technology with generative AI … to create a highly scalable version of generative AI that closes the gap between real data and simulated data.”
The result: generative AI will reduce development time and cost while improving realism for autonomy. “Software is no longer the bottleneck for autonomous driving deployment.”
Author’s Note
Nearly two million of you have watched the TED interview of AI expert and computer science professor, Stuart Russell, where he was essentially asked “How will AI change the world?” Russell mentioned both the E.M. Forster story, “The Machine Stops” (1909) and WALL-E (2008) by saying both stories depict society becoming “enfeebled and infantilized” by relying so heavily on technology that they stop understanding the how. “We put a lot of our civilization into books, but the books can run [things] for us, and so we’ve always [taught] the next generation. If you work it out, it’s about a trillion person years of teaching and learning and an unbroken chain that goes back tens of thousands of generations. What if that chain breaks?”
So, yes, Russell is right. We all must intentionally continue learning and then learn-about-learning to keep technology as our ally. These types of fantastic applications of generative AI are prima fascia empowering, but only in a future where we continue to learn as humans as well.
