Future of AI according to Nvidia
Waymo factory building a robotaxi future
When ChatGPT broke an entire field
Future of AI according to Nvidia
ChatGPT moment for robotics is coming,’ Huang said. Here’s why. The CEO went into detail regarding Nvidia’s plans to invest heavily in “physical AI,” a burgeoning field focused on the use of AI to create and simulate real-world physics data for use in robots and self-driving cars. Speaking to financial analysts, Huang explained his belief that companies should focus on developing humanoid, bipedal robots, because the terrain that they operate in doesn’t need to be altered the way it would for a wheeled or stationary robot. “The ChatGPT moment for robotics is coming,” said Huang.
The only problem? In order to reach that ChatGPT moment, robots need to accomplish physical tasks without falling over, and to do that they need AI models that have been trained on massive amounts of physics-based data. Plus, before an AI model can be uploaded to a real-life robot body, it needs to be improved by running simulations in a digital environment. To that end, at the keynote, Nvidia announced Cosmos, a new platform that gives developers access to “world foundation” AI models, designed to generate huge amounts of physics data. Once a robotics model has been trained on the synthetic data, it can be dropped in Nvidia’s Omniverse platform, which is used to create the virtual environments where the models simulate various tasks, learning from each failed attempt until they finally succeed.
Huang was quick to note that the company’s focus on physical AI is inherently tied to its cash-cow data center business, since creating synthetic data and high-quality simulations requires “racks and racks” of Nvidia accelerator chips. The Nvidia CEO also revealed that he’s ordered the company’s engineers to walk their big talk. Speaking to the analysts about his belief that all knowledge workers will soon have their own AI assistants, Huang said that “every software engineer at Nvidia has to use AI assistants next year, that’s just a mandate. Otherwise, they’re not coding fast enough!”….
***
Waymo factory building a robotaxi future
An exclusive look inside the facility turning Jaguar EVs into robotaxis with the AI-driven fleet’s custom computing system, cameras, lidar and radar. Soon, tens of thousands of robotaxis will be rolling off the line annually. At Phoenix’s Sky Harbor airport to the rideshare zone on a hot spring day and you’ll catch a glimpse of a fast-approaching future: driverless Waymo robotaxis queueing alongside human-driven Ubers and Lyfts to take waiting passengers to their next destination. The service just launched in Austin and continues to expand in San Francisco, Los Angeles and Silicon Valley, but Phoenix has been its home turf for years.
About 20 minutes east of Phoenix’s airport in Mesa, Arizona, is a 239,000-square-foot factory that opened in October. Every day, it churns out several battery-powered, white Jaguar I-PACE electric SUVs loaded up with the company’s custom-designed computer, cameras, radar and laser lidar sensors on a single production line. But the plan is to dramatically scale up the pace and automate output to keep up with growth plans, said Kent Yiu, Waymo’s head of vehicle manufacturing, who previously managed production operations for Apple and General Motors.
The production scale is small compared to traditional auto plants that make hundreds of thousands of vehicles a year. But the 1,500 robotaxis Waymo has provide more than 250,000 paid rides a week or about 24 a day per vehicle, vastly more use than personal cars and trucks that are driven only a few times a day. And by the time the Mesa factory gets 10,000 Waymos on the road, the fleet could be booking 250,000 rides a day. That’s well over 1.5 million a week. At that scale, Waymo’s annual revenue could jump to $2 billion, up from a Forbes estimate of $100 million last year. The company declined to comment on those estimates.
After years of testing and pilot programs stretching back to 2009, powered by three funding rounds that raised over $11 billion — not to mention the untold billions more Google poured into the program between 2009 and 2020 — Waymo finally became a real business last year. “We’ve been laser-focused and will continue to be on building the world’s best driver,” Alphabet CEO Sundar Pichai said on the company’s April 24 results call. “I think doing that well really gives you a variety of optionality and business models across geographies, etcetera.” The factory is run with Magna, a leading auto engineering and manufacturing company that produced the Jaguar I-PACE Waymo uses at its Graz, Austria, plant. It replaces a smaller Detroit assembly facility Waymo opened in 2019, also with Magna, and closed at the end of 2024.
The work pace is steady but not high volume. Raw cars roll into the building at one end, with plastic covers on the body panels over precut sections where sensors will be installed. They enter a manual assembly line where dozens of workers remove those covers, bumpers and other exterior components to begin the process of carefully installing an electrical wire harness, computers, sensors on each corner of the vehicle and Waymo’s telltale “top hat” unit–housing the main laser lidar for 3D vision, multiple cameras and audio sensors….
***
When ChatGPT broke an entire field
The goal of natural language processing is right there on the tin: making the unruliness of human language (the “natural” part) tractable by computers (the “processing” part). A blend of engineering and science that dates back to the 1940s, NLP gave Stephen Hawking a voice, Siri a brain and social media companies another way to target us with ads. It was also ground zero for the emergence of large language models — a technology that NLP helped to invent but whose explosive growth and transformative power still managed to take many people in the field entirely by surprise.
By 2017, neural networks had already changed the status quo in NLP. But that summer, in a now-seminal paper titled “Attention Is All You Need,” researchers at Google introduced an entirely new kind of neural network called the transformer that would soon dominate the field. Not everyone saw it coming.
Google had organized a workshop in New York for academics to hang out with their researchers, and Jakob Uszkoreit, one of the authors on that paper, was presenting on it. He was making a really clear point about how aggressively this model was not designed with any insights from language. Almost trolling a bit: I’m going to just talk about all these random decisions we made, look how absurd this is, but look how well it works. There had already been a feeling of the neural nets taking over, and so people were very skeptical and pushing back. Everyone’s main takeaway was, “This is all just hacks.”
It was sort of interesting, but it wasn’t an immediate breakthrough, right? It wasn’t like the next day the world changed. I really do think it’s not conceptually the right model for how to process language. I just didn’t realize that if you trained that very conceptually wrong model on a lot of data, it could do amazing things.
I clearly remember reading “Attention Is All You Need” in our NLP reading group. Ray actually ran it, and we had this very lively discussion. The concept of attention had been around for a while, and maybe that’s why Ray’s reaction was kind of, “Meh.” But we were like, “Wow, this seems like a turning point.”
During that summer, I vividly remember members of the research team I was on asking, “Should we look into these transformers?” and everyone concluding, “No, they’re clearly just a flash in the pan.”
Soon after it was introduced in October 2018, Google’s open-source transformer BERT (and a lesser-known model from OpenAI named GPT) began shattering the performance records set by previous neural networks on many language-processing tasks. A flurry of “BERTology” ensued, with researchers struggling to determine what made the models tick while scrambling to outdo one another on benchmarks — the standardized tests that helped measure progress in NLP.
There was this explosion, and everybody was writing papers about BERT. I remember a discussion in the [research] group I was in: “OK, we will just have to work on BERT because that’s what’s trending.” As a young postdoc, I just accepted it: This is the thing that the field is doing. Who am I to say that the field is wrong? When people submitted benchmark results and wanted to appear on the leaderboard, I was often the one who had to check the result to make sure it made sense and wasn’t just someone spamming our system. So I was seeing every result come in, and I was noticing how much of it was just, increasingly, old or simple ideas scaled up. The background assumption was, “Transformers aren’t going to get much better than BERT without new breakthroughs.” But it was becoming clearer and clearer for me that scale was the main input to how far this is going to go. You’re going to be getting pretty powerful general systems. Things are going to get interesting. The stakes are going to get higher. So I got very interested in this question: All right, what happens if you play that out for a few years?….