A MESSAGE FROM OUR SPONSOR

How can healthcare organizations capture the next wave of AI value?

Many healthcare systems and hospitals have deployed AI to address specific clinical and operational challenges. But without enterprise-wide integration, these tools can create fragmentation and limit impact. Real transformation will require a more connected approach.

In The coming evolution of healthcare AI toward a modular architecture, McKinsey explores how shifting to a modular, interoperable AI architecture can unlock scale, improve data governance, and enable holistic innovation. The article outlines the strategic moves leaders must make now to build a more resilient, AI-enabled healthcare system.

Read the full article here

And now for the main event…

Late last year, Yann LeCun, the Turing Award–winning A.I. pioneer, left Meta after a long-simmering disagreement about machine learning exploded into public view. LeCun had become increasingly convinced that large language models—the foundational technology for most modern frontier labs—was a technical dead end. The future, he posited, belonged to what industry researchers refer to as world models: A.I. systems trained to simulate the physical world. A few weeks later, LeCun quietly launched his own company, explicitly focused on world models: Advanced Machine Intelligence Labs.

Earlier today, LeCun announced that AMI Labs had raised over $1 billion in seed funding, at a stunning $3.5 billion valuation—not bad for a Paris-based company with a dozen employees. “My prediction is that ‘world models’ will be the next buzzword,” C.E.O. Alex LeBrun, another former Meta employee, told TechCrunch, noting the appeal for robotics manufacturers and heavy industry. “In six months, every company will call itself a world model to raise funding.”

The LeCun–LeBrun manifesto is indeed part of an emerging industry hypothesis that the road to more advanced A.I. will be paved by more than just large language models, the primary technology underpinning Anthropic and OpenAI. Judea Pearl, another Turing Award–winning pioneer in A.I., said recently that L.L.M.s “have mathematical limitations that cannot be crossed by scaling up.” It’s a contention shared by Richard Sutton, yet another Turing winner. “To mimic what people say is not really to build a model of the world at all,” Sutton noted in September. “They have the ability to predict what a person would say. They don’t have the ability to predict what will happen.”

Dr. Fei-Fei Li, another pioneer in the space, has made a similar argument. “Language is purely generated signal,” she has said. “You don’t go out in nature and there’s words written in the sky for you.” Li’s startup, World Labs, is focused on solving the problem of spatial, physical intelligence. The company recently raised $1 billion at an undisclosed valuation following a $230 million round in 2024.

Investor excitement for world models isn’t limited to research titans, either. Runway, the A.I. video-generation company that recently completed a $315 million Series E round at a $5.3 billion valuation, is intently focused on building world models. And then there’s Odyssey—a company backed by Nvidia and Samsung—whose founders believe we are on the cusp of a moment, akin to the release of ChatGPT in 2022, in which a new world model is about to change everything. (Yes, their name is an homage to 2001: A Space Odyssey.)

Before co-founding Odyssey in 2023, C.E.O. Oliver Cameron and C.T.O. Jeff Hawke spent years developing driverless cars. Hawke was a founding engineer at Wayve, a juggernaut in the self-driving space in Europe, while Cameron ran a self-driving firm called Voyage until its acquisition by the (now defunct) self-driving giant Cruise in 2021. When I met up with the duo last week, Cameron explained that driverless cars have developed what he described as a “narrow world model.” In other words, he said, self-driving A.I. has “learned from the world to predict how it evolves over time.”

A MESSAGE FROM OUR SPONSOR

How can healthcare organizations capture the next wave of AI value?

Many healthcare systems and hospitals have deployed AI to address specific clinical and operational challenges. But without enterprise-wide integration, these tools can create fragmentation and limit impact. Real transformation will require a more connected approach.

In The coming evolution of healthcare AI toward a modular architecture, McKinsey explores how shifting to a modular, interoperable AI architecture can unlock scale, improve data governance, and enable holistic innovation. The article outlines the strategic moves leaders must make now to build a more resilient, AI-enabled healthcare system.

Read the full article here

The problem is that the tech “only works for driverless cars.” The Waymo driver, for example, or even Tesla’s misnamed Full Self-Driving software, isn’t also capable of simulating next-gen, interactive video games. “We started [Odyssey] in late 2023 to solve for a general world model, something that would predict how the world evolves over time in all applications,” Cameron said, “not just for driverless cars.”

Of course, it’s not perfect yet—as a few simple tests confirmed. But they think it’ll keep getting better. Cameron described the world models of today as analogous to the era of GPT-2: “It could do impressive things, but very quickly you could break it, and you’d find that it didn’t do everything as you wanted it to.” But Cameron is convinced that world models are on the brink of their own ChatGPT-style breakthrough. “I think this year you’ll see a GPT-3–equivalent model, something that captures the imagination of the world,” he said. A model, in short, that “just changes everything” for a given field.

We had been speaking for about 20 minutes when Cameron gestured around the room we were in and contended that if we “tried to describe this room in text, we’d probably be able to write out a few sentences, but we’d probably miss a ton of detail,” like the way the shadows from the rain-slicked window splintered around the tabletop. As powerful as L.L.M.s have become, he said, “there is something about learning from language that misses the world.” If we instead took a video of that room, he continued, the model would have access to a far richer information source.

At a very basic level, the world models concept is not unlike how babies learn by observing and interacting with their physical environment. At Odyssey, Cameron said, the goal is to continue improving their models’ ability to approximate the properties of objects and motion until they “get closer and closer to a full realization of physics. Think about putting a machine in front of a TV screen that plays every video of anything that you could possibly imagine, learning the world that way.”

As with their L.L.M.-evangelizing competitors, the Odyssey guys adhere to the gospel of scale. “On evals, it’s consistently better across all dimensions as you scale,” Cameron said. To illustrate this point, he described an internal experiment in which they asked Odyssey 2—a smaller model than 2-Pro, which came out a few months later—to mix blue and yellow paints in a simulation. The result was that the model randomly selected an incorrect color. But the Odyssey 2-Pro model, which Cameron said is five times larger than its predecessor, successfully came up with green. Infant-level intelligence, perhaps, but he’s certain this trend will continue: “As you scale way beyond the scale of today, they’re going to learn things about the world that are very interesting and not explicit in the training data.”

Odyssey’s co-founders claim their advantage is in the vast reserves of training data that the company has spent the past two years gathering, in addition to “unique” training approaches and custom algorithms. Cameron declined to elaborate further, but suggested that their algorithmic approach—not the algorithms, themselves—was heavily inspired by their work in self-driving. “I think in five years’ time, we’ll look back and say there was a straight line from self-driving cars to general world models,” he said.

He described world models as a “step in the right direction” toward human-level, energy-efficient intelligence. Since the training data is far richer than text, he said, the models “just don’t need to be as large” as L.L.M.s. “That means world models will be more efficient than language models,” Cameron said. “Maybe the world models just keep getting more efficient. Or there’s some other thing that needs to be invented to get closer to human levels of intelligence and energy efficiency.”

Though skeptical researchers have told me that world models, while interesting, seem to lack a genuine near-term use case, Cameron is confident they will soon begin to emerge. He enumerated a range of applications for Odyssey’s technology, arguing that world simulators will “reinvent gaming” as well as robotics, with implications stretching into defense, healthcare, and live events. Hawke also contended that L.L.M.s have given world models a head start. “There are some obvious use cases that we talked about here, but I bet there’s a ton we haven’t thought about,” he said. “And in many respects, the advantage of treating it this way is that people can and will explore those, and frankly, we’ll learn from it. If we look at the major use cases of L.L.M.s today, I suspect half of them were predicted back then, but half weren’t, and I think it’ll be the same for us.”

There are plenty of known unknowns when it comes to all the ways such models can be misused. It’s also somewhat more difficult to devise safeguards for world models, due to their real-time nature. Still, Cameron said, it’s an area “we’ve made really great progress on.” Yes, the frightening notion of an autonomous robot equipped with world-model intelligence might distract from conversations about investment and efficacy, but Cameron has made peace with this reality. “It’s almost impossible to avoid the debate,” he said. “I just think about the good that this model will do, and I think that’s the thing we optimize for.”

Then there’s the seductive pull of discovery and invention for its own sake. “My motivation stems from conviction that this should be a thing,” Hawke added. “This just should exist. There are so many things about our lives as consumers and individuals that I think would be positively improved, and if I can spend my time developing that kind of foundational technology, why would I not?”

That’s all for today. I’ll see you on Thursday.

Ian

Need help? Review our FAQ page or contact us for assistance. For brand partnerships, email ads@puck.news.

You received this email because you signed up to receive emails from Puck, or as part of your Puck account associated with {{customer.email}}. To stop receiving this newsletter and/or manage all your email preferences, click here.

Puck is published by Heat Media LLC. 107 Greenwich St., New York, NY 10006