Understanding AI World Models with AMI’s Alex LeBrun

Understanding AI World Models with AMI’s Alex LeBrun

The current AI boom is really built on the huge scaling of LLMs, but are they're getting it wrong? Alex LeBrun, Co-founder and CEO of Advanced Machine Intelligence (AMI Labs), argues that if AI is going to understand the world, model cause and effect, and reason across time, it may need something different: world models. His new company has raised more than $1 billion to try to prove it. He isn't making a technical argument – it's a challenge to the whole logic of the current AI boom. So today, our hosts Autria Godfrey, Stephen Horn and Laila Rizvi are asking Alex LeBrun whether the industry is over-invested in scaling just one dominant paradigm, and whether world models are a real alternative or just a better theory in a market that may have already chosen its winner. Autria kicks things off by asking Alex where LLMs fall short. According to Alex, “If you want to understand or manipulate the real world, then LLMs are not good…across the board.” He explains that LLMs are good in language-first tasks: everything that is discreet and recognized, like mathematics, coding, and information retrieval. But that’s “only one class of problems. It's not everything in the world.” Alex defines what world models are, how they are trained, and what they are most useful for. You’ll hear about how AMI chairman Yann LeCun and his team developed a concept called JEPA, Joint Embedding Predictive Architecture, which is a way to train world models through self-supervised learning. We explore how AMI is using data-rich video to train their world models – and why the vast majority of the videos on YouTube just won’t cut it as source material. Laila asks about the possibility of scaling LLMs to the point where they can lead to implicit world models emerging. Alex disagrees. He points to diminishing returns on the core progress of LLMs in spite of spending billions of dollars. Instead, he sees world models as complimentary to LLMs the way physicians need real world training in addition to only reading books. He even compares how LLMs and world models can mirror the way the human brain works, with different areas of the brain responsible for different functions while working together in tandem. We dive into the economics of AI, from whether further investment in scaling LLMs makes sense to whether world models will catch on with investors. Alex shares some of the challenges he’s faced competing for compute and raising capital in an investment climate where Anthropic has a $30 billion run rate. When Stephen wonders how world models will fit in with agentic AI in areas like healthcare, Alex points out that agentic models are very brittle and usually break when it comes to long term planning. He says 99% accuracy isn’t enough if it means killing 1 out of every 100 patients. World models build an internal representation of the world, which is more deterministic, and will allow for longer term planning with more accuracy. Finally, it’s time for our Rapid Fire segment, where our hosts ask our guests a series of three questions. Autria asks Alex where people will draw the line with AI in their personal lives; Laila asks him for something that's universally accepted in his field that he disagrees with; and Stephen asks Alex what will happen in the future that people aren't talking about now. In our post interview discussion, Autria. Stephen and Laila discuss whether Alex made the case for including world models in the AI economy, including diverting some of the capital being invested in LLMs into world models. We also want to know what you think. Are world models the new frontier in the world of AI, or is scaling LLMs still the best bet forward for seeing all the potential that AI has to offer? Tell us in the comments.

AI and Productivity at Work with Microsoft’s Matt Firestone

AI at work has been sold as a productivity revolution. But Microsoft's latest Work Trend Index points to a more complicated story. Workers are using AI, companies are buying the tools, and agents are beginning to take on more of the execution – but many organizations are not yet built to capture the value. The question is no longer just whether AI can help people work faster. It's whether companies know how to redesign work around it. Is enterprise AI finally moving from experimentation to measurable impact? Or are companies still buying the promise before they know how to change the work? In this episode of Agents of Tech, we’re talking the current state of AI, workplace productivity, and the rise of agents with Matt Firestone, General Manager of Frontier Function at Microsoft, who leads their work on Microsoft 365 Copilot and agents. Matt, who has been closely involved in the company's research on “The Frontier Firm,” starts off by explaining how Microsoft looked at native AI companies to figure out how existing companies can reorganize themselves to be more like them and reach the frontier. He and hosts Autria Godfrey, Stephen Horn and Laila Rizvi discuss frontier professionals, who use AI agents for multi-step workflows and multi-agent systems. According to the Microsoft Work Trend Index, 16% of people self-identify that way. Matt describes the “Transformation Paradox,” where frontier professionals do great things using AI but get resistance from their organizations. The group talks about how the rate of change is much faster than the Industrial Revolution, and how leaders need to listen to and rely on their employees who are ahead of them on the AI adoption curve. They explore barriers to adoption, including technology hurdles, organizational culture, and the relationship of individuals to AI and their career development. When Autria brings up the topic of AI replacing entry level workers, Matt points to data from the MWTI that shows people are using AI for different things than rote, repetitive task work. “49% of the usage of M365 Copilot was higher-order cognitive tasks. So things like deep data analysis, asking multi-shot questions, more sophisticated types of long-running research.” Which means, Matt says, that AI is actually raising the level, scope, breadth, and depth of what an entry level worker can do in their first few years. Laila asks Matt about how to make AI less of a product, where employees use it within limits, and more of a substrate, where frontier professionals can use AI to extract the most value out of it. Matt says that AI is “raising individual ambition” and shifting from task-based knowledge work to outcomes-based thinking. Stephen and Matt talk about how enterprise AI can unlock value in employees as well as organizations. Matt describes the 15x year-on-year growth in agents across their ecosystem, and explains Microsoft’s Customer Zero Program. In our Rapid Fire segment, Autria asks Matt about where people will draw the line with AI in their personal lives; Laila asks Matt for something that's universally accepted in your specialist field that he disagrees with; and Stephen asks Matt what he’s optimistic about? Be sure to stick around for the conversation between Autria, Stephen and Laila about how not every company is as engaged with AI adoption as Silicon Valley may think, and what the cost of falling behind may be. What are you seeing in your offices? Is AI at work genuinely changing how your organization operates, or is it still mostly experimentation at employees' leisure? If you are using AI tools, where are you seeing the real value? Has your company successfully implemented AI strategies that are supported by the proper framework, or are they woefully under-prepared on how to execute AI in your industry? Share some successes and some shortcomings in the comments below.

Who Controls Medical AI and What Do They Want?

Disclaimer: The conversation in this video is for information purposes only and does not constitute medical advice. Always consult a licensed healthcare professional before making any health decisions. Dr. Eric Topol believes AI can predict disease decades before symptoms appear, and he's argued that AI without doctors may outperform doctors with AI. But someone has to set the rules. And right now, the people most likely to write them aren't doctors or patients, but rather insurers, tech companies, and hospital administrators. This week on Agents of Tech we’re exploring prevention, power, and who really controls medical AI, with Dr. Eric Topol, cardiologist, bestselling author, and Founder and Director of the Scripps Research Translational Institute. We start with Autria asking Dr. Topol about accountability when it comes to medical AI. “It has to be accountable,” he says, but “The problem most people don't realize is there's lots of errors by physicians. In the US, you know, 800,000 serious diagnostic errors a year that result in disability or death. So we're trying to improve accuracy.” He describes studies that show that AI without doctors outperforms doctors using AI, and offers some possible reasons that the studies came to those conclusions. Stephen brings up a Swedish breast cancer study that showed a 29% improvement when AI was brought into detecting, and wonders why that kind of result hasn’t led to widescale adoption. Eric breaks down some of the issues with adopting AI, and he and Laila discuss the implementation problem, including the impact of income disparities on access. Dr. Topol and Autria consider the differences between using AI to look for a cure, and what Eric feels is more promising, using AI for disease prevention. He explains to Stephen how AI has already been used not only to predict disease, but also when it will show up! With Alzheimer’s, for instance, there are layers of data we’re not yet using. “There are biomarkers, like the breakthrough one for Alzheimer's disease, p-tau217, that tells us in advance 15 or 20 years about people who are destined to have a high risk.” Later, he goes into more detail about how AI can make a difference in preventing Alzheimer’s and slowing down the “brain clock.” The conversation shifts to who controls medical AI and what their goals are. Dr. Topol describes hospital administrators who only want to use AI to “increase revenue and to use AI to maximize productivity…My biggest concern about the AI era in medicine is we have this great chance to restore a remarkable patient-doctor relationship. We may never see it again for a long, long time, if ever. And we could blow it because of business-centric issues.” Eric tells Laila how AI can give doctors back time they spend on writing notes, and how China is using opportunistic AI – finding things that were not the reason why abdominal and chest CTs were done that doctors miss – to pick up pancreatic cancer before it’s too late. Another question the team addresses is whether medical AI will benefit everyone or just the rich. Dr. Topol says it will be hard work to ensure the democratization of healthcare, and that it’s one of his primary worries. Finally, it’s our lightning round. Autria asks where people will draw the line with AI, and Eric talks about the current public backlash to AI that he feels will fade over time as we resolve their issues. Laila asks Eric what’s widely accepted in his field that he disagrees with, and he says it’s ludicrous that people in genomics say we shouldn't be using polygenic risk scores. And Stephen asks Eric what people aren't talking about now, and he says “no one's really talking about this prevention opportunity. I'm kind of the lone wolf out there.” What about you? If AI could predict your disease decades early, but the price was that that data is sitting in the hands of tech companies and insurers, not necessarily your doctor, would you still want it? Tell us in the comments.

Can AI Create Materials That Never Existed?

CuspAI was co-founded by Max Welling, a pioneer of modern AI. But instead of building chatbots, he's creating and harnessing AI to discover new materials for everything from carbon capture to water purification, plastic alternatives, and more efficient batteries. If this works, it could change everything. But can it? And if so, when? This week, we’re talking science, scale, and when CuspAI will be able to deliver with company co-founder Max Welling. Before talking to Max, hosts Autria Godfrey, Stephen Horn, and Laila Rizvi discuss why AI-powered material creation is so exciting… and why it needs to be addressed with little skepticism. Autria kicks off the interview by asking Max Welling about CuspAI’s plans for 2026. He explains that they’re currently building their platform, and that 2026 will be a time when CuspAI will work with customers to actually design new materials, synthesize them, and put them into the real world. In 2026, they also plan to connect to high throughput self-driving lab experimental facilities to increase the speed of experimentation. Max unpacks how the process works, starting with a customer describing the real world material that they need, to AI agent assessment to see whether anything already exists that can fit the task, to generating entirely new materials that never existed before. They typically generate “hundreds of thousands to millions of those,” which are first tested by a digital twin consisting of very cheap property predictors that quickly assess whether the material could exist in this world, followed by sophisticated molecular dynamic simulations to assess their actual properties in high accuracy. Laila and Max discuss the process of verification in materials science, and Max lays out some of the difficulties that lay between moving from simulation into the real world. Autria asks where their first successes will come. Max predicts that it will be in the semiconductor space, partly because of partners like Hyundai that have the need and the capability to manufacture the new materials at scale. Max explains how semiconductor lithography has gotten to a point where they’re creating the smallest structures possible that are the size of a few atoms, and in order to make that happen they really need new materials. “We are now creating chips that grow in the third dimension, so they become sort of taller.” Stephen brings the conversation around to scalability, the importance of finding partners that can power that growth, and whether that scale can even happen in Europe. Max makes a distinction between making materials at scale and scaling up a company, but says that Europe has the right companies for both. Laila raises the issue of commercial demand versus public good, and Autria asks about the pressure around using AI for the betterment of humanity. Max’s answer: “For me, the only reason I do this is because I do want to make a positive net impact… And I think the same goes for all our employees.” As always, we end with our lightning round questions. Autria asks where people will draw the line with AI, and Max says “We don't want AI to invade our privacy. We don't want AI to be used for mass surveillance. We don't want AI to get us addicted. We don't want AI to manipulate our opinions.” Laila asks Max for something that's widely accepted in his field that he disagrees with. He says that AI superintelligence replacing humans might be a little overhyped at this point. And Stephen asks Max what will happen in 18 months that people aren't talking about now? Max’s answer: “That we can design materials that feel quite exotic right now, with properties that you could not imagine, and it could completely change the world, and hopefully for the better. That's what we are shooting for.” What do you think? Will CuspAI be able to deliver on their promises in 2026? Will AI help us create new materials that benefit humankind in less than a year? Or will it take them longer? Tell us in the comments.

Will AI Help You Live 50 More Years? Immunologist Derya Unutmaz Weighs In

If you survive the next five years, says immunologist Derya Unutmaz, MD, you will live for the next 50 thanks to what AI can accomplish in medical research. But is AI really a silver bullet that solves humanity's most difficult problems? Or does that kind of thinking get us into trouble? Professor Unutmaz is an NIH funded immunologist, with 35 years of published research and more than 100 papers, and a professor at the Jackson Laboratory. But when it comes to what he calls the bio singularity – the moment when the convergence of AI and biotechnology radically transforms human biology – he’s ahead of most of his fellow scientists and researchers. He’s also ahead of most predictions about AGI and SGI, or super general intelligence, by at least a couple of years. But what if he’s right? “In the US alone,” Dr. Unutmaz tells hosts Autria Godfrey, Stephen Horn, and Laila Rizvi, “there are 12 million misdiagnoses every year. Only about 70% of the diseases are diagnosed correctly by medical professionals. And about 700,000 people die or…become sick because of misdiagnoses, okay?... with AI, if you could improve that by 10%, you are saving hundreds of thousands of lives.” For Derya, AI means the totality of artificial intelligence, including LLMs, agentic systems, world models, and more. He likens the different areas of AI to how the human brain works, with different areas managing different tasks. He even likens his own scientific training to the way AI models are trained. He explains that there are multiple levels of Artificial General Intelligence, where general means that you can generalize knowledge. “So for example, if you learn something on one topic, we can somehow generalize that information to learn something completely different, or understand something completely different.” According to Derya, we’re already at AGI Level One, where a system is as good as the top 1% of humanity. He thinks there will eventually be three or four levels of AGI. Dr. Unutmaz says that when Demis Hassabis is describing AGI, he really means ASI, or Artificial Super Intelligence, which is better than human. Derya says that “the Einstein test” – where we train an AI models on pre-1911 knowledge and ask it to recreate the General Theory of Relativity – is unfair, because only a few people in the history of humanity have been capable of such incredible insight. He also thinks we’ll reach a point where most jobs that depend on using a “computer, or your intelligence, or your experiences or expertise could be eventually replaced by AI.” In fact, he says, “I’m a scientist…and I can tell you that current AI models like GPT-5.2 PRO [are] simply better than me.” Derya says we are going to have to rethink the whole fabric of society and the impact of AI will be extremely disruptive. Whether he is right or not about the timeframe for AGI or the bio singularity, Stephen, Laila and Autria agree that the disruption from AI is here, now, and not enough people are addressing it. Do you think AI will cure disease within a decade, or that it's just a dangerous thing for a scientist to claim? Tell us in the comments. CHAPTERS: 00:00 - Will AI Cure All Disease Within 10 Years? 00:55 - AI, Disease and Bio Singularity with Immunologist Derya Unutmaz 01:21 - Why So Much Hype and Uncertainty Around AI and Science? 03:06 - We Can’t Leave This to the Scientists… or the Tech Bros 03:55 - If You Survive the Next 5 Years, Will AI Help You Live 50 More? 04:20 - Tech and AI Capabilities Doubling Every Few Months 07:31 - What is AGI? 08:44 - Are We Prepared for the Biggest Transformation in History? 10:48 - Is AI Hitting a Wall? Is the Einstein Test Fair? 13:58 - I’m a Scientist, and GPT-5.2 PRO Is Better Than Me 15:29 - Will There Always Be a Place for Humans In Science? 18:17 - Is It Unethical for Doctors and Scientists Not To Use AI? 24:55 - Aging Can Be Reversed Says Dr. Derya Unutmaz 25:26 - Is There Any Point in Publishing Scientific Papers Now?