This AI Godfather Says Business Tools Built on LLMs Are Doomed

Silicon Valley firms and countless other businesses across the country are spending billions of dollars to develop and adopt artificial intelligence platforms to automate myriad workplace tasks. But top global technologist Yann LeCun warns that the limited capabilities of the large language models (LLM) those apps and chatbots operate on are already well-known, and will eventually be overmatched by the expectations and demands users place on the systems. And when that happens, LeCun says, even more investment will be required to create the superintelligence technology that will replace LLM-based AI—systems he says should already be the focus of development efforts and funding. While that may seem like an outlier view, LeCun, 65, is far from a tech outsider. The Turing Award winner ran Meta’s AI research unit for a decade, only leaving last November to launch his own Paris-based startup, Advanced Machine Intelligence Labs. In addition to disliking the managerial duties that came with the research-rooted Meta job, LeCun said his departure was motivated by his view that Silicon Valley has prioritized short-term business interests over far more important and attainable scientific objectives. Top of those commercial concerns he cites was developing and marketing LLM-based AI chatbots and apps with limited capabilities, rather than superintelligence systems with virtually boundless potential. In contrast to current AI, which uses collected data to provide responses to questions or perform necessary tasks, superintelligence systems take in all kinds of surrounding information they encounter, including text, sound, and visual input. They use all of this not only to teach themselves how to respond to data feeds effectively, but also to predict what’s coming next—a requisite for truly self-driving cars, say, or robots that reason and react as humans would. The vast differences in what current LLM-based AI and emerging superintelligence systems can offer mean that countless businesses are now buying and adapting a technology LeCun predicts is destined to be replaced by something better. And not because it’s more effective—and certainly not less expensive—but because that’s how the tech sector decided the fastest profits were to be made. Human-level intelligence “There is this herd effect where everyone in Silicon Valley has to work on the same thing,” LeCun told the New York Times recently. “The entire industry has been LLM-pilled… [but] LLMs are not a path to superintelligence or even human-level intelligence.” To be sure, AI apps like OpenAI’s ChatGPT, Microsoft’s Copilot, and Anthropic’s Claude have continually been improving over time, as they automate workplace tasks like emailing, content composition, and research. But LeCun says the fact that their LLM models rely on gathering, digesting, and working from word-based data limits how far they can evolve to approach—much less surpass—human thinking and response capabilities. By contrast, he and fellow researchers at AMI Labs are creating “world models” also trained with sound, video, and spatial data. Over time, they are expected to be able to observe, respond to, and even predict user activity and physical environments in countless workplace settings. And that’s expected to allow them to collect both more and broader ranges of information than humans can and react in ways people would if they had those capabilities. “We are going to have AI systems that have humanlike and human-level intelligence, but they’re not going to be built on LLMs,” LeCun told MIT Technology Review this month, describing the models AMI Labs and other researchers are working on. “It learns the underlying rules of the world from observation, like a baby learning about gravity. This is the foundation for common sense, and it’s the key to building truly intelligent systems that can reason and plan in the real world.” But what does that mean for business owners—not to mention investors—spending huge sums to develop, acquire, and use LLM-based AI apps? If LeCun is correct, all those tools being marketed as the future of workplace productivity will become obsolete in several years and be replaced by the superintelligence tech he believes should have been prioritized in the first place. There’s already some evidence backing LeCun’s view that Silicon Valley has focused on the shorter-term profit objectives of rushing capacity-limited LLM apps to market, despite being aware of the limitations of their effectiveness. For example, a study published last August by MIT Media Lab’s Project Nanda estimated that despite the $30 billion to $40 billion that’s been invested since 2023 to develop or purchase AI platforms, only 5 percent of businesses that bought those automating tools have reported any return on that spending. “The vast majority remain stuck with no measurable [profit or loss] impact,” it said. And despite increasing investment in AI tech by businesses—and swiftly rising use by workers—there’s considerable disagreement on how effective the platforms actually are. According to a Wall Street Journal study, 40 percent of C-suite managers credited the work-automating apps with saving them considerable time each week. By contrast, two thirds of lower-level workers said the tech saved them little or no time at all. LeCun doesn’t appear to regard any ROI or performance questions during this still-early era of AI tech as the problem. He even thinks LLM-based apps are valuable—up to a point. For example, he compliments most apps and chatbots Silicon Valley has developed and sold to businesses as being very useful to “write text, do research, or write code.” AI’s unscalable apps But LeCun says the enormous fortunes and business strategy commitments Silicon Valley has made in what he views as a relatively short-term technological solution ignore the bigger, long-term potential of automating technology’s next phase. Meaning, in cumulative terms, it will make the broader effort to produce and perfect AI more expensive. In his view, much of the money and froth that’s inflated what critics call today’s AI bubble will likely vanish when the models of today’s apps and chatbots can’t be used to build tomorrow’s revolutionary tech. “LLMs manipulate language really well,” LeCun told MIT Technology Review. “But people have had this illusion, or delusion, that it is a matter of time until we can scale them up to having human-level intelligence, and that is simply false.” Ironically, even LLM-based apps using available data concur that superintelligence systems will offer huge advantages when (not if) they supplant today’s AI tools. “While LLMs are incredibly powerful tools for generating text and interacting with humans, a true superintelligence would represent a leap beyond these current systems in terms of understanding, autonomy, adaptability, and practical real-world impact,” ChatGPT replied when asked about its eventual replacement—providing eight major improvements superintelligence tech will offer. When those systems do come online, LeCun says, businesses recognizing their far wider range of applications will have no choice but to buy them to replace outdated LLM-based AI tools they’ve just recently acquired. “Think about complex industrial processes where you have thousands of sensors, like in a jet engine, a steel mill, or a chemical factory,” LeCun told MIT Technology Review. “There is no technique right now to build a complete, holistic model of these systems. A world model could learn this from the sensor data and predict how the system will behave. Or think of smart glasses that can watch what you’re doing, identify your actions, and then predict what you’re going to do next to assist you. This is what will finally make agentic systems reliable.” And superintelligent systems hopefully won’t generate photos of people with six fingers or endless volumes of workplace slop for employees to plow through. BY BRUCE CRUMLEY @BRUCEC_INC