The AI world is racing forward with near-religious confidence in large language models, yet one of the field’s most influential pioneers is already drafting their obituary. Yann LeCun – the Turing laureate whose ideas underpin much of today’s AI revolution argues that the systems dominating headlines and valuations may be little more than an evolutionary cul-de-sac, destined for obsolescence within five years.
His claim forces a harder, more uncomfortable question than most are willing to ask: are LLMs like ChatGPT, Gemini, Claude, and Llama truly the future of machine intelligence, or simply an impressive transitional phase before more powerful, world-grounded architectures take their place?
I believe so as the seeds of the new era of the death of LLMs are already beginning.
Let’s delve into LeCun’s provocative forecast and examines why the next era of AI might require something fundamentally different from the models we already hold in such high regard.
Context and Background
The landscape is dominated by towering language models – ChatGPT, Gemini, Llama, and Claude – systems that have kicked off an international arms race of investment, data center construction, and corporate evangelism.
These models have proved themselves remarkably effective, generating an unprecedented amount of hype, speculation about the imminent arrival of AGI, and stratospheric valuations. And this is exactly what makes LeCun’s skepticism so striking: he sees a technological ecosystem straining at the limits of its architecture just as the world assumes it’s about to reach its peak.
Technical Limitations of Current LLMs
Large language models are amazing at being fluent, not understanding – that’s tied into their fundamental limitation: token-by-token generation. An LLM doesn’t plan, it doesn’t build a mental model, and it certainly does not reason through cause and effect – discretely, all it can do is predict the most statistically probable next token.
This is very much like, as Daniel Kahneman points out, what he calls System 1: rapid, automatic, and reactive pattern recognition as opposed to slow and deliberate thinking. The end result is a system that can make articulate responses without ever performing the cognitive equivalent of “thinking ahead,” which explains how something like LLMs can sound smart even as they bomb on any task requiring actual reasoning, constraint satisfaction, or hierarchical planning.
This restriction is more transparent when we consider the underlying problem – LLMs lack an actual grounded understanding of the world. Their intelligence is all constructed out of text – dumb symbols, scraped abstractions, descriptions of the reality rather than the thing itself.
Understanding in humans is based on ongoing perception: seeing, sensing the position of one’s own body and objects in space, navigating spatially, and experiencing touch. Word-trained machines do not.
This gap echoes Moravec’s paradox: the counterintuitive observation that the “hard” intellectual tasks we consider advanced – algebra, logic, standardized tests – are trivial for computational systems, while the “easy” skills that even a toddler masters – grasping an object, balancing, reading emotional cues – require a rich, embodied world model.
LLMs inherit this asymmetry. They can summarize a physics textbook but cannot feel gravity; they can explain how to drive a car but cannot build an internal simulation of traffic dynamics; they can describe an emotion but cannot map it to live experience.
The results are small islands of brilliance that are floating on a huge sea of missing fundamentals. When LLMs pass the bar exams or Olympiad math problems, the success reflects dominance over textual patterns – not general intelligence.
Such achievement takes place in domains that are previously formalized, structured, and symbolically encoded. But when you take LLMs outside of those manicured settings, they flounder.
They cannot generalize reliably to novel scenarios, infer causal structure from raw perception, or build the kind of flexible, embodied understanding that humans acquire through direct interaction with the world. The models may dazzle with coherent prose or rigorous-looking logic, but beneath the surface lies an intelligence constrained by its training modality – brilliant with symbols, blind to reality.
The Emerging Next-Generation AI Paradigm
The single most interesting contender on today’s LLM-dominated landscape is the Joint Embedding Predictive Architecture, or JEPA – a method that is much closer to the machinery of real intelligence than any product of scaling token predictors.
JEPA does not rely on the belief that intelligence results from trying to predict the next word or reconstruct every pixel. It doesn’t force a system to paint detailed pictures of the world, but rather, it forces a system to build abstract compressed representations of the world and then predict what those representations will do next.
This shift is profound.
Instead of being buried in raw sensory data – text or pixels – JEPA learns the structure behind that data: spatial layout, object permanence, causality, force, motion, intent. It’s not predicting surface phenomena; it’s a prediction of this underlying state that you observe.
In practical terms, this means a JEPA-based model might learn that a cup exists after someone puts their hand over it, or that when a ball rolls off the table, it will fall, not evaporate – concepts that today’s LLMs and even vision-language models commonly misconstrue due to their lack of grounding in physical continuity.
This architectural change simultaneously removes the fundamental limitations hampering LLMs. As JEPA works in a representation space, it does not throw away the capacity for modelling noise or irrelevant information. It is instead all about training world models – inner simulations that enable an AI system to think through potential outcomes, plan out a sequence of actions, and generalize across new scenarios.
Just as Kahneman’s System 1 – are caught in Kahneman’s System 1 – reaction, association, fluency, but superficiality – JEPA is embracing System 2 – deliberate, goal-directed cognition that can look ahead at what it plans to do and adjust its plan and think about unseen possibilities.
An agent based on JEPA can query, implicitly: “If I do this action, what effect will it have in the world? The LLMs can only inquire: “What word typically follows this string of words?” The distinction is between imitation and understanding.
The benefits of such a paradigm are broad. An AI system that has learned a world model can traverse environments, interact with objects, reason about strategies, and react to unknowns – all this far beyond what even the largest LLMs can do.
Scaling LLMs leads to fluency; scaling JEPAs may result in competence. That’s why, increasingly, researchers are looking at JEPA and other types of predictive‑representation architectures as a blueprint for the next generation of AI: ones that don’t just speak convincingly about reality but can model, reason about, and eventually act within it.
Comparative Perspectives in the AI Field
More widely in the AI world, LeCun’s critique is echoed and contested in the views of other influential thinkers who have been cautioning for years against thinking about intelligence as a pure computational abstraction separated from the physical world.
Robotics pioneer Rodney Brooks has often made the point that actual popular culture narratives of AI assume too much about what disembodied systems can achieve and perform, stressing the dependent nature of real intelligence upon sensorimotor grounding. For Brooks, intelligence arises from interaction, perception, feedback, iteration – not linguistic pattern‑matching or statistical fluency.
Neuroscientist David Eagleman expresses this view from a biological standpoint, arguing that the organ we carry about in our noggins is basically a prediction machine that has been crafted by long-term, extended over time and space (embodied) experience.
Any system developed on text alone or sanitized data inherits a narrow, desiccated slice of reality and will be challenged to practice the flexible, common-sense intelligence that humans get from their ability to live in – and constantly learn about – the actual physical world.
These realistic views are sharply different from the risk narratives that exist in some corners of the AI community. Long-term collaborator and fellow Turing Award winner Geoffrey Hinton has been growing more outspoken about the potential for advanced AI systems to outstrip human control in ways that raise existential fears about misaligned superintelligence. The threat, Hinton posits, has to do with the runaway nature of capability: When systems are smarter than human beings, their goals could splinter off in directions we can’t forecast or control.
LeCun is on the other side of this divide. His optimism is not naive or complacent – it’s also grounded in a technical truth, intelligence on its own does not equal motivation, power to dominate and threaten. A more powerful machine is no more inherently dangerous, he says, than a better-designed tool. Whether an AI system does good or bad is entirely a function of the purposes we instruct it to pursue, the constraints with which we equip it, and the social mechanisms we adopt for its use. Rather than cautioning against hypothetical apocalypses, LeCun foresees an era of augmented intelligence: systems that can help us reason, enhance human decision‑making, and act as collaborative partners rather than competitors.
The contrast between those two visions has enriched the debate about AI’s own future. Brooks and Eagleman are reminding us that without a body, intelligence is thin. Hinton cautions that capability without sober restraint could be dangerous. And LeCun maintains that intelligence without agency is only a tool, and the true task, he says, isn’t to forestall superintelligence so much as it is to make sure the systems we create comprehend the world as richly as humans do, even as they meet some end goal we set for them.
Societal Implications and Future Directions
The next wave of AI systems is set to transform human work not by replacing it with automation, but by elevating us up the cognitive stack – from operators to orchestrators. As AI models improve at planning, reasoning, and interacting with the world, including discrete objects within it, humans are more likely to serve as managers, supervisors, and strategic deciders directing sets of smart tools. Instead of being pitted against machines in increasingly operation-specific competitions, we will marshal them, set goals for them, audit and control their outputs, and include their powers within larger human projects.
This change echoes earlier technological upheavals: our industrial machines didn’t put us out of work, but rather made it easier for us to act as managers over machines doing the physical labour; correspondingly, world‑model‑based AI might raise human cognition by allowing us to sit in positions where judgment, ethics, and context – not raw computation – are what separates success from failure.
But this doesn’t mean LLMs will go away entirely just yet. Even if, because of their limitations, LLMs are not suitable as engines for general reasoning, there may still be a role for them as linguistic interfaces: systems that can translate what is in our mind into well-formed language; break down complex thoughts into actionable pieces of information (for example by summarizing long articles into headlines); and might serve as intermediaries between humans and more ground-worlded AI modules.
In a hybrid model, language would be that thin layer on top – a communication interface rather than the basis of intelligence. Similar to how Broca’s area in the human brain processes production rather than comprehension, LLMs may remain as special-purpose processors even inside a larger cognitive ecosystem that does articulate summarization, drafting, and knowledge retrieval work while deeper reasoning takes place elsewhere.
Yet in order for such a future to be good for all, the path that AI takes needs to remain open, ethical, and accessible. The centralization of power among a few corporations jeopardizes the possibility that systems will be designed to serve the common interest, rather than commercial or geopolitical imperatives.
Open-source models such as Meta’s Llama are essential to counter this tendency, allowing researchers, start-ups, and civil society institutions alike to build on a common base rather than acquiescing in a closed technological monoculture.
Responsible innovation isn’t nice to have – transparent data sets, auditable architectures, safety protocols, and governance are not indulgences we can add when it gets convenient for us with ever more autonomous systems. Why this should matter is that the next generation of AI, unconstrained by such safeguards, could exacerbate inequality, entrench surveillance, and shore up autocratic rule.
The societal consequences of new AI paradigms therefore depend not only on what we invent but on how we choose – the values, principles, and applications, and who is included in the decision-making itself. If we manage it responsibly, next‑gen AI will augment human capacity, leading to the democratisation of intelligence itself.
Left unaddressed, it might only grow more powerful and further limit the range of those who benefit from it. The stakes are not just technological; they are civic, ethical, and human.
Conclusion
In the end, the tension between today’s LLM‑driven breakthroughs and tomorrow’s world‑model‑based architectures reveals a field on the edge of its next transformation. The limitations of current language models – rooted in token prediction, lack of grounding, and narrow symbolic proficiency – stand in stark contrast to the promise of systems like JEPA, which aim to learn the underlying structure of reality and reason within it.
As these next‑generation models grow up, they will move AI’s role from clever text generator to genuinely adept cognitive partner, with humans in lanes of supervision, interpretation, and strategic direction rather than head‑to‑head competition. And the machine won’t replace humans, but augment them – the machine can do the dirty cognitive work – weighing options, drawing predictions, and so on, while people will provide judgment, ethics, and direction. The question now is whether we are going to guide this transition actively with open hands and due care, or be guided by it defensively.
Readers should stop not just to marvel at the technology, but to consider what kind of AI‑infused world we want to make, and each of us will be required to help steer that future.
Found this article interesting?
1. Follow Dr Andrée Bates LinkedIn Profile Now
Revolutionize your team’s AI solution vendor choice process and unlock unparalleled efficiency and save millions on poor AI vendor choices that are not meeting your needs! Stop wasting precious time sifting through countless vendors and gain instant access to a curated list of top-tier companies, expertly vetted by leading pharma AI experts.
Every year, we rigorously interview thousands of AI companies that tackle pharma challenges head-on. Our comprehensive evaluations cover whether the solution delivers what is needed, their client results, their AI sophistication, cost-benefit ratio, demos, and more. We provide an exclusive, dynamic database, updated weekly, brimming with the best AI vendors for every business unit and challenge. Plus, our cutting-edge AI technology makes searching it by business unit, challenge, vendors or demo videos and information a breeze.
- Discover vendors delivering out-of-the-box AI solutions tailored to your needs.
- Identify the best of the best effortlessly.
Anticipate results with confidence.
Transform your AI strategy with our expertly curated vendors that walk the talk, and stay ahead in the fast-paced world of pharma AI!
Get on the wait list to access this today. Click here.
4. Take our FREE AI for Pharma Assessment
This assessment will score your current leveraging of AI against industry best practice benchmarks, and you’ll receive a report outlining what 4 key areas you can improve on to be successful in transforming your organization or business unit.
Plus receive a free link to our webinar ‘AI in Pharma: Don’t be Left Behind’. Link to assessment here
5. Learn more about AI in Pharma in your own time
We have created an in-depth on-demand training about AI specifically for pharma that translate it into easy understanding of AI and how to apply it in all the different pharma business units — Click here to find out more.
