AI can write software, analyse documents, generate realistic video and answer questions across thousands of subjects. Yet some of the researchers who created modern AI argue that these systems are still not genuinely intelligent.
Yann LeCun, the Turing Award-winning computer scientist who previously led AI research at Meta, has compared today’s machines unfavourably with animals. His central argument is that a chatbot can discuss almost any subject while lacking the basic physical understanding that allows a young child—or even a rat—to navigate the world.
That criticism does not mean generative AI is useless or fraudulent. It means fluency, knowledge and pattern recognition are not necessarily the same as understanding. The next phase of artificial intelligence may therefore move beyond systems that primarily process language and toward machines that can observe reality, remember experiences, predict consequences and plan actions.
Today’s AI Can Be Brilliant Without Being Generally Intelligent
Large language models learn statistical relationships from enormous collections of text, code, images and other data. This training allows them to produce remarkably useful responses, but it does not guarantee that they have built a reliable model of the reality described by those materials.
A model may solve a difficult mathematics problem and then fail on a simple variation. It can explain how an object should move while making an obvious error when shown an unfamiliar physical situation. It may also generate a convincing answer when it lacks enough information to know whether that answer is correct.
The Stanford AI Index 2026 shows how contradictory the current moment has become. Frontier models now meet or exceed human baselines on several science, mathematics, coding and multimodal evaluations, while adoption continues to rise. At the same time, evaluations are struggling to measure real capability reliably, and documented AI incidents are increasing.
Calling these systems “not smart” is therefore deliberately provocative. Current AI is highly capable within many digital tasks, but its competence remains uneven. It does not consistently possess the adaptable, grounded common sense humans use when entering unfamiliar situations.
Yann LeCun Is Betting That Language Models Will Hit a Limit
LeCun left Meta in 2025 and founded AMI Labs, a company pursuing an alternative to the industry’s heavy dependence on increasingly large language models. The company raised $1.03 billion in 2026 to develop systems focused on real-world understanding, reasoning and planning.
The official AMI Labs research vision says future systems should understand the physical world, maintain persistent memory, reason over problems and remain controllable. Its researchers argue that intelligence begins with interaction with the world rather than language alone.
This does not necessarily mean that chatbots will disappear. Language remains an extraordinarily useful interface for sharing knowledge and giving instructions. LeCun’s argument is that language models may become one part of a larger architecture rather than the complete foundation for advanced intelligence.
A future assistant might use language to communicate, a world model to predict consequences, memory to learn from experience and a planning system to select actions. That combination would be closer to an intelligent agent than a system that produces a new answer from a prompt without retaining a dependable understanding of what happened previously.
World Models Could Become AI’s Next Major Breakthrough
A world model is an internal representation of how an environment works. It allows a system to estimate what may happen next and how different actions could change the outcome.
Humans use this ability constantly. A person expects an unsupported glass to fall, understands that a blocked doorway cannot be crossed and can imagine the likely result of touching a hot surface. Much of this knowledge is learned through observation and interaction rather than through written explanations.
AMI Labs intends to build models that learn abstract representations from cameras and other sensors. Instead of attempting to reproduce every detail of a future scene, the system would concentrate on information important for prediction and action. Its action-conditioned models are intended to simulate possible consequences before an agent makes a decision.
Meta has already demonstrated part of this approach through V-JEPA 2, a video-trained model designed to understand motion, predict future states and support robot planning. Meta reported that the model could perform certain picking and placement tasks with unfamiliar objects in new environments, although significant performance gaps remained between the models and humans on physical-reasoning benchmarks.
World models remain experimental. They have not yet proven that they can provide the breadth, reliability and scalability associated with leading language models. However, the level of investment now entering the field shows that researchers and investors increasingly view real-world prediction as a potential route beyond chatbots.
Robotics Will Test Whether AI Truly Understands Anything
Digital environments allow AI mistakes to remain partly hidden. A generated paragraph can sound reasonable even when its underlying assumptions are wrong. The physical world is less forgiving.
A robot that misunderstands distance, weight, balance or cause and effect may drop an object, damage equipment or injure someone. Robotics therefore offers a demanding test of whether an AI system possesses practical understanding rather than verbal knowledge.
Google DeepMind’s Gemini Robotics research combines language, vision, spatial reasoning and physical action. The company says its models can adapt to unfamiliar objects, respond when their environment changes and replan when an action fails. It has also demonstrated systems controlling different types of robotic hardware.
This broader movement is often called embodied AI or physical AI. Instead of learning only from internet data, machines learn from video, simulations, sensors and interactions with real environments.
Progress is likely to begin in controlled settings such as warehouses, factories, laboratories and industrial facilities. These environments provide predictable equipment, clearly defined tasks and stronger safety boundaries than an ordinary home. Reliable general-purpose household robots remain a considerably harder challenge.
AI Agents Are the Bridge Between Chatbots and Autonomous Systems
While world models and advanced robotics are being developed, AI companies are turning language models into agents that can use tools and complete multi-step tasks.
An agent can search websites, operate software, analyse files, write code and move information between applications. This makes it more useful than a chatbot that only returns text.
OpenAI’s computer-using agent research trained a model to interact with graphical interfaces through screenshots, mouse movements and keyboard actions. Its later ChatGPT agent system combined browsing, research, terminal access and application connections within one workflow.
Agents represent a major change because they allow AI output to produce real consequences. However, connecting an imperfect model to email, financial systems, company data or external websites also magnifies the cost of mistakes.
The next generation of agents will need more than stronger reasoning scores. They will require permission controls, reliable memory, confirmation before sensitive actions and monitoring capable of recognising when the system has misunderstood its objective.
Persistent Memory and Continual Learning Are Still Missing Pieces
Most AI interactions remain temporary. A model receives context, generates a response and then depends on external software to preserve useful information.
Human intelligence works differently. People gradually build an understanding of their environment, update beliefs after mistakes and carry lessons from one situation into another.
Persistent memory could allow an AI assistant to understand long-term projects, remember previous decisions and improve its performance without repeatedly receiving the same background information. Continual learning could help a robot adapt to a particular factory or household rather than remaining frozen at the point when its original training ended.
These capabilities also create serious risks. A system that remembers everything may retain private or inaccurate information. One that changes continuously may become harder to evaluate because its behaviour no longer matches the model originally tested.
That is why AMI Labs includes memory, controllability and safety within the same research vision rather than treating them as separate features.
The Future May Combine Several Types of AI
The next breakthrough is unlikely to come from a single architecture completely replacing every system used today.
Language models are powerful interfaces and knowledge tools. World models may provide physical prediction. Symbolic systems can enforce formal rules. Search engines and databases can supply verifiable information. Robots provide interaction with physical environments, while specialised smaller models can run efficiently on local devices.
The strongest future systems may combine these components. A robot could use a language model to understand an instruction, a vision system to recognise objects, a world model to predict movement, a planner to choose actions and a safety controller to prevent dangerous behaviour.
This hybrid direction would also explain why declaring language models either the final path to general intelligence or a technological dead end may be premature. Their role could change without disappearing.
AI’s Next Chapter Is About Understanding, Not Just Answering
Artificial intelligence is advancing too quickly to be dismissed as merely sophisticated autocomplete. Its scientific, creative and economic capabilities are already substantial. Yet impressive answers should not be confused with dependable understanding.
The race now extends beyond building models that know more words or generate longer reasoning traces. Researchers are trying to create machines that can recognise how the world changes, learn from experience, remember what matters and predict the consequences of their actions.
Whether world models, embodied systems or hybrid architectures will deliver that transformation remains uncertain. What is becoming clearer is that the future of AI will not be judged only by how convincingly a machine can talk.
It will be judged by whether it can act reliably when reality refuses to behave like its training data.
