Yan LeCun on the future of AI: Why Large Language Models are a dead end and what his startup is proposing

MarGib September 02, 2026

🌐 🇵🇱 Polski · 🇬🇧 EN

Yan LeCun, Turing Award laureate and Chief AI Scientist at Meta, argues that current language models have no chance of achieving true intelligence. His startup, Objective Reasoning Systems, is developing the revolutionary JEPA architecture, designed to solve problems with reasoning and efficiency. Is this the beginning of a new era in AI?

Abstrakcyjna wizualizacja architektury JEPA Yana LeCuna jako mózgu AI z modułami percepcji, przewidywania i planowania — JEPA architecture - a new vision for artificial intelligence from Yan LeCun

Why does Yan LeCun consider LLMs to be "dumb" models?

In June 2026, Yan LeCun gave an interview to the BBC in which he ruthlessly assessed the current generation of artificial intelligence. In his view, Large Language Models (LLMs), despite their impressive text generation capabilities, suffer from fundamental limitations:

Lack of common sense - they do not understand the basic principles governing the physical world
Lack of causality - they cannot logically explain the relationships between events
Excessive energy consumption - training models requires as much electricity as a small town
Lack of planning ability - they cannot predict the consequences of their actions

LeCun emphasizes that LLMs operate like "advanced autocomplete," predicting the next words based on statistics rather than actual understanding. This approach, in his opinion, will never lead to the creation of Artificial General Intelligence (AGI).

Criticism from other experts

LeCun's views are not isolated. In March 2026, Geoffrey Hinton expressed similar concerns regarding hallucinations and data manipulation in LLMs. Yoshua Bengio, meanwhile, points to the need for hybrid systems that combine machine learning with symbolic reasoning.

At the same time, there are voices of skepticism. Sam Altman of OpenAI believes that scaling existing models remains the most promising path for AI development. Anima Anandkumar of NVIDIA points to the lack of public benchmarks that would allow for an objective evaluation of alternative approaches.

Objective Reasoning Systems: the startup set to change AI

In November 2024, LeCun founded the startup Objective Reasoning Systems (ORS), which is intended to be the answer to the limitations of current AI systems. The company is developing an architecture called Joint Embedding Predictive Architecture (JEPA), described in a research paper from September 2025.

How does JEPA differ from LLMs?

Feature	LLMs (e.g., GPT-4)	JEPA (ORS)
Operational basis	Statistical text prediction	World representation learning
Reasoning	Lack of common sense	Hierarchical, causal
Input data	Text	Multimodal (text, video, sensory)
Efficiency	High energy consumption	10-100x more energy-efficient

How does JEPA work?

The JEPA architecture consists of three main modules:

Perception module - analyzes data from sensors, video, and text
Prediction module - simulates future states of the world (e.g., "what happens if I drop the glass?")
Planning module - generates sequences of actions based on predictions

The key innovation is hierarchical learning - the system first learns simple concepts and then builds more complex world models upon them. This approach is intended to mimic the way the human brain processes information.

Project status and funding

ORS raised $120 million in a seed round in April 2026 from investors such as Meta Ventures, Andreessen Horowitz, and Sequoia Capital. The company collaborates with leading research institutions, including MIT and ETH Zurich.

At the ICML 2026 conference in July, a JEPA prototype running on visual and textual data was presented. However, there is a lack of public benchmarks comparing the effectiveness of the new architecture with existing LLMs.

Potential applications of JEPA

The new technology could find application in many industries:

Robotics - collaboration with Boston Dynamics on robots capable of adaptive motion planning
Medicine - systems diagnosing diseases based on medical images and patient data (collaboration with Massachusetts General Hospital)
Autonomous vehicles - tests with Waymo on improving the understanding of other road users' intentions
Education - adaptive AI tutors for schools (partnership with Khan Academy)

Meta plans to integrate JEPA with Ray-Ban Meta smart glasses in 2027, which could revolutionize interaction with wearable devices.

Challenges facing ORS

Despite promising prospects, LeCun's startup faces serious challenges:

Technical

Scalability - will JEPA handle large datasets as well as LLMs?
Interpretability - how to explain the decisions made by the system?
Lack of benchmarks - lack of tests on datasets like MMLU (Massive Multitask Language Understanding)

Competitive

ORS must compete with giants such as:

DeepMind (Google) - Gemini 2.0 and AlphaFold 4
Anthropic - Claude 4 focusing on AI safety
OpenAI - GPT-5 and further scaling of LLMs
NVIDIA - Project GR00T for robotics

The advantage of ORS may be its focus on multimodality and reasoning, rather than just scaling text models.

Financial

Budget for 2026-2027 is approximately $200 million
Competition for talent - DeepMind and Google Brain offer higher salaries
Pressure for quick commercial results

AI development outlook until 2030

LeCun presents an optimistic but cautious vision of the future:

2026-2027 - first commercial applications of JEPA in robotics and medicine
2028 - hybrid systems combining JEPA with LLMs for better language understanding
2030 - possible achievement of "weak AGI" in narrow domains

At the same time, he emphasizes that human-level AGI is still decades away. Other experts have different forecasts:

Ray Kurzweil predicts a "technological singularity" by 2045
Stuart Russell warns of the risk of losing control over AI systems

Key milestones (2026-2030)

Year	Expected achievement	Probability
2026	First robots with JEPA in industrial environments	70%
2027	AI systems with common sense in narrow domains	50%
2028	Hybrid models (JEPA + LLMs) in consumer products	60%
2030	Weak AGI in selected applications (e.g., medicine)	40%

Summary: Is JEPA the future of AI?

Yan LeCun's JEPA architecture represents an ambitious attempt to overcome the fundamental limitations of current AI systems. By focusing on causal reasoning, multimodality, and energy efficiency, it could open up new possibilities in robotics, medicine, and education.

However, ORS faces many challenges - from scalability to competition with tech giants. The success of JEPA will depend on whether the startup can prove its advantage in practical applications and convince the market to move away from the currently dominant LLMs.

Regardless of the outcome of this competition, LeCun's criticism and his alternative approach prompt reflection on the direction of artificial intelligence development. Do we really need increasingly larger language models, or is it time for a fundamental paradigm shift? We will know the answer in the coming years.

In the context of these changes, it is also worth following other tools and platforms supporting AI development that may play a key role in the implementation of new architectures.