Nilekani “introduced me to India’s dynamic ecosystem,” he said in an interview with Mint on the sidelines of Meta’s first ‘Build with AI Summit’ held in Bengaluru on Wednesday.
“I recently saw innovative ideas during our hackathon (a 30-hour ‘AI Hackathon with Meta Llama’ held in Bengaluru this month). AI tools like Meta AI, accessible through WhatsApp and Messenger, have gained massive traction, with India hosting the largest user community globally,” LeCun said.
“We have also run experiments in rural India, introducing Meta AI, which people adopted immediately. This shows AI’s potential across sectors such as education, healthcare, agriculture, and business,” said LeCun, donning the Ray-Ban Meta glasses, which the company has not yet introduced in India.
Also read | The coming of age of India’s AI-powered humanoids
LeCun’s decision to visit India was also “reinforced by policy experts who urged me to explore the country’s scientific, developer, and governmental landscape, emphasizing that platforms like Llama could have a major impact here,” he said. Llama comprises large language models (LLMs) developed by Meta.
But, he underscored, “what is missing in India is the absence of world-class research labs outside universities”, which he said could motivate students to pursue AI careers within India.
LeCun pointed to a similar transformation in France “when we opened the FAIR (Fundamental AI Research) Lab in Paris. It catalyzed the local AI ecosystem, inspiring students to pursue graduate studies and encouraging companies to set up research labs”, he said.
FAIR Paris now produces around a dozen PhD graduates annually. LeCun believes India could replicate this model.
‘Godfather of AI’
Meta operates widely in India through Facebook, WhatsApp, and Instagram. It also supports small and medium-sized enterprises with digital tools and partners with universities on AI research. Nilekani’s Infosys Ltd, for instance, on Wednesday announced a partnership with Meta to advance innovation in generative AI through open-source projects.
Other than collaborations and provinding open-source AI models to the Indian ecosystem, Meta is also exploring metaverse opportunities.
LeCun, who is also a professor at New York University, is widely regarded for his contributions to deep learning—a machine learning technique. He is also referred to as a ‘Godfather of AI’, along with Geoffery Hinton, who won this year’s Nobel Prize for Physics, and Yoshua Bengio, professor of computer science at the University of Montreal.
All three of them were recipients of the 2018 ACM Turing Award, referred to as the ‘Nobel Prize in Computing’.
Achieving human-level AI will take time. But we’re going to get there. There’s no question.
LeCun is known for his optimistic view of AI, which contrasts sharply with more cautionary voices like those of Hinton or Elon Musk.
Acknowledging that AI systems will evolve to possess common-sense reasoning, planning abilities, and persistent memory, making future deployments easier, LeCun asserted that “achieving human-level AI will take time—optimistically within a decade… We’re going to get there. There’s no question.”
LeCun was categorical, though, that “we are not going to get there with large language models, or what we call autoregressive large language models (that generate text by predicting the next token based on previous ones, processing input sequentially from left to right)”.
Also read | LLM chatbots, search engines will co-exist, says Google’s Raghavan
As intelligent as a four-year-old
LeCun has often said that even the largest LLMs aren’t yet as intelligent as a four-year-old. Future AI, according to him, will need to understand the physical world, and be trained on video and real-world experiences, not just text.
This shift is essential since LLMs are limited in scope and won’t achieve the full range of human intelligence. Meta’s FAIR lab, LeCun said, is focused on developing next-generation AI models beyond LLMs, “which will still rely on deep learning but require new techniques”.
The real breakthrough will come when AI can become smart by understanding the world from firsthand interactions.
Currently, LLMs are trained on up to 20 trillion tokens, covering most public text, but this data is insufficient for them to master all languages, including India’s 700 dialects, according to LeCun.
This points to a need for models to go beyond vast datasets to learn from real-world experiences, similar to how children learn through vision and interaction.
“Research shows that by age four, a child’s visual cortex processes as much information as the largest LLM… but they achieve this in a far shorter timeframe. Therefore, achieving human-level AI will require grounding models in real-world experiences, especially from video and sensory data, just as humans and animals learn,” he said.
“Language will enhance these systems, but the real breakthrough will come when AI can become smart with less data by understanding the world from firsthand interactions. This shift will serve as a foundation for higher intelligence and better generalization,” LeCun elaborated.
Also read | Meet the four musketeers of India’s quantum computing dream
The superintelligent AI
That said, sceptics have warned that the rise of such AI could lead to superintelligent systems beyond human control, advocating for strict regulation to prevent such a scenario. LeCun, however, dismisses these fears as overblown.
LeCun believes these systems won’t threaten jobs or autonomy; instead, they will act as intelligent assistants, amplifying human intelligence. AI will empower people by amplifying their intelligence, he said.
Superintelligent AI could spark a new renaissance, similar to the transformative impact of the printing press.
“Imagine using smart glasses or a smartphone to access virtual assistants anytime, advising you on any topic. Like having a highly skilled staff, these systems can enhance decision-making in academia, business, or politics. Their intelligence, though superior in specific areas, won’t be threatening.
“These systems will be designed so that they do our bidding. They are not going to have any inkling of domination or anything like that. This is where I disagree with some of my friends, like Hinton. Such technology could spark a new renaissance, similar to the transformative impact of the printing press,” LeCun asserted.
Also read | Quantum computers may soon outclass traditional ones, says IBM fellow Jay Gambetta
Is Meta truly open source?
That said, Meta’s AI models are not truly “open source” but more “open weight” since they only share the weights (parameters) but not the source code.
LeCun acknowledged that the concept of “open” in AI is nuanced, involving models, inference code, training code, and data, with varying degrees of openness.
Currently, open weights and inference code are the most accessible, allowing users to download, fine-tune, or adapt models efficiently for various platforms. This approach fosters innovation and enables startups and companies to deploy AI quickly, he explained.
AI, according to him, “is evolving into a common infrastructure, much like how Linux powers the internet. Open-source platforms thrive because community involvement ensures security, faster development, and lower costs. Future AI systems will likely be trained in a distributed way, leveraging local data—essential for multilingual models like one capable of covering India’s 700 languages”.
Future AI systems will likely be leveraging local data—essential for multilingual models like one capable of covering India’s 700 languages.
Open-source hardware also plays a key role. Meta’s Open Compute Project (OCP) sets standards for open server designs, and similar frameworks could emerge for AI.
Training AI models requires powerful graphics processing units (GPUs), an area currently dominated by Nvidia Corp. However, the challenge lies in scaling inference infrastructure—especially in large markets like India—where AI systems must operate at minimal cost and energy consumption to reach millions.
Innovations in hardware and software optimization are crucial, and the incentives for making AI accessible to billions are already in place, making this evolution inevitable, according to LeCun.
Alternatives to the transformer model
LeCun also believes that while the transformer model (which most LLMs are based on) will remain a core building block for future AI, the current approach—using autoregressive, decoder-only models like LLMs—will likely evolve.
“These architectures may be replaced by more advanced systems such as Joint Embedding Predictive Architectures. JEPA, a non-generative model, is better equipped to handle tasks involving video, images, and long-term dependencies, while still incorporating transformers as components,” LeCun said. “These systems can build hierarchical models and run simulations, making them more effective for reasoning and planning.”
He added that AI technologies are evolving rapidly, and LLMs, for instance, might not dominate the landscape five years from now. He added that while deep learning, neural networks, and transformers will continue to play important roles, their configurations and applications will shift.
“As the technology continues to evolve your ability to adapt will be key. So, prioritize learning to learn—it’s this mindset that will ensure you stay ahead of the curve.”
