Meta AI boss says big language models won’t catch up to human intelligence

Meta’s AI chief said the large language models powering generative AI products like ChatGPT would never achieve the ability to reason and plan like humans, as he focused on a radical alternative approach to creating “superintelligence” in cars.

Yann LeCun, chief AI scientist at the social media giant that owns Facebook and Instagram, said LLMs had “very limited understanding of logic. . . they don’t understand the physical world, they don’t have continuous memory, they can’t reason in any reasonable definition of the term, and they can’t plan. . . hierarchically”.

In an interview with the Financial Times, he argued against relying on advanced LLMs to try to create human-level intelligence, as these models can only respond accurately to requirements if they are fed the right training data and , therefore, are “fundamentally unsafe. “.

Instead, he is working to develop an entirely new generation of AI systems that he hopes will power machines with human-level intelligence, though he said that vision could take 10 years to achieve.

Meta has poured billions of dollars into developing its own LLMs as generative AI has exploded, aiming to catch up with rival tech groups including Microsoft-backed OpenAI and Alphabet’s Google.

LeCun leads a team of about 500 employees at Meta’s Fundamental AI Research (Fair) lab. They are working toward creating AI that can develop common sense and learn how the world works in human-like ways, in an approach known as “world modeling.”

The Meta AI chief’s experimental vision is a potentially risky and costly gamble for the social media group at a time when investors are looking to see quick returns on AI investments.

Last month, Meta lost nearly $200 billion in value as CEO Mark Zuckerberg vowed to ramp up spending and turn the social media group into “the world’s leading AI company,” leaving Wall Street investors worried about growth of costs with little immediate income potential.

“We’re at the point where we think we’re on the cusp of possibly next-generation AI systems,” LeCun said.

LeCun’s comments come as Meta and its rivals continue with increasingly improved LLMs. Figures such as OpenAI chief Sam Altman believe they provide a vital step towards creating artificial general intelligence (AGI) – the point where machines have greater cognitive abilities than humans.

OpenAI last week released its new fastest model GPT-4o, and Google unveiled a new “multimodal” artificial intelligence agent that can answer real-time questions on video, audio and text called Project Astra, powered by an upgraded version of its Gemini model.

Meta also launched the new Llama 3 model last month. The company’s head of global affairs, Sir Nick Clegg, said her recent LLM had “immeasurably improved skills like reasoning” – the ability to apply logic to questions. For example, the system would assume that a person suffering from a headache, sore throat and runny nose had a cold, but it could also figure out that allergies could be causing the symptoms.

However, LeCun said this evolution of LLMs was superficial and limited, with models only learning when human engineers step in to train it on that information, rather than the AI ​​arriving at a conclusion organically like humans.

“That certainly seems to most people like reasoning — but mostly it’s leveraging knowledge accumulated from a lot of training data,” LeCun said, but added: “[LLMs] are very useful despite their limitations.”

Google DeepMind has also spent several years pursuing alternative methods for building AGI, including methods such as reinforcement learning, where AI agents learn from their surroundings in a game-like virtual environment.

At an event in London on Tuesday, DeepMind chief Sir Demis Hassabis said what the language models lacked was “they didn’t understand the spatial context you’re in. . . thus limiting their usefulness in the end.”

Meta created its Fair Lab in 2013 to look at AI research, employing leading academics in the space.

However, in early 2023, Meta created a new GenAI team, led by Chief Product Officer Chris Cox. He poached many AI researchers and engineers from Fair, and led work on Llama 3 and integrated it into products such as its new AI assistants and image generation tools.

The creation of the GenAI team came after some insiders argued that an academic culture within the Fair lab was partly to blame for Meta’s late arrival to the generative AI boom. Zuckerberg has sought more commercial applications of AI under pressure from investors.

Still, LeCun has remained one of Zuckerberg’s top advisers, according to people close to the company, because of his track record and reputation as one of the founding fathers of AI, winning a Turing Award for his work on neural networks.

“We’ve refocused the Fair toward the long-term goal of human-level AI, essentially because GenAI is now focused on things for which we have a clear path,” LeCun said.

“[Achieving AGI] it’s not a product design problem, it’s not even a technology development problem, it’s a very scientific problem,” he added.

LeCun first published a paper on his world modeling vision in 2022, and Meta has since released two research models based on the approach.

Today, he said Fair was testing different ideas to achieve human-level intelligence because “there’s a lot of uncertainty and exploration in that, [so] we can’t know which one will succeed or end up being taken.”

Among them, LeCun’s team is feeding the systems hours of video and deliberately skipping frames, then having the AI ​​predict what will happen next. This is to mimic the way children learn by passively observing the world around them.

He also said Fair was exploring building “a universal text encoding system” that would allow a system to process abstract representations of knowledge in text, which could then be applied to video and audio.

Some experts are doubtful whether LeCun’s vision is viable.

Aron Culotta, associate professor of computer science at Tulane University, said that common sense had long been “a thorn in AI’s side” and that it was challenging to teach causality models, leaving them “susceptible to these unexpected failures”.

A former Meta AI employee described the worldwide modeling push as “vague fluff”, adding: “It feels like huge flag planting”.

Another current employee said Fair had yet to prove itself as a real rival to research groups like DeepMind.

In the long term, LeCun believes the technology will power artificial intelligence agents that users can interact with through wearable technology, including augmented reality or “smart” glasses and electromyography (EMG) “bracelets.”

“[For AI agents] to be really useful, they have to have something like human-level intelligence,” he said.

Additional reporting by Madhumita Murgia in London