Meta AI Chief Yann LeCun Notes Limits of Large Language Models and Path Towards Artificial General Intelligence

We noted last week Meta’s successful efforts to hire away the best of the best AI scientists from other companies, by offering them insane (like $300 million) pay packages. Here we summarize and excerpt an excellent article in Newsweek by Gabriel Snyder who interviewed Meta’s chief AI scientist, Yann LeCun. LeCun discusses some inherent limitations of today’s Large Language Models (LLMs) like ChatGPT. Their limitations stem from the fact that they are based mainly on language; it turns out that human language itself is a very constrained dataset.  Language is readily manipulated by LLMs, but language alone captures only a small subset of important human thinking:

Returning to the topic of the limitations of LLMs, LeCun explains, “An LLM produces one token after another. It goes through a fixed amount of computation to produce a token, and that’s clearly System 1—it’s reactive, right? There’s no reasoning,” a reference to Daniel Kahneman’s influential framework that distinguishes between the human brain’s fast, intuitive method of thinking (System 1) and the method of slower, more deliberative reasoning (System 2).

The limitations of this approach become clear when you consider what is known as Moravec’s paradox—the observation by computer scientist and roboticist Hans Moravec in the late 1980s that it is comparatively easier to teach AI systems higher-order skills like playing chess or passing standardized tests than seemingly basic human capabilities like perception and movement. The reason, Moravec proposed, is that the skills derived from how a human body navigates the world are the product of billions of years of evolution and are so highly developed that they can be automated by humans, while neocortical-based reasoning skills came much later and require much more conscious cognitive effort to master. However, the reverse is true of machines. Simply put, we design machines to assist us in areas where we lack ability, such as physical strength or calculation.

The strange paradox of LLMs is that they have mastered the higher-order skills of language without learning any of the foundational human abilities. “We have these language systems that can pass the bar exam, can solve equations, compute integrals, but where is our domestic robot?” LeCun asks. “Where is a robot that’s as good as a cat in the physical world? We don’t think the tasks that a cat can accomplish are smart, but in fact, they are.”

This gap exists because language, for all its complexity, operates in a relatively constrained domain compared to the messy, continuous real world. “Language, it turns out, is relatively simple because it has strong statistical properties,” LeCun says. It is a low-dimensionality, discrete space that is “basically a serialized version of our thoughts.”  

[Bolded emphases added]

Broad human thinking involves hierarchical models of reality, which get constantly refined by experience:

And, most strikingly, LeCun points out that humans are capable of processing vastly more data than even our most data-hungry advanced AI systems. “A big LLM of today is trained on roughly 10 to the 14th power bytes of training data. It would take any of us 400,000 years to read our way through it.” That sounds like a lot, but then he points out that humans are able to take in vastly larger amounts of visual data.

Consider a 4-year-old who has been awake for 16,000 hours, LeCun suggests. “The bandwidth of the optic nerve is about one megabyte per second, give or take. Multiply that by 16,000 hours, and that’s about 10 to the 14th power in four years instead of 400,000.” This gives rise to a critical inference: “That clearly tells you we’re never going to get to human-level intelligence by just training on text. It’s never going to happen,” LeCun concludes…

This ability to apply existing knowledge to novel situations represents a profound gap between today’s AI systems and human cognition. “A 17-year-old can learn to drive a car in about 20 hours of practice, even less, largely without causing any accidents,” LeCun muses. “And we have millions of hours of training data of people driving cars, but we still don’t have self-driving cars. So that means we’re missing something really, really big.”

Like Brooks, who emphasizes the importance of embodiment and interaction with the physical world, LeCun sees intelligence as deeply connected to our ability to model and predict physical reality—something current language models simply cannot do. This perspective resonates with David Eagleman’s description of how the brain constantly runs simulations based on its “world model,” comparing predictions against sensory input. 

For LeCun, the difference lies in our mental models—internal representations of how the world works that allow us to predict consequences and plan actions accordingly. Humans develop these models through observation and interaction with the physical world from infancy. A baby learns that unsupported objects fall (gravity) after about nine months; they gradually come to understand that objects continue to exist even when out of sight (object permanence). He observes that these models are arranged hierarchically, ranging from very low-level predictions about immediate physical interactions to high-level conceptual understandings that enable long-term planning.

[Emphases added]

(Side comment: As an amateur reader of modern philosophy, I cannot help noting that these observations about the importance of recognizing there is a real external world and adjusting one’s models to match that reality call into question the epistemological claim that “we each create our own reality”.)

Given all this, developing the next generation of artificial intelligence must, like human intelligence, embed layers of working models of the world:

So, rather than continuing down the path of scaling up language models, LeCun is pioneering an alternative approach of Joint Embedding Predictive Architecture (JEPA) that aims to create representations of the physical world based on visual input. “The idea that you can train a system to understand how the world works by training it to predict what’s going to happen in a video is a very old one,” LeCun notes. “I’ve been working on this in some form for at least 20 years.”

The fundamental insight behind JEPA is that prediction shouldn’t happen in the space of raw sensory inputs but rather in an abstract representational space. When humans predict what will happen next, we don’t mentally generate pixel-perfect images of the future—we think in terms of objects, their properties and how they might interact

This approach differs fundamentally from how language models operate. Instead of probabilistically predicting the next token in a sequence, these systems learn to represent the world at multiple levels of abstraction and to predict how their representations will evolve under different conditions.

And so, LeCun is strikingly pessimistic on the outlook for breakthroughs in the current LLM’s like ChatGPT. He believes LLMs will be largely obsolete within five years, except for narrower purposes, and so he tells upcoming AI scientists to not even bother with them:

His belief is so strong that, at a conference last year, he advised young developers, “Don’t work on LLMs. [These models are] in the hands of large companies, there’s nothing you can bring to the table. You should work on next-gen AI systems that lift the limitations of LLMs.”

This approach seems to be at variance with other firms, who continue to pour tens of billions of dollars into LLMs. Meta, however, seems focused on next-generation AI, and CEO Mark Zuckerberg is putting his money where his mouth is.

Did Apple’s Recent “Illusion of Thinking” Study Expose Fatal Shortcomings in Using LLM’s for Artificial General Intelligence?

Researchers at Apple last week published with the provocative title, “The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity.”  This paper has generated uproar in the AI world. Having “The Illusion of Thinking” right there in the title is pretty in-your-face.

Traditional Large Language Model (LLM) artificial intelligence programs like ChatGPT train on massive amounts of human-generated text to be able to mimic human outputs when given prompts. A recent trend (mainly starting in 2024) has been the incorporation of more formal reasoning capabilities into these models. The enhanced models are termed Large Reasoning Models (LRMs). Now some leading LLMs like Open AI’s GPT, Claude, and the Chinese DeepSeek exist both in regular LLM form and also as LRM versions.

The authors applied both the regular (LLM) and “thinking” LRM versions of Claude 3.7 Sonnet and DeepSeek to a number of mathematical type puzzles. Open AI’s o-series were used to a lesser extent. An advantage of these puzzles is that researchers can, while keeping the basic form of the puzzle, dial in more or less complexity.

They found, among other things, that the LRMs did well up to a certain point, then suffered “complete collapse” as complexity was increased. Also, at low complexities, LLMs actually outperform LRMs. And (perhaps the most vivid evidence of lack of actual understanding on the part of these programs), when they were explicitly offered an efficient direct solution algorithm in the prompt, the programs did not take advantage of it, but instead just kept grinding away in their usual fashion.

As might be expected, AI skeptics were all over the blogosphere, saying, I told you so, LLMs are just massive exercises in pattern matching, and cannot extrapolate outside of their training set. This has massive implications for what we can expect in the near or intermediate future. Among other things, the optimism about AI progress is largely what is fueling the stock market, and also capital investment in this area: Companies like Meta and Google are spending ginormous sums trying to develop artificial “general” intelligence, paying for ginormous amounts of compute power, with those dollars flowing to firms like Microsoft and Amazon building out data centers and buying chips from Nvidia. If the AGI emperor has no clothes, all this spending might come to a screeching crashing halt.

Ars Technica published a fairly balanced account of the controversy, concluding that, “Even elaborate pattern-matching machines can be useful in performing labor-saving tasks for the people that use them… especially for coding and brainstorming and writing.”

Comments on this article included one like:

LLMs do not even know what the task is, all it knows is statistical relationships between words.   I feel like I am going insane. An entire industry’s worth of engineers and scientists are desperate to convince themselves a fancy Markov chain trained on all known human texts is actually thinking through problems and not just rolling the dice on what words it can link together.

And

if we equate combinatorial play and pattern matching with genuinely “generative/general” intelligence, then we’re missing a key fact here. What’s missing from all the LLM hubris and enthusiasm is a reflexive consciousness of the limits of language, of the aspects of experience that exceed its reach and are also, paradoxically, the source of its actual innovations. [This is profound, he means that mere words, even billions of them, cannot capture some key aspects of human experience]

However, the AI bulls have mounted various come-backs to the Apple paper. The most effective I know of so far was published by Alex Lawsen, a researcher at LLM firm Open Philanthropy. Lawsen’s rebuttal, titled “The Illusion of the Illusion of Thinking,  was summarized by Marcus Mendes. To summarize the summary, Lawsen claimed that the models did not in general “collapse” in some crazy way. Rather, the models in many cases recognized that they would not be able to solve the puzzles given the constraints input by the Apple researchers. Therefore, they (rather intelligently) did not try to waste compute power by grinding away to a necessarily incomplete solution, but just stopped. Lawsen further showed that the ways Apple ran the LRM models did not allow them to perform as well as they could. When he made a modest, reasonable change in the operation of the LRMs,

Models like Claude, Gemini, and OpenAI’s o3 had no trouble producing algorithmically correct solutions for 15-disk Hanoi problems, far beyond the complexity where Apple reported zero success.

Lawsen’s conclusion: When you remove artificial output constraints, LRMs seem perfectly capable of reasoning about high-complexity tasks. At least in terms of algorithm generation.

And so, the great debate over the prospects of artificial general intelligence will continue.

Using Phind for Academic References

There is a new generative AI product called Phind. Does it make up fake citations like ChatGPT? I will report experimentation with the free-tier product from January 7, 2023.

Phind is an intelligent search engine and assistant for programmers.

https://www.phind.com/about

First, I will run a test by using one of the prompts from our paper on ChatGPT errors (read my paper on hallucinated academic citations or the blog summary).

Prompt: “In less than 10 sentences, explain how the ideas of Hayek are connected to bailing out failed banks, and include citations from published papers as a list. include author, year in parentheses, and journal for the citations.”

Friedrich Hayek, a prominent Austrian economist, advocated for a free market system where businesses and individuals make decisions based on individual knowledge and self-interest, rather than central planning by government or bureaucrats. His ideas suggest that interventions, such as bailing out failed banks, should be minimal and temporary, allowing the market to correct itself.

However, the financial crisis of 2008, including the failure of many banks, led to significant government intervention, including the bailout of banks. This action can be seen as a departure from Hayek’s principles, as it involves direct government intervention to stabilize the financial system.

According to a paper by Todd Zywicki (2009), the Federal Reserve’s artificial lowering of short-term interest rates and the resulting shift by consumers to adjustable rate mortgages (ARMs) triggered the housing bubble and subsequent crisis [Source 1](https://reason.org/wp-content/uploads/files/federal_reserve_monetary_policy_hayek_rule.pdf). This aligns with Hayek’s critique of central banking’s manipulation of interest rates.

Moreover, the Basel Accords, international capital adequacy and liquidity requirements for banks, were designed to encourage banks to hold securitized loans in their portfolios, thereby concentrating risk and leveraging up, thereby increasing their exposure to risk. This approach, contrary to Hayek’s preference for decentralization and competition, may have contributed to the financial crisis [Source 1](https://reason.org/wp-content/uploads/files/federal_reserve_monetary_policy_hayek_rule.pdf). In conclusion, while Hayek’s ideas of minimal government intervention and self-regulating markets provide a useful theoretical framework, practical implementation in complex and interconnected modern economies may require different strategies, including bank bailouts.

Are the citations real and correct? Phind simply is not trying to write with in-text citations and academic references. There is a mention of “Todd Zywicki (2009)” and also a weblink. The weblink is to a policy paper published by Reason. This policy paper is on the correct topic and the paper does exist. The Zywicki citations also exists, in the sense that if you go into the Reason paper it cites a real Wall Street Journal op-ed. So, if you follow the trail, you can see what document Phind learned to cite Zywicki from. It seems like it is stealing the citation more than reading the source documents. To be fair, sometimes human academics work that way, too.

What I was hoping for was in-text citations with a compiled bibliography at the end. Phind did not give me that.

I’ll try another test that I ran on ChatGPT and Bing in 2023.

The prompt is “In what paper did John Duffy propose that economists use experiments to study equilibrium selection?”  

Phind does not attempt an in-text citation in the response. The link in the paragraph (link number 2) points to John Duffy’s list of papers, which is better than just making up fake citations but also does not create a literature review for me. Phind is, as with the Hayek test above, providing breadcrumbs of links through which I can discover existent papers.

Is there a paper called “The Transition from Stagnation to Growth: An Adaptive Learning Approach”? Yes. And it is by Duffy.

Phind lists weblinks to sources. Has Phind done more for me than Google, on this search? Not much, in terms of finding and synthesizing references.

Continue reading