Understanding the Core Limitations of Large Language Models: Insights from Gary Marcus

In a recent episode of ODSC’s Ai X Podcast, which was recorded live during ODSC West 2024, Gary Marcus, an influential AI researcher, shared a critical perspective on the limitations of large language models (LLMs), emphasizing the need for true reasoning capabilities in AI.

Gary Marcus is a leading voice in artificial intelligence and is well known for his challenges to contemporary AI. He is a scientist and best-selling author and was the founder and CEO of Geometric.AI, a machine learning company acquired by Uber. A Professor Emeritus at NYU, he is the author of five previous books, including the bestseller Guitar Zero, Kluge, and Rebooting AI, one of Forbes’s seven must-read books on AI.

Marcus’s views provide a deep dive into why LLMs, despite their breakthroughs, are not suited for tasks requiring complex reasoning and abstraction. This blog explores Marcus’s insights, addressing LLMs’ inherent limitations, the need for hybrid AI approaches, and the societal implications of current AI practices.

You can listen to the full podcast on Spotify, Apple, and SoundCloud.

Gary Marcus on The Core Mechanism of LLMs: Pattern Recognition, Not Reasoning

Gary Marcus argues that LLMs operate fundamentally on pattern recognition rather than true reasoning. Unlike human cognition, which combines experience with logical inference, LLMs rely on recognizing patterns in vast data corpora. Marcus likens this approach to the “Mad Libs” game, where language is produced based on statistical patterns rather than logical coherence. Although effective in generating text and even mimicking certain reasoning tasks, LLMs often falter with tasks requiring genuine logical reasoning.

For example, Marcus points to classic “river crossing” puzzles. While these problems have numerous variations in datasets, any small change in the setup often results in nonsensical answers from an LLM, as it lacks the cognitive structure to logically deduce solutions beyond its learned patterns. This, he says, underscores the distinction between “pattern-matching” (what LLMs excel at) and reasoning, which requires an ability to understand principles and abstract rules.

Scaling as a Mask, Not a Solution

As AI researchers continue scaling up models to increase data processing power, Gary Marcus raises a cautionary point: scaling does not address the core issue of understanding. Scaling may improve an LLM’s ability to handle more data but does not inherently confer the ability to reason. Marcus notes that even in areas with well-structured data, such as chess, LLMs fail at the fundamentals, occasionally making illegal moves because they don’t “understand” the rules in a conventional sense — they simply mimic patterns.

The scaling of LLMs also leads to increasingly unsustainable energy consumption. Marcus uses recent examples of large tech firms investing in nuclear power to sustain data centers as evidence of AI’s growing environmental cost. He points to the irony of “green” companies resorting to nuclear energy to support AI’s expansion, suggesting that instead of scaling, it may be more prudent to develop fundamentally different AI approaches that are both efficient and capable of reasoning.

The Case for Hybrid AI Models

A significant portion of Gary Marcus’s discussion revolves around hybrid AI as a necessary evolution. He argues for integrating symbolic reasoning systems with deep learning models to create AI systems that combine data-driven pattern recognition with logical reasoning capabilities. This hybrid approach can overcome LLMs’ limitations by adding reasoning modules that function alongside traditional LLMs.

Marcus illustrates the potential of hybrid AI with AlphaFold2, DeepMind’s protein-folding AI, which uses a combination of deep learning and symbolic algorithms to tackle complex problems in ways that neither approach could achieve alone. However, he notes that a fully general-purpose reasoning system — capable of applying logical rules to diverse situations — is still far from reality.

This vision of hybrid AI echoes cognitive scientist Daniel Kahneman’s concept of “System 1” (fast, intuitive thinking) and “System 2” (slow, deliberative reasoning). Marcus sees deep learning as akin to System 1: efficient, but lacking in depth and reliability. For AI to be truly effective, both systems need to be integrated, creating models that are not only quick but also capable of abstract thought and reliable inference.

The Ethical and Social Costs of Generative AI

Gary Marcus goes beyond the technical aspects to address the societal implications of current AI models. He is particularly critical of how companies like OpenAI and Google have shifted the ethical and environmental costs of AI onto society, citing their heavy data usage, environmental footprint, and copyright infringement issues. Drawing a parallel to historical environmental abuses, Marcus argues that AI companies “privatize the profits and socialize the costs.”

Generative AI, Marcus notes, has also contributed to the misinformation problem, with LLMs producing “deepfakes” and misleading content that can undermine public trust and democratic processes. This, according to Marcus, is just one of many social costs of generative AI. In a world where LLMs might scrape and misuse personal data, Marcus sees a growing need for robust regulations to prevent misuse and exploitation.

Regulatory Changes Needed for AI Safety

Gary Marcus highlights the recent veto of the California AI safety bill as a missed opportunity to address these ethical concerns. He argues for a pre-flight safety system for AI — similar to the FDA’s process for pharmaceuticals — where AI technologies must demonstrate their safety and efficacy before mass deployment. In his view, allowing AI companies to operate with minimal oversight risks unpredictable consequences, with AI firms holding unchecked power over critical aspects of society.

A key regulation Marcus advocates for is around data privacy, as LLMs pose serious risks of data misuse. Companies need to handle consumer data with transparency and responsibility, but he suspects some, like OpenAI, are on the path toward becoming surveillance companies, amassing data without adequate user consent.

Moving Forward: Toward Responsible AI Development

Gary Marcus closes by urging a shift from unrestricted AI development to responsible, regulated AI that prioritizes societal benefits over private gains. He suggests that a hybrid approach — one combining scalable data-driven systems with structured reasoning — could pave the way for more responsible and practical applications of AI.

As AI continues to evolve, Marcus’s insights are a valuable reminder that technological advancement must be balanced with ethical considerations and respect for human needs. Without thoughtful development and regulatory guidance, the promise of AI could be overshadowed by the social and environmental costs it incurs.