# A Sentence is Worth a Thousand Pictures: Can Large Language Models Understand Human Language?

**Authors:** Gary Marcus, Evelina Leivada, Elliot Murphy  
**arXiv:** [2308.00109](https://arxiv.org/abs/2308.00109)  
**Date:** 2023-07-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2308.00109v1/

## Key points

- Artificial Intelligence applications have demonstrated potential in language-related tasks relying on next-word prediction.
- Large language models are associated with claims about human-like linguistic performance, advancing understanding of human language.
- The study analyzes the contribution of large language models as theoretically informative representations compared to mechanistic tools.
- Large language models may lack deep insights into the mechanisms of human language processing and theoretical grounding.
- Further research is needed to bridge the gap between theoretical understanding and practical application.
- Despite limitations, large language models offer valuable insights into AI in language-related tasks and natural language processing advancements.
- Important abilities are missing from current models that need to be addressed for enhanced performance and representation of human linguistic abilities.
- More work is needed to develop theoretically informed representations and address limitations for further advancements in AI applications.

## Summary

Artificial Intelligence applications have demonstrated immense potential in language-related tasks that rely on next-word prediction. The current generation of large language models has been associated with claims about their human-like linguistic performance, making them a crucial step towards Artificial General Intelligence and a significant advancement in understanding the cognitive and neural basis of human language. In this study, conducted by Gary Marcus, Evelina Leivada, and Elliot Murphy, the authors analyze the contribution of these large language models as theoretically informative representations of a target system compared to atheoretical powerful mechanistic tools. They also identify the key abilities that are still lacking in the current state of development and exploitation of these models. The authors argue that while large language models have demonstrated impressive capabilities in predicting subsequent words in a sentence, they may not provide deep insights into the underlying mechanisms of human language processing. These models serve as powerful tools for generating text but may lack theoretical grounding. The study highlights the need for further research to bridge the gap between theoretical understanding and practical application. Despite their limitations, large language models offer valuable insights into the potential of AI in language-related tasks. They have paved the way for advancements in natural language processing and have sparked discussions about the future development of Artificial General Intelligence. However, there are still important abilities missing from these models that need to be addressed. This study sheds light on areas where improvements can be made to enhance their performance and make them more representative of human linguistic abilities. In conclusion, while large language models have shown promise in next-word prediction tasks and have contributed to our understanding of human language, there is still much work to be done. Further research is needed to develop more theoretically informed representations and address the limitations present in current models. By doing so, we can continue advancing AI applications in language-related tasks and move closer towards achieving Artificial General Intelligence.