# Generative AI vs. AGI: The Cognitive Strengths and Weaknesses of Modern LLMs

**Authors:** Ben Goertzel  
**arXiv:** [2309.10371](https://arxiv.org/abs/2309.10371)  
**Date:** 2023-09-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2309.10371v1/

## Key points

- The paper examines interactive Language Models (LLMs) as cognitive systems
- Focuses on LLMs developed around mid-2023 such as ChatGPT, GPT-4, Bard, and Llama
- Reviews the cognitive strengths exhibited by these LLMs
- Highlights significant differences between the cognitive capabilities of LLMs and humans
- Attributes practical weaknesses in LLMs to architectural limitations
- Argues against relying solely on incremental improvements in LLMs for achieving human-level AGI
- Acknowledges that studying and experimenting with LLMs can provide valuable insights into understanding human-level AGI
- Suggests that LLMs cannot achieve human-level AGI single-handedly but may play significant roles within AGI architectures alongside other ideas
- Discusses social and ethical implications of LLMs, including addressing issues like misinformation and potential economic disruptions
- Emphasizes the need for alternative approaches beyond incremental improvement to advance towards human-level AGI effectively.

## Summary

In the paper "Generative AI vs. AGI: The Cognitive Strengths and Weaknesses of Modern LLMs," author Ben Goertzel provides a detailed examination of interactive Language Models (LLMs) as cognitive systems, with a specific focus on those developed around mid-2023 such as ChatGPT, GPT-4, Bard and Llama. The paper begins by reviewing the cognitive strengths exhibited by these systems. However, it also highlights significant differences between the cognitive capabilities of LLMs and those of human beings. By analyzing the basic cognitive architectures upon which these AI systems are built, it becomes evident that many practical weaknesses in LLMs can be attributed to these architectural limitations. The paper argues against relying solely on incremental improvements in LLMs as a viable path towards achieving human-level Artificial General Intelligence (AGI). Given the constraints imposed by available computational resources, it is deemed impractical to expect substantial progress solely through this approach. Nonetheless, the author acknowledges that studying and experimenting with LLMs can still provide valuable insights into understanding human-level AGI. Additionally, they suggest that while LLMs may play significant roles within AGI architectures alongside other ideas, they cannot single-handedly achieve human-level AGI. The social and ethical implications of LLMs are also discussed from this perspective. While caution should be exercised to address issues like misinformation and potential economic disruptions resulting from powerful technologies like LLMs, the required policies for modern LLMs differ significantly from those that would be necessary if a more credible approximation of human-level AGI were attainable in the near future. Overall, Goertzel's paper offers an expanded understanding of interactive LLMs as cognitive systems by examining their strengths and weaknesses compared to human cognition. It emphasizes the need for alternative approaches beyond incremental improvement in order to advance towards human-level AGI effectively.