# Brain in a Vat: On Missing Pieces Towards Artificial General Intelligence in Large Language Models

**Authors:** Yuxi Ma, Chi Zhang, Song-Chun Zhu  
**arXiv:** [2307.03762](https://arxiv.org/abs/2307.03762)  
**Date:** 2023-07-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2307.03762v1/

## Key points

- Authors challenge the notion that Large Language Models (LLMs) represent artificial general intelligence
- LLMs excel at language processing but fall short of true general intelligence demonstrated by humans
- Importance of transparent evaluation methods for true generalization and concerns about data leakage from internet-trained models
- Proposal of the concept of unity of knowing and acting as crucial for artificial general intelligence, lacking in LLMs
- Emphasis on interactive environments with rich affordances to facilitate concept learning and knowledge acquisition beyond vision and language processing
- Call for meta-verse support for various reasoning tasks to enhance agent learning through interaction with objects
- Advocacy for development of a cognitive architecture integrating knowing and acting mechanisms for knowledge abstraction, accumulation, and application
- Stress on the need to focus on key challenges such as evaluation transparency, interactive environments, and unifying knowing and acting mechanisms to advance towards artificial general intelligence

## Summary

In this perspective paper, the authors critically evaluate the current state of Large Language Models (LLMs) and challenge the notion that they represent artificial general intelligence. They highlight existing failure cases and argue that LLMs may excel at language processing but fall short of true general intelligence demonstrated by humans. The authors emphasize the importance of transparent evaluation methods to ensure true generalization and address concerns about data leakage from internet-trained models. Furthermore, the authors propose the concept of the unity of knowing and acting as a crucial factor for artificial general intelligence, which is currently lacking in LLMs. They suggest that interactive environments with rich affordances can facilitate concept learning and knowledge acquisition beyond just vision and language processing. The meta-verse should support various reasoning tasks to enhance agent learning through interaction with objects in the environment. The paper also calls for the development of a cognitive architecture that integrates knowing and acting to enable knowledge abstraction, accumulation, and application. While acknowledging the practical advancements brought by LLMs, the authors stress that they do not embody artificial general intelligence. They urge the research community to focus on advancing towards this ultimate goal by addressing key challenges such as evaluation transparency, interactive environments, and unifying knowing and acting mechanisms. Overall, this paper provides a comprehensive analysis of LLMs' limitations in achieving artificial general intelligence and offers insightful suggestions for future research directions in this field. It serves as a thought-provoking piece that encourages researchers to strive towards developing more advanced AI systems capable of true general intelligence beyond language processing capabilities.