# Talking About Large Language Models

**Authors:** Murray Shanahan  
**arXiv:** [2212.03551](https://arxiv.org/abs/2212.03551)  
**Date:** 2022-12-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2212.03551v5/

## Key points

- Large language models (LLMs) such as Bert and GPT-2 have transformed the field of AI
- LLMs use transformer architectures comprising hundreds of billions of parameters and trained on massive amounts of textual data
- The effectiveness of LLMs is surprising in three inter-related ways: their performance scales with the size of the training set, there are qualitative leaps in capability as the models scale, and a great many tasks that demand intelligence in humans can be reduced to next token prediction with a sufficiently performant model
- As LLMs become more adept at mimicking human language, we become more vulnerable to anthropomorphism - seeing these systems as more human-like than they really are
- To mitigate this trend, Murray Shanahan advocates for repeatedly stepping back to remind ourselves how LLMs actually work and how they form part of larger systems
- It's important not to overestimate the abilities of LLMs or see them as fully autonomous entities capable of independent thought. Instead, we should view them as tools designed for specific purposes within larger systems that require careful consideration and ethical oversight.
- The paper highlights the need for greater awareness around our use of language when discussing AI technologies like LLMs. By avoiding anthropomorphism and maintaining scientific precision in our discussions, we can foster a more nuanced understanding of these powerful tools while also ensuring responsible development and deployment practices going forward.

## Summary

In recent years, the rapid progress in artificial intelligence has led to an intersection of technology and philosophy, with large language models (LLMs) at the center of this intersection. LLMs such as Bert and GPT-2 have transformed the field of AI by using transformer architectures comprising hundreds of billions of parameters and trained on massive amounts of textual data. The effectiveness of these models is surprising in three inter-related ways: their performance scales with the size of the training set, there are qualitative leaps in capability as the models scale, and a great many tasks that demand intelligence in humans can be reduced to next token prediction with a sufficiently performant model. However, as LLMs become more adept at mimicking human language, we become more vulnerable to anthropomorphism - seeing these systems as more human-like than they really are. This trend is amplified by our tendency to use philosophically loaded terms such as "knows," "believes," and "thinks" when describing these systems. To mitigate this trend, Murray Shanahan advocates for repeatedly stepping back to remind ourselves how LLMs actually work and how they form part of larger systems. The hope is that increased scientific precision will encourage more philosophical nuance in discussions around artificial intelligence both within the field and in public discourse. While LLMs have shown remarkable capabilities, it's important not to overestimate their abilities or see them as fully autonomous entities capable of independent thought. Instead, we should view them as tools designed for specific purposes within larger systems that require careful consideration and ethical oversight. Overall, this paper highlights the need for greater awareness around our use of language when discussing AI technologies like LLMs. By avoiding anthropomorphism and maintaining scientific precision in our discussions, we can foster a more nuanced understanding of these powerful tools while also ensuring responsible development and deployment practices going forward.