# A Philosophical Introduction to Language Models -- Part I: Continuity With Classic Debates

**Authors:** Raphaël Millière, Cameron Buckner  
**arXiv:** [2401.03910](https://arxiv.org/abs/2401.03910)  
**Date:** 2024-01-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2401.03910v1/

## Key points

- LLMs can reproduce information and creatively combine patterns to produce original outputs
- The flexible generalization capabilities of LLMs in semantic spaces contribute to their efficiency and resilience compared to rule-based systems
- The paper introduces LLMs to philosophers, emphasizing the importance of engaging with fundamental questions about artificial intelligence
- It discusses the origins of LLMs in early AI research and the divide between symbolic and stochastic approaches in natural language processing
- Noam Chomsky's transformational-generative grammar and the development of rule-based syntactic parsers are highlighted as influences on LLMs
- Renewed debates surrounding artificial neural networks, deep learning advancements, and the success of LLMs are explored
- The significance of behavioral evidence from benchmarks and targeted experiments in shaping discussions about LLMs is emphasized

## Summary

The authors assert that LLMs possess the ability to not only reproduce information from their training data but also creatively combine patterns and produce original outputs. They draw on empiricist philosophy and scientific research to propose that the flexible generalization capabilities of these models in semantic spaces may account for their efficiency and resilience compared to rule-based systems. This paper serves as a comprehensive introduction to LLMs for philosophers, emphasizing the importance of engaging with fundamental questions about artificial intelligence. It traces the origins of LLMs back to early AI research and the divide between symbolic and stochastic approaches in natural language processing. The influence of Noam Chomsky's transformational-generative grammar is highlighted, along with the development of rule-based syntactic parsers for sentence decomposition. Additionally, the authors delve into renewed debates surrounding artificial neural networks in light of advancements in deep learning and the success of LLMs. They emphasize the significance of behavioral evidence from benchmarks and targeted experiments in shaping these discussions. Overall, this paper aims to provide a nuanced understanding of LLMs for philosophers across various disciplines by offering insights into their architecture, achievements, and philosophical implications in contemporary AI research.