Intelligence as Language Games: Wittgenstein and the Rise of Large Language Models

Revision: v0.2.0

Introduction

Intelligence is not limited to isolated areas of expertise but shows up across a wide range of human activities, from science and art to everyday conversation. Defining it has always been challenging. However, Ludwig Wittgenstein’s concept of “Language Games” offers a unique way to understand intelligence. Through his lens, we might see intelligence as the ability to engage in these various “games” of language. This idea also provides a philosophical foundation for understanding the recent success of large language models (LLMs) like GPT-4.

What Are Language Games?

Wittgenstein’s notion of “Language Games” suggests that words gain meaning based on their use within specific social contexts. He argued that language isn’t just a list of definitions but a flexible system where meaning is generated through its use in diverse “forms of life”—activities that have their own rules and purposes.

In this view, intelligence could be seen as the ability to participate in these varied language games. Intelligence, then, is not just abstract reasoning or memory but is woven into the subtleties of everyday life. Each “game” (whether a scientific debate, an artistic critique, a board game, a business negotiation, or a casual chat) represents a distinct form of intelligence.

Language Models and the Game of Text

When we apply this perspective to language models, we see why they excel in the “game” of textual interaction. Trained on vast datasets, LLMs learn patterns, rules, and structures embedded in human language, equipping them to adapt to different contexts and generate responses that feel meaningful to us.

While LLMs don’t “understand” in the way humans do, they participate in an approximation of human intelligence through text. They play the language game by recognizing and responding to prompts with learned context-specific patterns, allowing them to mimic the structure and tone of human responses across various topics.

Why Language Games Matter for AI

Thinking of intelligence as language games reveals a few intriguing ideas about artificial intelligence:

1. Diverse Forms of Intelligence: Intelligence isn’t a monolithic skill but varies with the type of activity, whether logical reasoning, social nuance, or creativity. This view allows us to recognize different forms of intelligence, beyond what is traditionally measured.

2. Context-Driven Interactions: Language models adjust their responses based on context, reflecting how human language adapts across situations. This adaptability is key to their success; they can recognize and respond in ways that feel appropriate across different settings, from professional writing to informal chats.

3. A Path to General Intelligence?: If intelligence is the ability to navigate language games, LLMs represent a step toward artificial general intelligence—at least in text. By mastering countless small, context-specific rules, they provide a glimpse of what it might mean for AI to exhibit “general” capabilities in certain realms.

Conclusion

Wittgenstein’s concept of Language Games provides a compelling lens to view both human and artificial intelligence. Intelligence, when seen as the ability to engage in language games, becomes a more nuanced and adaptable concept that can account for a wide range of human activities. LLMs exemplify this potential by succeeding not through human-like thought but by skillfully playing the language game. In doing so, they offer us a novel way to think about intelligence, one that aligns well with the varied and context-driven nature of human communication.