The Wittgenstein’s Duality of Language, a reflection on generative AI

In this article, I’m further exploring the duality of language to its implication on the capabilities of Large Language Models (LLMs). Let’s start with the works of Ludwig Wittgenstein on the nature of language and the means of perceiving reality and rational thought. In his work, Wittgenstein himself represents a deep duality, mainly the “Tractatus Logico Philosophicus” [1] , “Tractatus” for short, as the work of the young Wittgenstein, and “Philosophical Investigations” [2] as the work of the older Wittgenstein.

Both works are of the most influence on philosophy of the 20th century and perhaps find a new application around generative AI and beyond in the 21st century. At a glance, it appears that the two works are in direct contradiction regarding the nature of language but a closer look reveals that they are linked together via a duality which sheds light on fundamental questions we have about us as the language-wielding species on Earth.

Both works give different essential meanings of language as a picture in the Tractatus and as a tool in the Philosophical Investigations.

?

Language as a Picture (Tractatus Logico-Philosophicus)

1. Representation of Reality: Wittgenstein initially proposed that language functions similarly to a picture, representing reality. Words and sentences mirror the structure and arrangement of the world.
2. Structural Similarity: There's an underlying assumption of structural similarity between language and the world it describes. The arrangement of words in a sentence corresponds to the arrangement of elements in the real world.
3. Logical Structure: This approach suggested a logical structure inherent in both the world and language, with a focus on elemental sentences depicting facts.
4. Metaphysical Lever: This viewpoint provided a metaphysical perspective where the structure of reality could be deduced from the structure of language.
5. Limitation on Meaning: According to this view, meaningful language was limited to fact-stating language, relegating other forms of expression to the realm of nonsense or the unsayable.

?

Language as a Tool (Philosophical Investigations)

1. Functional and Contextual Use: Wittgenstein later shifted to viewing language as a tool, emphasizing its use in various contexts. The meaning of words arises from how they are used in language.
2. Indefinite Extensibility and Family Resemblance: Language is seen as indefinitely extendable with no single essence binding all uses. Wittgenstein introduces the concept of ‘family resemblance’ to describe the overlapping and interrelated uses of words.
3. Social Practice and Rule-Governed Activity: Language is framed as a social practice governed by rules learned within a society. This perspective rejects the notion of a private language, emphasizing the communal aspect of language.
4. Forms of Life: Language is deeply connected with forms of life, indicating that its meaning is rooted in social contexts and practices.
5. Rejection of Theoretical Constraints: Wittgenstein moves away from seeking a general theory of language, focusing instead on the diverse, practical applications of language in various forms of life.

?

These two profoundly different usages of language have some philosophical implications such as:

• Shift The transition from language as a picture to language as a tool reflects a philosophical journey from a metaphysical to a more pragmatic and utilitarian understanding of language.
• Wittgenstein's later work challenges Cartesian dualism by situating language within social practices rather than as a medium to express internal states or represent external reality.
• The concept of language games further illustrates the diversity and contextual nature of language use, arguing against a one-size-fits-all approach to understanding language.
• Wittgenstein's ideas, particularly in his later period, have influenced fields beyond philosophy, such as literary criticism, anthropology, and political theory, underscoring the pervasive role of language in human thought and society.

In the next section, I’d like to link the capabilities and implications of Large Language Models (LLMs) such as GPT to Wittgenstein's philosophical perspectives on language as a tool versus language as a picture. This alignment may offer a better understanding of the current state of AI, particularly in terms of its capacity for understanding, reasoning, and creativity, as well as its limitations and potential.

?

Language as a Picture: LLMs' Limitations

1. Like Wittgenstein's picture theory, where language represents reality without necessarily engaging with its deeper complexities, LLMs mimic human language patterns without genuine understanding. They generate outputs based on statistical patterns, analogous to a picture that resembles the world but does not interact with it.
2. LLMs, like the picture theory, operate within the confines of their programming and data. They cannot transcend these structures and engage with the world in a meaningful, dynamic way. This limitation reflects Wittgenstein's notion of language as a static representation, unable to capture the fluidity and depth of human experience.
3. Just as language as a picture is limited to depicting surface realities, LLMs struggle with context, nuance, and the deeper implications of their outputs. They process information based on what is present in their training data. This is similar to a picture constrained by its frame.

?

Language as a Tool: LLMs' Potential

1. In alignment with Wittgenstein's later view of language as a tool, LLMs can be seen as tools that facilitate various functions, such as answering questions, generating text, and even creating art. Their utility is derived from how they are used in specific contexts, reflecting the pragmatic and utilitarian aspects of language.
2. Just as language as a tool adapts to different contexts and purposes, LLMs display a level of versatility in handling diverse tasks. They can be fine-tuned and applied in various domains, showcasing the adaptive nature of tools.
3. Advocates of LLMs' potential may argue that, like tools that evolve with use, LLMs could develop more advanced capabilities, including a form of understanding or problem-solving ability, as they are exposed to more data and diverse contexts.

?

Philosophical Implications and Considerations

1. Wittgenstein’s emphasis on the role of language in human life touches on the ethical considerations prompted by LLMs. Just as language shapes human reality and understanding, AI's influence on society necessitates responsible development and use, mindful of potential impacts and misuses.
2. Recognizing the limitations of LLMs, much like acknowledging the constraints of language as a representational tool, is crucial. It involves understanding that these potentially powerful models operate within the bounds of their programming and training, lacking the intrinsic understanding and consciousness of humans.
3. Wittgenstein’s focus on the social aspect of language as a tool aligns well with the idea of collaborative AI, where humans and AI systems work in tandem, leveraging the strengths of both. This perspective aligns with the view that AI should augment rather than replace human capabilities.

?

To put it all together, aligning Wittgenstein's philosophies of language with the capabilities and limitations of LLMs like GPT offers a different framework for understanding generative AI's current state and future potential using human-generated language data. It moves us away from admiring the almost unusual effectiveness of the state-of-the-art transformer architectures at the core of LLMs. It removes the reliance on conditional probabilities to a more “human” aspect of creating tools from our language arguably for the first time in history. It also underscores the importance of recognizing AI as both a tool with significant functional capabilities and a system with inherent limitations reminiscent of language as a mere representation.

In future articles, I’ll explore other dualities in philosophy, biology, mathematics, and physics to link them to the new summer of AI.

?

References:

1. https://www.wittgensteinproject.org/w/index.php/Tractatus_Logico-Philosophicus_(English)
2. https://archive.org/details/philosophicalinvestigations_201911/page/n123/mode/2up