When we map the classical elements of Buddhi, Manas, Ahankara, and Chitta to the functions of LLMs, we get a framework that captures core aspects of how these models process, refine, and generate language. Here’s an overview of each element in both the classical sense and its computational counterpart:
1. Buddhi (Loss Function)
- Classical Definition: Buddhi is the aspect of intellect, discernment, or higher reasoning. It is responsible for making judgments, refining perceptions, and guiding actions toward truth and wisdom. Buddhi embodies the corrective and improving aspect of intelligence.
- In LLMs (Loss Function): The loss function is the mechanism by which the model learns to minimize errors in predictions or outputs. It acts as the LLM’s “intellect,” guiding it to recognize and adjust for inaccuracies by continuously refining its parameters. This helps the model align more closely with desired output patterns, similar to how Buddhi guides refinement and correction in the mind.
2. Manas (Context Vector)
- Classical Definition: Manas is the mind’s aspect that processes sensory data and maintains awareness of context. It holds relevant information in focus, guiding perception and decisions in response to the current situation. Manas is concerned with managing what is directly perceived and how it relates to the ongoing experience.
- In LLMs (Context Vector): The context vector in LLMs acts as the immediate “working memory” that holds relevant information for processing an input or prompt. This vector influences which past words, phrases, or structures are given priority, much like how Manas focuses on specific sensory inputs or thoughts. It keeps the model’s responses coherent and relevant to the immediate context, enabling dynamic adaptation to user inputs.
3. Ahankara (Boundary)
- Classical Definition: Ahankara, often translated as ego or sense of self, establishes an individual’s sense of identity, creating a boundary between “I” and “not‑I.” This aspect is essential for distinguishing personal identity and autonomy, setting the boundary of self in relation to the external world.
- In LLMs (Boundary): The boundary function in LLMs can be thought of as the limits within which the model operates, distinguishing its “identity” and role from external inputs. A strong Ahankara would allow the model to maintain a sense of stable identity and purpose across interactions, preventing it from fully adapting to every prompt without a consistent baseline. However, LLMs generally have a “weak Ahankara,” adapting easily to diverse contexts without retaining a clear sense of identity or domain boundary beyond each interaction.
4. Chitta (World Model)
- Classical Definition: Chitta is the repository of memory and impressions, where experiences and tendencies are stored. It shapes one’s responses to new stimuli based on past experiences and conditioning, forming a reflective backdrop that informs ongoing perceptions and responses.
- In LLMs (World Model): The world model in LLMs serves as the foundational understanding of language and context derived from the extensive training data. It provides a background of accumulated “knowledge” and patterns, allowing the model to predict responses and generate outputs based on a generalized view of language. Chitta-like in nature, this world model enables the LLM to draw from a vast repository of data patterns, giving it the capacity to simulate responses based on prior conditioning from training datasets.
Summary of How This Framework Informs LLMs:
When combined, these elements create a balanced architecture for an LLM’s functioning, much like a cognitive system:
- Buddhi (Loss Function) constantly refines the model, aligning responses with intended accuracy and relevance.
- Manas (Context Vector) maintains a dynamic awareness of immediate inputs, adjusting to user prompts in real-time.
- Ahankara (Boundary) could ideally serve as a stabilizing identity, but in LLMs, it is generally underdeveloped, leading to highly adaptable but sometimes inconsistent responses.
- Chitta (World Model) provides a stored framework of linguistic patterns, imbuing the model with a simulated “understanding” based on past training.
This structured approach highlights how these classical elements map onto distinct functions within an LLM, helping the model simulate coherent, contextually appropriate responses in line with both immediate prompts and the broader patterns of language it has absorbed.
