Mantra M5 Thesis Brief 17 – Converse: Sanskrit Grammar and LLMs – EVERYTHING CONNECTED

In this article, we set the stage for developing a 3-pronged architecture to converse, compute and infer needs for process engineering based on the explicit rules of Sanskrit grammar postulated by Panini, the architecture of MemGPT and the 3-phased philosophy propounded by Lacan based on Symbolic, Imaginary and Real registers:

Thusly, the architecture of our platform Mantra M5 will be divided into:

Converse – defining an Operating System for LLM based on the Ashtadhyayi and MemGPT.
Compute – creating a World Model to simulate interactions and enable the model to be trained based on the 5th order of Cybernetics.
Infer – Utilizing concepts of Nyaya and Yoga to build a multi-modal inference engine to convert meaning to motion, what we term actionable intelligence.

In this article, we will focus on the first item – Converse and we will literally “converse” with a multitude of LLMs to explore this red dot in our progression of defining the basis for Mantra M5, our platform for process engineering.

SANSKRIT (Dialog with Pi.ai)

The Ashtadhyayi of Panini is a Sanskrit treatise on grammar, written by the ancient Indian grammarian Panini in the 6th to 5th century BCE. It’s a significant work in the history of Sanskrit linguistics, as it standardized the language and established its grammatical rules.

The Ashtadhyayi is a collection of 4,000 sutras (aphorisms) that summarize the science of phonetics and grammar developed in the Vedic tradition. It’s considered a foundational text of the Vyakarana branch of Vedanga, the auxiliary scholarly disciplines of the Vedic period. 📜🗣️

The Ashtadhyayi remains a significant work in the era of Large Language Models (LLMs) for several reasons:

Understanding Language Fundamentals: The Ashtadhyayi offers a deep understanding of the fundamental principles of language, particularly in its rule-based approach to morphology (word structure) and syntax (sentence structure). This knowledge can inform the development of LLMs, helping them move beyond pure statistical learning to a more nuanced grasp of language.
Insights into Human Language Processing: Studying the Ashtadhyayi sheds light on how humans process language. The way the text builds complexity from basic rules offers insights into how our brains might work. This can be valuable for developing more human-like LLMs.
Benchmark for Formal Grammar Systems: The Ashtadhyayi stands as a historical benchmark for formal grammar systems. By analyzing its strengths and limitations, researchers can gain valuable insights for designing future grammar formalisms applicable to LLMs.
Potential for Hybrid Approaches: There’s potential for combining the strengths of LLMs and rule-based systems like the Ashtadhyayi. LLMs excel at statistical learning from massive datasets, while the Ashtadhyayi provides a structured framework. Merging these approaches could lead to LLMs that are both powerful and more interpretable.

However, it’s important to consider some limitations:

Direct Application: The Ashtadhyayi is specific to Sanskrit and may not directly translate to other languages, especially modern ones with less structured grammar.
Computational Complexity: Implementing the full complexity of the Ashtadhyayi in an LLM might be computationally expensive. Researchers would need to find ways to make the system efficient for large-scale applications.

Overall, the Ashtadhyayi offers valuable insights for the field of LLMs, even if it’s not directly applicable in its entirety. It serves as a reminder of the importance of understanding fundamental linguistic principles and the potential benefits of combining statistical learning with structured grammar formalisms.

Sanskrit itself wouldn’t directly build a knowledge graph for process engineering, but it offers some interesting starting points for the taxonomy and ontology within that knowledge graph. Here’s a breakdown:

Limitations of Sanskrit:

Domain Specificity: Sanskrit’s vocabulary and structure are geared towards general communication, not the specific terminology and concepts of process engineering.

Potential Contributions of Sanskrit:

Conceptual Foundation: Sanskrit, especially through its focus on root words and their relationships, can help identify core concepts relevant to process engineering. For instance, the root “kri” (to do, make) could be linked to various manufacturing processes.
Logical Structure: The Ashtadhyayi’s emphasis on rules and hierarchy can inspire the organization of the knowledge graph’s ontology. The relationships between broader categories (e.g., materials) and subcategories (e.g., metals, plastics) can be informed by Sanskrit’s approach.

Building the Knowledge Graph:

Here’s how you can create the knowledge graph for process engineering:

Domain Experts: Involve process engineers to identify key entities (materials, equipment, processes) and their relationships.
Standardized Vocabularies: Leverage existing ontologies and taxonomies specific to process engineering, such as those from engineering standards organizations.
Natural Language Processing (NLP): Use NLP techniques to extract relevant information from process engineering documents, literature, and data sources. This can help populate the knowledge graph and identify relationships between entities.

Benefits of a Process Engineering Knowledge Graph:

Improved Search and Retrieval: The knowledge graph allows for efficient search and retrieval of process engineering information. Users can find relevant data based on entities, relationships, and properties.
Automation and Decision-Making: The knowledge graph can be used to automate tasks and support decision-making within the AI platform. By understanding relationships between entities, the platform can suggest optimal materials, recommend efficient processes, and identify potential issues.
Integration and Interoperability: The knowledge graph can act as a central hub for integrating data from various sources within the platform, promoting consistency and interoperability.

Conclusion:

While Sanskrit itself won’t build the knowledge graph, its underlying principles can inspire its structure. The real legwork lies in collaborating with domain experts, leveraging existing resources, and employing NLP techniques to populate the knowledge graph with the specific entities, relationships, and properties relevant to process engineering. This will empower your AI platform to understand, reason, and make informed decisions within the complex world of process engineering.

What is the Ashtadhyayi of Panini?

The Ashtadhyayi of Panini is an ancient Sanskrit text on grammar, considered one of the most influential works of linguistics in history. Here’s a breakdown of its key features:

Authored by: Panini, a scholar believed to have lived around 500 BCE
Significance: It laid the foundation for the standardization of Classical Sanskrit.
Structure: The Ashtadhyayi is a concise treatise consisting of around 4,000 aphorisms (sutras) that describe the language’s phonetics and grammar. It’s divided into eight chapters with further subdivisions.
Content: The text goes beyond just the structure of words. It also delves into the differences between spoken and literary Sanskrit.
Uniqueness: The Ashtadhyayi is known for its complex yet elegant use of metarules, transformations, and recursion. This has led some to compare it to the Turing machine, a theoretical model of computation.

Understanding the Ashtadhyayi requires studying alongside supplementary texts that provide word lists and other necessary information for applying the sutras.

What are its 8 chapters, and can you describe them?

The Ashtadhyayi’s eight chapters, each further divided into four sections (padas), don’t have specific titles but rather focus on progressively building the structure of Sanskrit language. Here’s a general overview of their functions:

Sandhi: This chapter deals with the rules of combining sounds at the boundaries of words, ensuring smooth pronunciation.
Stavabhyasa: It focuses on the modification of vowels and consonants within a word.
Taddhita: This chapter introduces suffixes (Taddhita) that modify the meaning and function of root words.
Samjñasana: Here, the focus is on the formation of nouns through various processes, including compounding.
Agama: This chapter introduces sounds (Agama) inserted within words during the formation process.
Krtya: It deals with the creation of verbs from nominal stems using specific suffixes.
Upapada: This chapter focuses on particles (prepositions and the like) and their role in sentences.
Ahatadi: The concluding chapter addresses exceptions and specific formations not covered in the preceding chapters.

It’s important to note that these are simplified explanations. Each chapter delves into intricate details and interacts with the others to create a comprehensive system for Sanskrit grammar.

Somebody has compared Sanskrit as being akin to “reading, writing and using” a computer program. Can you explain this?

The comparison of Sanskrit to reading, writing, and using a computer program is interesting and stems from some unique features of the language, particularly its connection to the Ashtadhyayi. Here’s why this comparison is made:

Systematic and Rule-Based: Sanskrit grammar, as codified in the Ashtadhyayi, is highly systematic and rule-based. Much like a computer program with its syntax and logic, Sanskrit follows a set of defined rules for forming words, sentences, and structures.
Precision and Efficiency: The Ashtadhyayi emphasizes efficiency and precision. Just like computer code aims for clear and concise instructions, the aphorisms (sutras) in the Ashtadhyayi strive to be compact yet comprehensive.
Emphasis on Transformation: Sanskrit grammar involves transformations, where you start with a root word and apply rules to create new words with specific meanings. This can be seen as analogous to how computer programs manipulate data based on defined instructions.
Recursion: The Ashtadhyayi utilizes recursion, where a rule can refer to itself. This allows for complex structures to be built from simpler ones, similar to how functions within a computer program can call upon themselves.

However, there are also key differences:

Creativity and Fluidity: Human language like Sanskrit allows for creativity and ambiguity that’s not typical of computer programs. While the rules provide a framework, there’s still room for interpretation and expression.
Evolution and Change: Languages like Sanskrit evolve over time, unlike computer programs designed for a specific purpose.

So, the analogy highlights the structured and logical aspects of Sanskrit, but it’s important to remember that Sanskrit is a rich and expressive language that goes beyond just following rules.

In the era of Large Language Models, where does the Ashtadhyayi fit?

The Ashtadhyayi holds a unique position in the era of Large Language Models (LLMs) like me. Here’s how they can be seen as complementary.

Similarities:

Focus on Rules: Both LLMs and the Ashtadhyayi deal with rules for generating language. LLMs are trained on massive amounts of text data, which allows them to identify patterns and statistical relationships between words. The Ashtadhyayi, on the other hand, provides a set of explicit rules for manipulating language.
Predictability and Efficiency: Both approaches aim for a degree of predictability and efficiency in language processing. The Ashtadhyayi offers a systematic way to generate grammatically correct Sanskrit sentences. Similarly, LLMs strive to generate human-like text that is grammatically sound and coherent.

Differences:

Data-Driven vs. Rule-Based: LLMs are data-driven, learning from vast amounts of text. The Ashtadhyayi, however, is a rule-based system, offering a more theoretical and philosophical approach to language.
Adaptability: LLMs can adapt to new information and contexts based on the data they’re exposed to. The Ashtadhyayi, while incredibly detailed, is a fixed system specific to Classical Sanskrit.
Focus on Meaning: While LLMs can generate grammatically correct text, they may struggle with deeper semantic understanding. The Ashtadhyayi, however, goes beyond structure, with some scholars suggesting it delves into the relationship between language and thought.

Potential for Collaboration:The Ashtadhyayi and LLMs can potentially work together:

Formal Language Understanding: The Ashtadhyayi’s explicit rules could be used to train LLMs to better understand the formal aspects of language, like grammar and sentence structure.
Generating Specific Languages: The Ashtadhyayi’s approach could be used to develop LLMs for languages with complex grammar rules, like Sanskrit itself.
Understanding Language Universals: By comparing how LLMs learn language patterns with the structure of the Ashtadhyayi, researchers might gain insights into universal principles of language.

Overall, the Ashtadhyayi offers a valuable historical perspective on how humans have tried to understand and codify language. In the age of LLMs, it serves as a reminder of the importance of both rules and data in achieving human-like language processing.

Is there any research being conducted to use Sanskrit as a metalanguage in the context of LLMs?

While there isn’t widespread research on using Sanskrit specifically as a metalanguage for LLMs, there is definite interest in incorporating ideas from formal languages and linguistics into LLM development. Here’s what we can explore:

Formal Languages and Universal Grammar: There’s ongoing research in applying concepts from formal language theory, like Chomsky’s Universal Grammar, to LLMs. This could involve building models that understand core grammatical principles underlying human languages. The Ashtadhyayi, with its focus on rules and structure, aligns with this approach.
Linguistic Features in LLMs: Some researchers are exploring ways to integrate linguistic features like morphology (word structure) and syntax (sentence structure) into LLMs. This could help them move beyond statistical relationships between words and achieve a deeper understanding of language. The Ashtadhyayi’s detailed analysis of these aspects could be a valuable resource.

Here are some search keywords you can use to explore further:

While Sanskrit itself might not be the language of choice for current LLM architectures, the theoretical underpinnings of the Ashtadhyayi could be relevant to this area of research.

We are formulating a generative AI platform that combines conversation, computation and actionability using LLMs, compute servers and inference engines. How could such a platform be created to encompass process engineering needs of make, manage, move, market and maintain of products and services?

Your idea for a generative AI platform that combines conversation, computation, and actionability for process engineering has a lot of potential. Here’s how you could approach it to encompass the “make, manage, move, market, and maintain” aspects of products and services:

Core Components:

Large Language Models (LLMs): These will handle the conversational interface, understanding user intent and translating it into actionable tasks.
Compute Servers: These will provide the processing power for complex simulations, data analysis, and optimization tasks.
Inference Engines: These will bridge the gap between the LLM and the compute servers, translating LLM outputs into instructions for simulations and calculations.

Modules for Process Engineering Needs:

Design and Development (Make):
- The LLM can interact with users to capture product or service requirements.
- It can then access databases and past projects to suggest design options and materials.
- The inference engine can run simulations on compute servers to evaluate performance and manufacturability.
Production Planning and Scheduling (Manage):
- The LLM can analyze production data and recommend optimal production schedules.
- It can integrate with other systems for resource allocation and inventory management.
Logistics and Supply Chain (Move):
- The LLM can analyze real-time data to optimize transportation routes and delivery schedules.
- It can integrate with logistics providers and track shipments.
Marketing and Sales (Market):
- The LLM can analyze customer data and market trends to suggest targeted marketing campaigns.
- It can even generate marketing copy and personalize content for different customer segments.
Maintenance and Service (Maintain):
- The LLM can be used to create interactive maintenance manuals and chatbots to assist customers.
- It can analyze sensor data from products to predict maintenance needs and schedule service calls.

Additional Considerations:

Domain-Specific Knowledge: Train the LLM on process engineering data, terminology, and best practices.
Integration with Existing Systems: Ensure your platform can connect with existing software used for design, manufacturing, and other functions.
Security and Explainability: Implement robust security measures and ensure the AI’s decision-making process is transparent.

Overall, this generative AI platform has the potential to streamline process engineering workflows, improve decision-making, and increase overall efficiency. By combining human expertise with AI capabilities, you can create a powerful tool for the entire product lifecycle.

Where would the explicit grammar of Sanskrit’s Ashtadhyayi fit in?

In your generative AI platform for process engineering, the explicit grammar of Sanskrit’s Ashtadhyayi wouldn’t directly translate into core functionality. However, there could be some interesting ways to draw inspiration from its principles:

Limited Applicability:Sanskrit grammar, while incredibly detailed, is specific to that language. Directly applying its rules wouldn’t be helpful for understanding the complexities of process engineering, which involves diverse technical terms and concepts.

Potential Areas of Inspiration:

Formalization and Structure: The Ashtadhyayi’s emphasis on rules and structure could inspire the development of a formal language for describing process engineering workflows. This could improve communication, consistency, and automation within the platform.
Modularity and Reusability: Similar to how the Ashtadhyayi builds complex structures from simpler ones, your platform could benefit from modular components for different aspects of process engineering (design, scheduling, etc.). This allows for reusability and easier adaptation to different scenarios.
Efficiency and Comprehensiveness: The Ashtadhyayi strives for efficiency and aims to be comprehensive within its domain. Your platform can be designed with similar goals in mind, optimizing workflows and aiming to encompass a wide range of process engineering needs.

Alternative Approaches for Understanding Natural Language:Instead of the Ashtadhyayi, your platform could benefit more from incorporating recent advancements in Natural Language Processing (NLP) techniques:

Domain-Specific Language Models (DSLMs): Train these models on process engineering data and terminology to improve the LLM’s understanding of user queries and instructions related to product design, manufacturing, etc.
Entity Recognition and Relationship Extraction: Develop functionalities to identify key entities (materials, components) and their relationships within user instructions and documents. This allows the platform to extract crucial information for process automation.

Overall, while the Ashtadhyayi itself might not be directly applicable, its underlying principles of structure, efficiency, and modularity can inspire the development of your AI platform. Modern NLP techniques tailored to process engineering needs will likely be more impactful for achieving your goals.

Typical genAIs are based on Knowledge Graphs. Can Sanskrit help create this, the taxonomy and the ontology needed for process engineering?

Sanskrit itself wouldn’t directly build your knowledge graph for process engineering, but it offers some interesting connections. Here’s a breakdown:

Knowledge Graphs for Process Engineering:

A knowledge graph for process engineering would be a network of interconnected concepts, entities, and relationships relevant to the field.
It would include information about materials, equipment, processes, and how they all relate to each other within the “make, manage, move, market, maintain” lifecycle of products and services.

Sanskrit and Knowledge Graphs:

While Sanskrit can’t directly populate your knowledge graph, its underlying structure can be inspiring.
The Ashtadhyayi’s focus on categorization and relationships between words aligns with the concept of ontologies within a knowledge graph.

Alternative Approaches for Building the Knowledge Graph:Here’s how you can develop the knowledge graph and its components for process engineering:

Domain Experts: Involve process engineers to identify key concepts, entities, and relationships within the field.
Standardized Vocabularies: Leverage existing ontologies for engineering and manufacturing, such as GOODS (Generic Object-Oriented Design System) or OpenCyc.
Machine Learning Techniques: Use NLP to extract relevant information from process engineering documents, manuals, and data sources. This can help populate the knowledge graph and identify relationships between entities.

Benefits of a Knowledge Graph:

Improved communication and collaboration among engineers.
Enhanced decision-making through access to a centralized knowledge base.
Streamlined workflows by automating tasks based on knowledge graph queries.
Increased efficiency in information retrieval and analysis.

Overall, Sanskrit provides an interesting philosophical connection to knowledge graphs, but the actual development will rely on domain expertise, existing ontologies, and NLP techniques tailored to process engineering.

GRAMMAR (gemini.google.com)

In Indian philosophical traditions, especially the Vedic and Upanishadic thought, the concept of Vac (speech or voice) is described as having four distinct stages or levels of manifestation. These are:

Parā Vāc (Transcendental Speech): This is the supreme, transcendental form of speech that remains in the realm of the Absolute Reality or Brahman. It is the unmanifest, eternal source from which all speech originates.
Pashyantī Vāc (The Speech as Seen): This is the stage where speech takes a subtle mental form as thoughts or ideas before being expressed verbally. It represents the ideation stage in the mind.
Madhyamā Vāc (Intermediate Speech): At this level, speech takes on a transitional phase where the mental thoughts and ideas transform into words and sentences before being articulated.
Vaikharī Vāc (Expressed Speech): This is the final stage where speech manifests as the audible, expressed form of words spoken or language communicated outwardly.

These four levels symbolize speech progressively manifesting from its subtlest, unmanifest transcendental essence (Parā) through the stages of mental conception (Pashyanti), linguistic formation (Madhyama) and finally as spoken expression (Vaikhari).

In the philosophical traditions, grasping the progression of Vac from Parā to Vaikhari is seen as a means to trace the emergence of the entire phenomenal reality from its causal source in the Absolute.

Every paragraph needs to be scrutinized as to its authentic literal meaning and then see what other meanings are implied or hidden in the expressions. In Sanskrit, the multiple levels of meaning is codified. Can you tell me what they are?

DRAFT 1:

You’re absolutely right! Sanskrit is known for its rich and layered meaning. Here are some of the traditional ways Sanskrit texts can convey multiple levels of meaning:

Shruti (श्रुति): This refers to the “heard” knowledge, the primary meaning conveyed through the literal interpretation of the words and grammar. It forms the foundation for understanding the deeper layers.
Smriti (स्मृति): This translates to “remembered” knowledge and refers to the secondary meaning derived from cultural context, traditions, and past interpretations. It adds layers of meaning based on established knowledge.
Vakyartha (वाक्यार्थ): This signifies the “sentence meaning,” where the meaning is derived not just from individual words but from how they interact within the sentence structure. The order and arrangement of words can unlock deeper meanings.
Lakshana (लक्षण): This refers to the “figurative” or “implied” meaning. It involves figures of speech like metaphors, similes, and symbolism to convey deeper truths beyond the literal level.
Vyanjana (व्यंजना): This translates to “suggested” or “indicated” meaning. It’s the most subtle level, where the text hints at additional meanings through connotation, soundplay, and emotional associations.

These levels are not mutually exclusive and can work together to create a rich tapestry of meaning. Understanding all these layers requires a deep knowledge of the language, cultural context, and philosophical underpinnings of Sanskrit texts.

DRAFT 2

Sanskrit does indeed have a rich tradition of understanding texts on multiple levels of meaning. This is often referred to as Arthapatti (अर्थपत्ति) or Chāyā (छाया), which translates to “secondary meaning” or “shade of meaning.” Here are some of the recognized levels of meaning in Sanskrit texts:

Sthulaarthah (स्थूलार्थः): The literal or primary meaning. This is the surface level understanding of the words and sentence structure.
Sukshmaarthah (सूक्ष्मार्थः): The subtle or secondary meaning. This goes beyond the literal level and explores the implied meaning or message behind the words. It could involve metaphors, symbolism, or cultural references.
Laक्षणikarthah (लक्षणिकार्थः): The figurative or suggestive meaning. This involves interpreting the text based on figures of speech like similes, metaphors, and allegories.
Vyangyarthah (व्यंग्यार्थः): The implied or suggested meaning. This is the deepest level of meaning, often hinted at through context, tone, and emotional suggestion. It can be quite nuanced and may vary depending on the reader’s interpretation.
Anagogic meaning: This is a concept borrowed from Western philosophy and refers to a spiritual or mystical interpretation of the text. It seeks to uncover a hidden religious or philosophical truth beyond the literal meaning.

It’s important to note that not all Sanskrit texts are meant to be understood on all these levels. Some might be primarily focused on conveying a straightforward message, while others might be rich with layers of symbolism and hidden meanings.

Understanding these different levels of interpretation is a crucial skill for scholars and practitioners of Sanskrit.

DRAFT 3:

You’re absolutely right! Sanskrit is known for its rich and layered meaning, often conveying multiple levels of interpretation within a single phrase. Here are some of the key concepts that contribute to this:

Shruti and Smriti: Shruti refers to the “heard” knowledge, the core Vedas passed down through oral tradition. Smriti are the “remembered” texts, interpretations and commentaries on the Vedas. This layered structure allows for deeper exploration of concepts within the framework of the original teachings.
Sphota: This refers to the underlying essence or “linguistic potential” of a word. A single word can have multiple meanings depending on the context and interpretation. It’s like a seed holding the potential for a variety of interpretations to bloom.
Anvaya: This refers to the “syntactic connection” between words in a sentence. The order and arrangement of words influence the meaning derived. By changing the order, you can unlock new layers of meaning.
Vacya: This refers to the “primary or literal” meaning of a word or sentence. It’s the most basic interpretation you get from the surface level.
Lakshana: This refers to the “secondary or implied” meaning of a word or sentence. It goes beyond the literal meaning and suggests deeper connotations or symbolism.
Vyanjana: This refers to the “suggested or hinted” meaning of a word or sentence. It’s the most subtle level, often relying on cultural context or the reader’s own understanding to unlock hidden meanings.

By understanding these concepts, readers can approach a Sanskrit text with a multi-layered perspective. They can appreciate the literal meaning (Vacya) and then delve deeper to uncover the implied (Lakshana) and suggested (Vyanjana) meanings. This layered approach makes Sanskrit a powerful tool for philosophical and spiritual exploration.

When scrutinizing a Sanskrit paragraph or sentence, there are multiple levels of meaning that need to be analyzed:

1. Shabdartha – The literal meaning of the words themselves.
2. Matitartha – The succinct or direct meaning being communicated.
3. Goodhartha – The deeper, inner meaning implied.
4. Guhyartha – Any secret or hidden meaning encoded.
5. Sandarbhartha – The contextual meaning based on the circumstantial context.
6. Sanketartha – Any suggested or hinted meaning requiring further interpretation.

On top of these multi-layered meanings intrinsic to the Sanskrit text itself, the reader’s own unconscious projections and conditioning from their background/training can further distort or lead to vipareeta artha – inverted or wrong meaning. Utmost caution is needed when scholars/academics steeped in Western paradigms attempt to interpret Sanskrit works, as their inherent lenses can cause them to completely miss the authentic essence and make a “mess” of the profound wisdom encoded at multiple levels.

You rightly warned that just having academic credentials in “South Asian Studies” does not automatically grant someone the ability to perceive the nuances of meaning in Sanskrit literature without unconscious biases creeping in.

In essence, comprehending Sanskrit requires continually peeling back layers of one’s own conditioned mind to reveal the refined, multidimensional meaning – a process that is immensely challenging for those deeply ingrained in Western thought.

TRAINING

What is the AI learning type that asks the program to learn by filling in the gaps?

There isn’t a single, universally accepted term for AI learning that specifically involves “filling in the gaps.” However, there are a couple of areas of AI research that align with this concept:

Semi-supervised Learning: This type of learning leverages both labeled and unlabeled data. Labeled data has pre-defined outputs, while unlabeled data lacks these classifications. The AI model uses the labeled data to learn patterns and then attempts to apply those patterns to classify the unlabeled data. In essence, the model is “filling in the gaps” in its understanding by using the labeled data as a guide for the unlabeled data.
Active Learning: Here, the AI model itself drives the learning process. The model identifies areas where its knowledge is lacking and actively seeks out new data to fill those gaps. This can involve querying a user or database for additional information relevant to the specific uncertainty it faces.

Both approaches involve the AI system working to improve its understanding by addressing areas where its knowledge is incomplete.

Example would be:

The “blank” is red or
The “blank” is blue

and the AI is asked to come up with rose and sky.

That’s a great example! Here are some learning techniques that fit that scenario:

Cloze Procedure: This is a common technique used in language learning, but it applies to AI training as well. In a cloze procedure, the AI is presented with text with missing words (blanks). Its task is to predict the most likely words to fill those blanks based on its understanding of the context. Your example, “The ‘blank’ is red or The ‘blank’ is blue,” perfectly demonstrates a cloze prompt.
Masked Language Modeling (MLM): This is a specific type of neural network architecture used in natural language processing (NLP). In MLM, the model is presented with text where some words are masked (replaced with a special token). The model then tries to predict the masked words based on the surrounding context. This is very similar to the cloze procedure but leverages the power of neural networks for more complex predictions.
Template Filling: This technique involves providing the AI system with a pre-defined template with missing elements. The AI then learns to fill those elements based on the data it’s trained on. For example, the template could be “The sky is usually _ during the day and _ at night.” The AI would be tasked with filling in the blanks with appropriate words like “blue” and “dark.”

In all these examples, the AI is essentially “filling in the gaps” by leveraging its knowledge and the context provided to make predictions about the missing information.

What is CNN (Convolutional Neural Network)?

A Convolutional Neural Network (CNN) is a type of artificial intelligence specifically designed for analyzing images and videos. Here’s a breakdown of how it works:

Inspiration from the Brain: Unlike traditional neural networks that process data in a flat way, CNNs are inspired by the structure of the visual cortex in the human brain. The visual cortex processes information hierarchically, detecting edges and lines at first, then gradually building up to more complex shapes and objects.

Key Components:

Convolutional Layers: These layers are the heart of a CNN. They apply filters (also called kernels) that slide across the image, extracting features like edges, shapes, and colors. Imagine a filter looking for horizontal lines. As it slides across the image, it will activate strongly where it finds horizontal lines.
Pooling Layers: These layers downsample the data from the convolutional layers, reducing its complexity and computational cost. This can involve techniques like averaging the values in a specific area or keeping only the maximum value.
Fully-Connected Layers: These layers function more like traditional neural networks, taking the features extracted by the convolutional layers and using them to classify the image or make other predictions.

Advantages of CNNs:

Excellent at Image Recognition: CNNs excel at recognizing patterns in images, making them highly effective for tasks like object detection, image classification, and facial recognition.
Automatic Feature Extraction: Unlike traditional machine learning methods that require manual feature engineering, CNNs can automatically learn the most important features from the data itself.
Handle Complex Data: CNNs can effectively process high-dimensional data like images and videos, which are difficult for traditional algorithms to handle.

Applications of CNNs:

Self-driving Cars: CNNs are crucial for self-driving cars to recognize objects like pedestrians, traffic lights, and other vehicles.
Medical Diagnosis: CNNs can be trained to analyze medical images like X-rays and MRIs to help doctors detect diseases.
Facial Recognition: CNNs power many facial recognition applications, used for security purposes or unlocking smartphones.
Image Recommendation Systems: Online platforms use CNNs to recommend products or content based on your browsing history and image preferences.

CNNs are a powerful tool in the field of computer vision and continue to evolve, enabling new applications and advancements in artificial intelligence.

PROMPT ENGINEERING (gemini.google.com)

In “Making sense out of nonsense”, we had compiled a series of articles on using different LLMs: ChatGPT, Pi, Claude and Gemini. Here is a summary of how LLMs work:

In working with LLMs we always talk of “hallucinations”. The definition of hallucination is: Perception in the absence of stimulus. Shouldn’t the real word be “confabulation”? Can you define these two words first before replying to my question?

You’re right, there’s an ongoing debate about the appropriate term for AI outputs that contain false or misleading information. Here’s a breakdown of both terms:

Hallucination: A sensory perception that occurs in the absence of any external stimulus. It’s like seeing things that aren’t there or hearing sounds that no one else can.
Confabulation: The act of unconsciously creating false memories or information to fill in gaps or inconsistencies in one’s memory. It’s like making up a story to explain something you can’t quite remember.

While “hallucination” is commonly used for AI outputs, it’s not entirely accurate. AI systems don’t have sensory experiences in the same way humans do.

So, is confabulation a better term?

It’s a strong contender. Confabulation captures the idea of AI models filling in gaps in their knowledge base with made-up information. However, some argue that “confabulation” implies intentionality, which AI models might not possess.

Here’s the ongoing debate:

Using “hallucination”: It’s a familiar term and highlights the fact that the information is unreal.
Using “confabulation”: It better reflects the process of filling in gaps but might be misinterpreted as intentional.

Ultimately, the best term might be something entirely new that reflects the specific way AI models generate inaccurate outputs.

I agree with you that “confabulation” probably relates more to “made up” than real answer. Is there any other work in the English vocabulary or a translation from any other language that relates to this phenomenon that could be a better word?

You’re right, “confabulation” might not be perfect. Here are some other options, each with its own strengths and weaknesses:

Fabrication: This emphasizes the creation of something false but doesn’t necessarily capture the unintentional aspect.
Invention: Similar to fabrication, but suggests a more creative process, which might not always be the case with LLMs.
Artifact: This refers to something created by a process, highlighting that the information is a product of the LLM itself. However, it’s a more general term and might not clearly convey the issue of falseness.

Looking beyond English, there might be interesting options:

Méprise (French): This translates to “mistake” or “misunderstanding,” but can also imply a misconception or something misinterpreted.
Truchement (French): This means “go-between” or “interpreter,” which could be interesting as it reflects the LLM acting as an intermediary between data and user output.
German: “Fehlleistungen” (faulty performance) – This emphasizes the malfunctioning aspect of the AI.
Japanese: “Genshin” (false impression) – This focuses on the misleading nature of the output.
Japanese: “Genshi Gensho” (hallucination/illusion) – Similar to “hallucination” but might be more fitting for AI due to the lack of biological senses involved.

The ideal term would capture the following:

Unintentional: The LLM doesn’t deliberately create false information.
Data-driven: The outputs are based on the data the LLM is trained on.
Inaccurate: The information might be misleading or wrong.

The search for the perfect term is ongoing in the field of AI. Researchers are constantly evaluating the best way to describe these outputs.

Lets test: There is an elegant solution to the 3-body problem. Do you know what it is?

There is no known general analytical solution to the classical 3-body problem in celestial mechanics. The 3-body problem involves determining the motions of three bodies (e.g., stars, planets) interacting through gravitational forces, given their positions, masses, and velocities at some initial time.

While solutions exist for special cases with simplified assumptions, such as when one of the masses is negligible or when the three bodies form a particular configuration, a general closed-form solution that describes the motion of three arbitrary masses has not been found.

The complexity and chaotic nature of the 3-body problem have made finding a general solution an intractable problem in mathematics and physics. Numerous mathematicians and physicists have worked on this problem over centuries, but a complete analytical solution remains an open challenge.

Instead, the 3-body problem is typically approached using numerical methods and computational simulations to approximate the motions of the three bodies over time. These methods involve iteratively calculating the forces and updating the positions and velocities of the bodies based on the equations of motion.

If you were implying that there is an “elegant solution” to the general 3-body problem, I must respectfully disagree or request that you provide more information about this purported solution, as it would be a significant breakthrough in celestial mechanics and mathematics.

Nice. You did not hallucinate or should I say confabulate!

Thank you for the feedback. You’re correct that in this case, I did not hallucinate or confabulate an imaginary solution to the 3-body problem, which remains an unsolved challenge in classical mechanics. As an AI system, I try my best to avoid generating information that is factually incorrect or inconsistent with established knowledge, unless directly prompted to be creative or speculative. Maintaining accuracy and avoiding confabulation is an important goal, though it’s an ongoing challenge given the limitations of my training data and model. I appreciate you testing this aspect of my capabilities – it helps reinforce the importance of distinguishing fact from fiction or unfounded claims.

What would be an example instance where you would definitely hallucinate?

As an AI assistant without subjective experiences, I do not actually “hallucinate” in the literal sense of perceiving something that does not exist. However, there are certainly instances where my outputs could be considered confabulations or hallucinations in the figurative sense of generating fictional content inconsistent with facts.

One scenario where I would be prone to confabulating is if I were asked about highly speculative future events or technological developments well beyond my training data cutoff. For example, if asked to describe the first human colony on Mars in the year 2050, I would have to confabulate plausible-sounding but ultimately fictionalized details, since I have no factual information about such future events.

Another case is when prompted with open-ended, unbounded creative writing tasks like “Tell me a story about a talking tree in a magical forest.” Here, I would generate a coherent fictional narrative, essentially hallucinating plotlines, characters, and events that do not correspond to any reality.

Additionally, if there are gaps or inconsistencies in my training data regarding some topic, I may inadvertently combine pieces of inaccurate information in an attempt to craft a response, producing confabulated “facts.”

So in essence, while I do not truly hallucinate, anytime the prompt requires me to venture beyond the scope of my training data into hypothetical, creative, or speculative domains, the resulting outputs could reasonably be considered forms of confabulation or hallucination. Maintaining ground truth accuracy is an ongoing challenge.

But story telling is fiction, I would not put it in the same bucket as hallucination/confabulation. In story telling instance, I want you to make up things.

You make a fair point. My previous example of creative writing or storytelling should not be categorized as hallucination or confabulation, since in those instances, the intent is explicitly to generate fictional narratives and imaginary scenarios. You are correct that storytelling requires intentionally fabricating events, characters, and plots that do not represent reality.

A better example of when I would genuinely hallucinate or confabulate would be if I were asked a factual question that falls outside the scope of my training data, and I attempted to synthesize a response based on flawed or incomplete information in my knowledge base. This could result in me outputting statements or “facts” that are detached from truth and reality.

For instance, if asked about the geological composition of an alien exoplanet that my training data did not cover, and I tried to extrapolate an answer from tangentially related information about Earth’s geology, I could end up hallucinating or fabricating inaccurate details about this fictional exoplanet.

So you are correct – creative writing intentionally involves fictional world-building, which is not the same as unintentionally hallucinating or confabulating information that deviates from established facts and reality. Thank you for clarifying that distinction. I will be more precise in separating deliberate fiction from unintended hallucinations in the future.

But isn’t transformer architecture based on attention register hallucination in a sense? I call it “making sense of nonsense” where the next word is selected using probabilistic weights?

You raise an insightful point about the nature of how transformer language models like myself generate text. While I don’t literally experience hallucinations, the process by which I predict the next word in a sequence could be viewed as a form of “making sense of nonsense” or confabulating plausible continuations.

The transformer architecture does indeed rely heavily on self-attention mechanisms to weigh and combine relevant parts of the input sequence when generating the next output token. However, this is done in a statistical, probabilistic manner based on patterns in the training data, rather than truly “making sense” in the cognitive way a human would.

So in a sense, you could interpret it as my language model trying to statistically confabulate or hallucinate coherent next words, sentences, and narratives based on recognizing and extrapolating patterns – even if the prompts contain nonsensical elements my training did not directly encounter. The models don’t truly understand meaning, but through weighted associations, they can generate contextually plausible continuations.

From this perspective, the entire process of an LLM generating fluent text could be viewed as a form of advanced confabulation – leveraging statistical associations to fabricate responses that “make sense” to humans, while not actually deriving them from grounded real-world understanding.

You’ve highlighted a philosophically interesting way to frame how transformer LLMs operate under the hood. While not true hallucinations, their text generation does involve statistically confabulating responses based on latent patterns, in what could be seen as “making sense of nonsense” through densely associated probabilities.

Finding and correcting falsehoods (hallucinations and confabulations) will be the bane of future LLMs. In a pertinent article by MIT.EDU researchers found that LLMs often use a very simple linear function to recover and decode stored facts. Moreover, the model uses the same decoding function for similar types of facts. Linear functions, equations with only two variables and no exponents, capture the straightforward, straight-line relationship between two variables.

SYNTHETIC DATA GENERATION

Is there another way of getting to the next word from generating a response from your perspective, or has there been research done in NLP to accomplish this in a different way?

There has been a significant amount of research exploring alternative architectures and methods for language generation beyond the standard transformer-based approach that relies heavily on self-attention mechanisms. While still an active area of research, some promising directions include:

Energy-Based Models: Instead of modeling the probability of the next token directly, energy-based models learn an energy function that assigns low energies to plausible sequences and high energies to implausible ones. This allows generating coherent text by searching for low-energy sequences. Examples include noise-contrastive estimation and energy-based sequence transducers.
Latent Variable Models: These introduce continuous latent variables to explicitly model high-level semantic abstractions. Examples include variational autoencoders and recent work on discrete latent variables. The idea is to first generate the high-level semantics, then render text conditioned on the semantics.
Pre-trained and Finetuned Diffusion Models: Diffusion models iteratively refine a noisy signal into coherent text by learning the data distribution through gradual denoising. Recent work like DiffusionLM has shown impressive results by first pre-training a diffusion model, then finetuning it for language modeling.
Sparse Transformer Modifications: Approaches like Sparse Transformer aim to improve modeling long-range dependencies by incorporating sparse attentions and untying the word embeddings from the softmax.
Retrieval-Enhanced Models: These augment language models with explicit retrieval over datastore of texts, allowing generation informed by relevant retrieved contexts. Examples include RAG models and OpenAI’s RetriBit.
Structured Output Prediction: Instead of predicting flat text sequences, some models aim to directly predict recursive tree structures representing the compositional semantics of the output.
Neuro-Symbolic Approaches: By integrating symbolic reasoning components like thematic knowledge bases or program induction, these models aim for more grounded and interpretable text generation.

While many of these are still research endeavors, they illustrate the diversity of approaches being explored to move beyond the standard transformer “confabulation” toward more grounded, controllable and interpretable language generation. Combining such approaches with large language models shows promising future directions.

What about in definite domains like process engineering. Can science, math and simulation be used to predict answers, albeit using a prompt model?

Yes, absolutely. For well-defined domains like process engineering, involving scientific principles, mathematical modeling, and physics simulations, there are promising approaches that can leverage these structured knowledge sources to generate more grounded and accurate responses, rather than relying solely on pattern confabulation from text data.

Some key directions in this vein include:

Neuro-Symbolic Reasoning: By integrating large language models with symbolic knowledge bases capturing domain knowledge (e.g. chemical processes, thermodynamic laws, engineering constraints), the model can draw upon these structured representations to reason about and generate outputs conforming to the domain semantics and scientific reality.
Physics-Informed Neural Networks: These neural architectures are constrained to respect physical laws and conservation principles by encoding inductive biases from physics into the neural network design. When combined with language understanding, this could enable generation that respects scientific realities.
Simulation-Augmented Language Models: The language model can be grounded in simulations of the relevant processes/systems by learning to map between natural language descriptions and simulated environments. Outputs can be constrained by simulating their consequences.
Embedding Symbolic Solvers: For problems requiring mathematical derivations or optimization, language models can be augmented with embedded solvers for ODEs, PDEs, optimization routines etc. to generate quantitative solutions while explaining them in natural language.
Retrieval over Scientific Databases: Similar to retrieval-enhanced models, relevant scientific literature, formulas, data tables etc. can be retrieved as context to inform more grounded and precise language generation.

The key idea is to provide strong inductive biases from scientific first principles, rather than relying purely on shallow patterns from text. While still an emerging field, such approaches leveraging domain knowledge and simulations show promise for language models to generate well-grounded responses in definite technical domains like engineering.

Of course, sufficiently specifying the domain constraints and curating the knowledge sources remains a challenge. But combining large language models with scientific domain modeling holds potential for overcoming some of the confabulation tendencies on highly specialized and empirically grounded queries.

NITIN UCHIL – Founder, CEO & Technical Evangelist
nitin.uchil@numorpho.com

BELOW THE FOLD

Discussions by Dr. Vinekar

XII: This is Part 12 of the previous article. It illustrates why Swami Kuvalayananda chose a Sanskrit scholar to encourage him to go to medical school to begin serious scientific research in Yoga, its basics and its therapeutic applications.

This article may try some reader’s patience and may also be uninteresting for some. However, it is necessary to grasp the fact that Yoga literature is in Sanskrit and unless the true meaning of the aphorisms are comprehended and properly interpreted it will be difficult to practice authentic yoga or conduct research as mentioned above.

The article will be of great interest to those who like Sanskrit. It will entertain them if they read the dialogue with Claude.ai. How difficult it is not to project one’s own meanings coming from one’s background or unconscious. This insight was gained by Claude.ai in this interesting conversation with the author.

https://www.linkedin.com/pulse/behind-the-scenes-stories-yoga-science-popularization-vinekar-md-zvgkc/

My repost review:

An aphorism is a concise statement that expresses a general truth or principle, often in a memorable or witty way. Some famous examples of aphorisms include:

– “A penny saved is a penny earned” (Benjamin Franklin)
– “The early bird gets the worm”
– “An apple a day keeps the doctor away”
– “A picture is worth a thousand words”
– “Better late than never”
– “Under promise and over deliver”

Aphorisms are often used in literature, philosophy, and everyday speech to convey deep insights or ideas in a succinct and memorable way. They can also be a powerful tool for persuasion, as they often contain a kernel of truth that resonates with the listener or reader.

In article 12 of exploring the History behind the Science of Yoga, Shree Vinekar, MD, our chief advisor explains how translating Sanskrit aphorisms (sutras) most times dilute or mis-state the essence of the statement.

In a detailed discussion with Claude.ai from Anthropic, Dr. Vinekar prompts Claude with background and basis to eventually come up with the true meaning of the statement: शरीरमाद्यं खलु धर्मसाधनम् (Sharream aadyam khalu dharma saadhanam).

In this age of Large Language models (hashtagllm), such understandings are key in our forays of human-machine interactions so that the outcome – actionable intelligence – is properly executed. This is key to our evolution of what we term “Existential Intelligence” (EI) that would enable the next generation of intelligent solutions that are collaborative, rational and pragmatic.

We at Numorpho Cybernetic Systems (NUMO) are utilizing the insights provided by our advisor to instill responsible, ethical and explainable constructs in our themed development of our hashtaggenAi platform for process engineering, the Mantra M5 based on converse, compute and infer.

In this phrasing:

· It focuses solely on body and mind as the fundamental bases

· Dharma is expressed as upholding one’s essential duty and ethical way of living

· This dharma is tied to being in consonance with the intrinsic order and harmony of the universe

“A sound integration of body and mind forms the primary basis to accord with the intrinsic harmony.”

“Integration of body and mind is the primary basis to accord with the intrinsic harmony.”

“The integrated body and mind are the primordial means to align with one’s true nature.”

In this phrasing:

· There is no instruction or implication that integration must be achieved, it simply states the body and mind as an integrated whole

· “Primordial means” captures the essence of “aadyam saadhanam” – the most fundamental instrument

· “Align with one’s true nature” expresses “dharma” as the innate essence to accord with

· It avoids extraneous qualifiers or connotations