AI Summary

[TOC]

Artificial Intelligence (AI) refers to the simulation of human intelligence in machines, allowing them to perform tasks that typically require human cognitive functions such as learning, reasoning, problem-solving, perception, and decision-making.

AI Types

Types of AI Based on Functionalities

Basis	Reactive AI	Limited Memory AI	Theory of Mind AI	Self-Aware AI
Level of Intelligence	Basic response-based	Learned from data	Social/emotional (emerging)	Conscious-level (hypothetical)
Learning Ability	No learning	Learns from past data	Limited/experimental social learning	Self-reflective learning (theoretical)
Decision-Making	Rule-based, immediate response	Data-driven, probabilistic	Context + emotion + intention-based	Fully autonomous, self-directed
Interaction Complexity	Simple, fixed responses	Context-aware interaction	Human-like social interaction	Highly advanced, human-equivalent (theoretical)
Real-World Status	Fully exists	Widely used today	Research stage	Not yet achieved

Types of AI Models

Developer AI Toolkit

Hierarchy

Workflow

Machine Learning (ML)

Supervised Machine Learning

Unsupervised Machine Learning

Reinforcement Machine Learning

Deep Learning

Neural Networks

Artificial Neural Networks (ANNs)

Convolutional Neural Networks (CNNs)

Recurrent Neural Networks (RNNs)

Long Short-Term Memory Networks (LSTMs)

Generative Adversarial Networks (GANs)

Autoencoders

Natural Language Processing (NLP)

Phase

Computer Vision

Generative AI

Agentic AI

Types of Agents

Simple Reflex Agents

Simple reflex agents act only on the current perception of the environment using predefined condition–action rules. They do not rely on past experiences or predict future outcomes and respond directly using simple “if–then” logic.
Model-Based Reflex Agents

Model-based reflex agents maintain an internal model of the environment to handle situations where full information is not directly available. This helps them make better decisions by considering changes in the environment and the impact of their actions.
Goal-Based Agents

Goal-based agents choose their actions by focusing on a specific objective and evaluating how different choices can help achieve it. Instead of reacting only to the current situation, they plan ahead and consider possible future outcomes.
Utility-Based Agents

Utility-based agents go beyond simply achieving goals by evaluating how beneficial each action is using a utility function, which measures the overall “value” or satisfaction of an outcome. This helps them choose the best option when dealing with trade-offs or uncertainty.
Learning Agents

Learning agents improve their behavior over time by using feedback from past actions. They continuously refine their internal models to make better decisions in future situations.
Multi-Agent Systems (MAS)

Multi-agent systems consist of multiple autonomous agents that interact within a shared environment, where they may cooperate, compete, or do both depending on the situation.
Hierarchical Agents

Hierarchical agents organize decision-making in layers, where higher levels focus on planning and lower levels handle execution. This structure helps manage complex tasks by separating strategy from operational details.

AI Agent Roadmap

Other

Knowledge Representation

Knowledge Representation (KR) in AI focuses on how machines store and organize real-world information so they can reason, learn, and make intelligent decisions like humans.

Declarative Knowledge: Knowledge about facts and concepts it answers what something is . Example: Paris is the capital of France.
Procedural Knowledge: Knowledge about how to perform a task or solve a problem. Example: Steps to sort numbers using an algorithm.
Meta knowledge: Knowledge about other knowledge or how knowledge is used. Example: Knowing that a certain rule works better for solving math problems.
Heuristic Knowledge: Experience-based knowledge or rules of thumb used by experts. Example: A doctor using past experience to guess a possible disease.
Structural Knowledge: Knowledge that shows relationships between concepts. Example: A car is a type of vehicle.

Inference

In AI, inference rules serve as guiding principles for deriving valid conclusions from existing data. These rules underpin the construction of proofs, which constitute chains of reasoning leading to desired outcomes. Within these rules lie key terminologies that delineate relationships between propositions connected by various logical connectives:

Implication

Symbolized by A → B, implication denotes that proposition A implies proposition B, suggesting a cause-and-effect relationship.
Converse

Flipping the implication, placing B on the left and A on the right (B → A), though the converse doesn't ensure the original implication's validity.
Contrapositive

The negation of the converse (¬B → ¬A), offering an equivalent implication with both propositions negated.
Inverse

Symbolized by ¬A → ¬B, the inverse represents the negation of the original implication, albeit not guaranteeing its truth.

Hardware

Library/Tools

NLP Python Library

AI Algorithms

Challenge

Ignorable Problems

These are minor issues that have little or no effect on the AI system’s overall performance. They are often harmless and do not require immediate attention.

Examples:

Small inaccuracies in predictions that don't affect the final result like a slight error in image classification.
Minor issues in data preprocessing that don't change the outcome.

Recoverable Problems

Recoverable problems are those where the AI system encounters an issue but can be fixed with intervention, either automatically or manually, such as error-handling functions.

Examples:

Missing data that can be filled in using statistical methods.
System crashes that can be fixed by restoring from a backup.

Irrecoverable Problems

These are severe issues that cause permanent damage or failure, making it impossible for the system to recover. They can lead to significant performance loss.

Examples:

Corrupted training data that causes bias and reduces the model's effectiveness.
Adversarial attacks that make the model untrustworthy.
Overfitting where the model becomes too specialized and cannot adapt to new data.

Constraint Satisfaction Problems (CSP)

A Constraint Satisfaction Problem is a mathematical problem where the solution must meet a number of constraints. In CSP, the objective is to assign values to variables such that all the constraints are satisfied. Many AI applications use CSPs to solve decision-making problems that involve managing or arranging resources under strict guidelines.

Application

AI in Manufacturing

AI in Transportation

Tools and Framework

Framework

LangChain

LangChain is a framework for building agents and LLM-powered applications. It helps you chain together interoperable components and third-party integrations to simplify AI application development — all while future-proofing decisions as the underlying technology evolves.

LangChain enables Retrieval-Augmented Generation (RAG) by combining document processing, vector storage, and LLMs to generate accurate, context-aware responses. It connects embeddings, vector databases, and models into a smooth workflow.

LangGraph

LangGraph is an open-source framework from RAG#LangChain designed to build and manage AI agent workflows using graph-based structures. It allows developers to define workflows as nodes and edges, making complex agent interactions more structured, scalable, and easier to control.

Start

The process begins with the agent (Assistant) initiating an interaction or task.
Assistant

The process begins with the agent (Assistant) initiating an interaction or task.
Enter Write Sequence

If the task requires writing assistance like generating content, the workflow enters a dedicated writing sequence.
Write Assistant

This specialized module focuses on the writing process. It may loop with tools for refining or editing before completing the sequence.
Leave Write Sequence

Once the writing task is complete the system exits the write mode.
Writer Sensitive Tools and Assistant Tools

These nodes provide specialized capabilities. Depending on the state the Assistant routes tasks to tools that enhance writing or perform sensitive operations.
End

The process concludes once the desired outcome is achieved and all necessary tools have been executed.

AutoGen

AutoGen allows you to develop AI agents that can chat with each other or involve humans in the loop. It is like a collaborative workspace where agents can run code, pull in data from tools, or get human feedback to complete a task.

CrewAI

As the name suggests, CrewAI is all about teamwork. It orchestrates teams of AI agents with roles like writers and editors, processing tasks in a structured workflow. It utilizes LLMs and tools (APIs, Internet, code, etc) to efficiently manage complex task execution and data flow.

LlamaIndex

LlamaIndex is an open-source framework that helps connect private and domain-specific data with large language models to build context-aware AI applications. It simplifies data ingestion, indexing and querying for better and more efficient outputs.

Semantic Kernel

Semantic Kernel connects AI services (OpenAI, Calude, Hugging Face models, etc) with a plugin-based architecture that supports skills, templates, and API integrations for flexible workflows. It supports text search and custom workflows for applications.

Libraries

NLTK

Natural Language Processing (NLP) plays an important role in enabling machines to understand and generate human language. Natural Language Toolkit (NLTK) stands out as one of the most widely used libraries. It provides a combination of linguistic resources, including text processing libraries and pre-trained models, which makes it ideal for both academic research and practical applications.

Unstructured

The Unstructured open source library offers an open-source toolkit designed to simplify the ingestion and pre-processing of diverse data formats, including images and text-based documents such as PDFs, HTML files, Word documents, and more. With a focus on optimizing data workflows for Large Language Models (LLMs), the Unstructured open source library provides modular functions and connectors that work seamlessly together. This cohesive system ensures efficient transformation of unstructured data into structured formats, while also offering adaptability to various platforms and use cases.

Tools and Platforms for Model Deployment

Kubeflow is an open-source machine learning toolkit built on top of Kubernetes. It is utilized for coordinating, delivering, and operating machine learning workloads. By making the deployment procedure straightforward, adaptable, and scalable, it makes machine learning workload deployment simple. Kubeflow can run in a Kubernetes cluster on-premises or the cloud.

MLflow is an open-source platform designed to manage and streamline the entire machine learning lifecycle. It provides a set of tools for tracking experiments, packaging models and deploying them, making it easier to manage the various stages of ML workflows. Whether we are a data scientist, ML engineer or DevOps engineer, MLflow offers a robust solution to track our experiments and manage models throughout their lifecycle.

YOLO (You Only Look Once)

YOLO was proposed by Joseph Redmond et al. in 2015 to deal with the problems faced by the object recognition models at that time, Fast R-CNN was one of the models at that time but it had its own challenges such as that network could not be used in real-time because it took 2-3 seconds to predict an image and therefore could not be used in real-time. Whereas in YOLO we have to look only once in the network i.e. only one forward pass is required through the network to make the final predictions.

Input Preprocessing

The model accepts an image as input. It resizes the input image to 448×448 pixels ensuring that the aspect ratio is preserved using padding. This ensures uniformity of input dimensions across the network which is essential for batch processing in deep learning.
Backbone Convolutional Neural Network (CNN)

After preprocessing the image is passed through a deep CNN architecture designed for object detection:
- The model consists of 24 convolutional layers and 4 max-pooling layers.
- These layers help in extracting hierarchical spatial features from the image.
Use of 1×1 and 3×3 Convolutions:
- To reduce the number of parameters and compress channels, 1×1 convolutions are employed.
- These are followed by 3×3 convolutions to capture spatial patterns in the feature maps.
This design pattern i.e 1×1 followed by 3×3 improves computational efficiency while maintaining expressive power.
Fully Connected Layers

Following the convolutional layers, the architecture has 2 fully connected layers. The final fully connected layer produces an output of shape (1, 1470).
Cuboidal Prediction Output

The output vector of size 1470 is reshaped to (7, 7, 30). Here, 7×7 represents the grid cells, and 30 represents the prediction vector for each cell: $$ 30 = (2 \text{ bounding boxes} \times 5) + (20 \text{ class probabilities}) $$
Activation Functions

The architecture predominantly uses Leaky ReLU as its activation function. The Leaky ReLU is defined as: $$ f(x) = \begin{cases} x, & \text{if } x > 0 \ 0.01x, & \text{if } x \leq 0 \end{cases} $$ This activation allows a small gradient when the unit is not active, preventing dead neurons.
Output Layer Activation

The last layer uses a linear activation function, suitable for making raw predictions like bounding box coordinates and confidence scores.
Regularization Techniques
1. Batch Normalization is employed across layers to stabilize and accelerate training.
2. Dropout is also incorporated to prevent overfitting by randomly deactivating neurons during training, encouraging the network to learn more robust features.

YOLO uses sum-squared error loss function which is easy to optimize. However, this function gives equal weight to the classification and localization task. The loss function defined in YOLO as follows: $$ \lambda_{coord} \sum_{i = 0}^{S^2} \sum_{j = 0}^{B} \mathbb{1}_{ij}^{obj}[(x_i - \hat{x_i})^{2} + (y_i - \hat{y_i})^{2}] \

\lambda_{coord}\sum_{i = 0}^{S^2}\sum_{j = 0}^{B} \mathbb{1}_{ij}^{obj}[(\sqrt{w_i} - \sqrt{\hat{w_i}})^2 + (\sqrt{h_i} - \sqrt{\hat{h_i}})^2] \
\sum_{i = 0}^{S^2}\sum_{j = 0}^{B}\mathbb{1}_{ij}^{obj}(C_i - \hat{C_i})^2 \
\lambda_{noobj}\sum_{i = 0}^{S^2}\sum_{j = 0}^{B} \mathbb{1}_{ij}^{noobj}(C_i - \hat{C_i})^2 \
\sum_{i = 0}^{S^2} \mathbb{1}{i}^{obj} \sum{c \in classes}(p_i(c) - \hat{p_i}(c))^2 $$ where:

$l_{i}^{obj}$ denotes if object is present in cell $i$.
$l_{ij}^{obj}$ denotes $j_{th}$ bounding box responsible for prediction of object in the cell $i$.
$\lambda_{coord}$ and $\lambda_{noobj}$ are regularization parameter required to balance the loss function.

In this model, we take $\lambda_{coord}=5$ and $\lambda_{noobj}=5$.

The first two parts of the above loss equation represent localization mean-squared error, but the other three parts represent classification error:

Localization Error
1. The first term calculates the deviation from the ground truth bounding box.
2. The second term calculates the square root of the difference between height and width of the bounding box. In the second term, we take the square root of width and height because our loss function should be able to consider the deviation in terms of the size of the bounding box.
3. For small bounding boxes, the little deviation should be more important as compared to large bounding boxes.
Classification Loss
1. The first term calculates the sum-squared error between the predicted confidence score that whether the object present or not and the ground truth for each bounding box in each cell.
2. Similarly, the second term calculates the mean-squared sum of cells that do not contain any bounding box and a regularization parameter is used to make this loss small.
3. The third term calculates the sum-squared error of the classes belongs to these grid cells.

At test time we multiply the conditional class probabilities and the individual box confidence predictions. We define our confidence score as follows: $$ \kern 6pc P_{r}\left( \text{Object} \right) * \text{IOU}_{\text{pred}}^{\text{truth}} $$

This results in combination of bounding boxes from each grid like this:

Each grid also predicts $C$ conditional class probability, $P_r$ ($Class_i$ | Object):

This probability were conditional based on the presence of an object in grid cell. Regardless the number of boxes each grid cell predicts only one set of class probabilities. These prediction are encoded in the 3D tensor of size S * S * (5 * B +C).

Now, we multiply the conditional class probabilities and the individual box confidence predictions:

Summary

AI Agent vs MCP vs RAG

An AI agent is a software program that can interact with its environment, gather data, and use that data to achieve predetermined goals. AI agents can choose the best actions to perform to meet those goals.

Model Context Protocol (MCP) is an open standard that allows AI models (like Claude) to connect to databases, APIs, file systems, and other tools without needing custom code for each new integration.

RAG (Retrieval-Augmented Generation) is about what the model knows at runtime. The model stays frozen. No retraining. When a user asks a question, a retriever fetches relevant documents (PDFs, code, vector DBs), and those are injected into the prompt.

Traditional AI vs Agentic AI

Feature	Traditional AI	Agentic AI
Core Function	Performs specific, preprogrammed tasks	Executes tasks autonomously using predefined goals
Typical output	Deterministic results—answers, classifications, predictions	Actions, decisions, multi-step workflows
Autonomy	Low as it requires explicit instructions, operates within set boundaries	High as it plans, adapts, and makes decisions with minimal human direction
Learning	Learns from labeled data, often needs retraining for new situations	Learns from experience, adapts strategies and workflows in real time
Use cases	Data sorting, image recognition, basic diagnostics	Workflow automation, dynamic planning, virtual assistants, problem solving
Scalability	Requires manual oversight as systems grow.	Oversees and coordinates whole systems hence reducing manual monitoring.
Adaptability	Struggles with unexpected changes and may need retraining.	Adjusts strategies and learns in real time and is best suited for fast-changing situations.
Business value	Automates simple, rule-based jobs, increases consistency	Automates complex operations, reduces manual work, and enables personalized tasks

AI vs ML vs DL

Aspect	AI	ML	DL
Scope & Application	Broad – includes ML, DL, expert systems, robotics, computer vision, NLP, symbolic AI, etc.	Narrower – focuses on data-driven algorithms and statistical learning.	Narrowest – focuses specifically on deep neural networks.
Core Techniques	Rule-based systems, search algorithms, expert systems, ML, DL, reinforcement learning, NLP.	Supervised learning, unsupervised learning, reinforcement learning, regression, classification, clustering.	CNNs (Convolutional Neural Networks), RNNs (Recurrent Neural Networks), LSTMs, Transformers, GANs.
Data Type	Can work with structured, semi-structured or unstructured data depending on the approach.	Mainly structured and labeled data (though some algorithms handle unstructured data).	Primarily unstructured data (images, audio, text, video).
Learning Dependency	May or may not involve learning (AI can be purely rule-based).	Always involves learning from historical data.	Fully dependent on large-scale learning with neural networks.
Model Complexity	Can range from simple decision trees to complex hybrid AI systems.	Relatively simpler – linear models, trees, SVMs, ensemble methods.	Very complex – multi-layer neural networks with millions to billions of parameters.
Computation Power	Low to high, depending on the AI technique (expert systems vs DL).	Moderate – runs well on CPUs for most algorithms.	Very high – requires GPUs/TPUs for training large models.

Generative AI vs Traditional AI

Aspect	Traditional AI	Generative AI (GenAI)
Definition	Traditional AI focuses on analyzing existing data, identifying patterns and making predictions or decisions based on predefined logic or learned patterns.	Generative AI focuses on creating new and original content—such as text, images or audio—by learning patterns from existing data.
Purpose	To analyze data and assist in decision-making or automation of specific tasks.	To generate new data or content that mimics human creativity.
Output Type	Predictive or analytical outputs (e.g., classification, recommendation).	Creative or generative outputs (e.g., text, images, videos, code).
Data Usage	Uses data to train models that make accurate predictions.	Uses data to learn structure and generate novel examples from it.
Examples	Spam detection, credit scoring, medical diagnosis.	ChatGPT, DALL·E, Midjourney, Gemini.
Techniques Used	Machine learning algorithms, decision trees, regression models, and rule-based systems.	Deep learning models like Transformers, Diffusion models, and GANs.
Interaction Type	Task-specific and rule-driven.	Conversational, open-ended and creative.

Generative AI vs Agentic AI

Aspect	Generative AI (GenAI)	Agentic AI
Definition	Generative AI focuses on creating new content such as text, images, videos or code by learning from existing data.	Agentic AI focuses on autonomous decision-making and goal-oriented actions by interacting with environments, tools or systems.
Core Function	Generates creative outputs based on learned data patterns.	Acts independently to plan, reason and execute tasks to achieve objectives.
Primary Goal	Content creation and idea generation.	Task automation and problem-solving.
Dependency	Works based on prompts or input from users.	Can operate with minimal human input once a goal is set.
Key Components	Large Language Models (LLMs), Diffusion Models, GANs.	LLMs combined with reasoning, memory, planning and tool-use capabilities.
Examples	ChatGPT (for text), DALL·E (for images).	AutoGPT, LangGraph Agents, ReAct-based Agents.
Output Type	Static (produces content).	Dynamic (takes actions, makes decisions, adapts to outcomes).
Human Involvement	High – relies on user input and direction.	Low – can self-direct and manage workflows.

Generative AI vs Multimodal AI

Aspect	Generative AI	Multimodal AI
Definition	AI systems that generate new content or data outputs	AI systems that integrate and process multiple modalities simultaneously
Primary Function	Create new content, images, text, etc.	Combine and process information from multiple modalities
Input Modality	Typically operates within a single modality (e.g., text or images)	Processes inputs from multiple modalities (e.g., text, image, audio)
Output Type	Outputs new content based on learned patterns	Outputs integrated information from different modalities
Examples	Deep Generative Models (GANs, VAEs), text generation models	Smart assistants (e.g., Alexa, Google Assistant), systems handling image-text data
Use Cases	Creative tasks (art generation, music composition), text synthesis	Information retrieval, intelligent assistants, multimedia processing
Key Technologies	GANs, VAEs, language models (GPT, BERT)	Speech recognition, computer vision, natural language understanding

Variational Autoencoder (VAE) vs Standard Autoencoder

Aspect	Standard Autoencoder	Variational Autoencoder (VAE)
Latent Space	Deterministic – each input maps to a single point.	Probabilistic – each input maps to a distribution (mean and variance).
Purpose	Dimensionality reduction, feature learning, reconstruction.	Generative modeling, data synthesis, learning underlying distributions.
Output	Reconstructed input.	Reconstructed input or novel samples generated from latent distribution.
Training	Minimizes reconstruction loss (e.g., MSE).	Minimizes reconstruction loss plus KL divergence to regularize latent space.
Generative Ability	Cannot generate truly new data outside training set.	Can generate novel data by sampling from latent distribution.
Use Cases	Denoising, compression, feature extraction.	Image/text generation, anomaly detection, creative AI tasks.

Diffusion Model vs GANs

Aspect	GANs (Generative Adversarial Networks)	Diffusion Models
Definition	GANs are generative models where a generator creates data and a discriminator evaluates it, improving both through adversarial training.	Diffusion models generate data by iteratively denoising random noise, learning how to reverse a noise process step by step.
Performance (Quality of Output)	Can produce highly realistic data quickly, but sometimes suffers from mode collapse (limited diversity).	Usually generates high-quality and diverse outputs, often sharper than GANs, especially for complex data.
Training Stability	Training is often unstable due to adversarial nature; careful tuning is required.	Training is generally more stable, as it is based on a likelihood optimization rather than adversarial competition.
Speed of Generation	Fast at inference once trained.	Slower, as generation involves multiple iterative denoising steps.
Data Diversity	May generate less diverse samples if mode collapse occurs.	High diversity due to probabilistic sampling in latent space.
Use Cases	Image synthesis, video generation, data augmentation.	Image generation, audio/video synthesis, high-fidelity generative tasks.

Cross-Attention vs Self-Attention

Aspect	Cross-attention	Self-attention
Definition	Allows the decoder to use information from the encoder.	Allows each word in a sequence to focus on other words in the same sequence.
Uses	Used in tasks like translation, where the model must look at the entire input (encoder) while generating output.	Used to understand relationships within a single input (e.g. text) without external context.
Focus	Focuses on picking useful information from another part of the model.	Focuses on how words in the same sentence relate to each other.
Data Flow	Involves interaction between two different data parts (encoder and decoder).	Data flows within the same sequence (internal attention).
Example	In translation, cross-attention helps the decoder choose the right words from the encoder.	In text analysis, self-attention helps the model understand how words in the sentence connect.

Fine-tuning vs Transfer Learning

Aspect	Fine-Tuning	Transfer Learning
Definition	Fine-tuning is the process of taking a pre-trained model and further training it on a specific task or dataset to improve performance for that task.	Transfer learning is a broader concept where knowledge learned from one task or domain is applied to a different but related task or domain.
Scope	Usually focuses on a specific downstream task.	Can be applied to multiple tasks or domains, not limited to a single task.
Training Requirement	Often requires task-specific labeled data.	May require less data for the new task since it uses existing learned knowledge.
Model Modification	Can involve adjusting weights of the entire model or only certain layers.	Often involves reusing pre-trained model features, sometimes freezing layers and only training a few new layers.
Goal	Optimize a pre-trained model to maximize performance on a target task.	Uses prior knowledge to accelerate learning and improve performance on a new, related task.
Use Cases	Fine-tuning GPT for legal document summarization, BERT for sentiment analysis.	Using ImageNet pre-trained CNNs for medical image classification, BERT for different NLP tasks.

RAG vs Fine-tuning

	RAG	Fine Tuning
Nature of Task	RAG is ideal for tasks requiring contextual understanding and the incorporation of external knowledge, like question answering or content summarization, financial report generation, etc.	Fine-tuning is suitable for tasks where adaptation to specific patterns within a domain is crucial, like sentiment analysis, document classification, or for more creative tasks (music or novel generation).
Data Availability	RAG always requires a knowledge base for effective retrieval, which may limit applicability in domains with sparse external information.	Fine-tuning is more adaptable to scenarios with limited task-specific data, leveraging pre-existing knowledge during the pre-training phase.
Computational Intensity	RAG is very computationally intensive, particularly during the retrieval process, potentially affecting real-time applications.	Fine-tuning generally less computationally demanding, making it more suitable for applications with strict latency requirements.
Output Diversity	RAG excels in generating diverse and contextually relevant outputs due to its knowledge retrieval mechanism.	Fine-tuning can only efficiently adapt to specific domains during training, and we need to perform overall re-training for working in new domains.
Knowledge Source	RAG fully depends on external knowledge sources, which may introduce biases or inaccuracies depending on the quality of the retrieved information.	Fine-tuning can't be biased but limited to the knowledge encoded during pre-training, with potential challenges in adapting to entirely new or niche domains.
Use Cases	RAG is well-suited for tasks that benefit from a blend of generative capabilities and access to external information, like chatbots in customer support or ChatGPT.	Fine-tuning is effective for domain-specific applications like healthcare document analysis or sentiment analysis in specific industries.
Training Complexity	RAG involves joint training for both generative and retrieval components, adding complexity to the training process.	Fine-tuning involves simpler training procedures, especially when leveraging pre-trained models with readily available task-specific datasets.

MCP vs A2A Protocol

Google’s Agent-to-Agent (A2A) Protocol enables AI agents to communicate and collaborate, allowing them to delegate tasks, share results, and enhance each other’s capabilities.

Feature	Agent2Agent (A2A)	Model Context Protocol (MCP)
Primary Focus	Facilitates communication and collaboration between autonomous agents.	Enables interaction between a model and external tools or data sources.
Originator	Google	Anthropic
Key Technical Concepts	Agent Cards, Tasks, Messages (Parts), HTTP/JSON-RPC, SSE for real-time streaming.	Host, Client, Server, Tools, Resources, Prompts.
Communication	Task-based, asynchronous communication with potential natural language tasks.	Structured requests for accessing external tools and contextual data, typically using specific schemas like JSON Schema.
Primary Use Case	Supports collaborative workflows across independent agents in various systems.	Facilitates AI models' access to external data, files, and APIs.

Symbolic AI vs Connectionist AI

Aspect	Symbolic AI	Connectionist AI
Definition	AI based on explicit rules and logic to represent knowledge.	AI based on neural networks, learning patterns from data.
Knowledge Representation	Uses symbols, facts and logic statements (e.g., “IF…THEN…” rules).	Uses distributed representations across nodes in a network.
Learning	Limited learning; mostly pre-programmed rules.	Learns from data; adapts over time.
Example	Expert systems, Prolog-based reasoning systems.	Neural networks for pattern recognition, speech or image recognition.
Strengths	Good at reasoning, explainable, interpretable.	Good at handling noisy or unstructured data.
Limitations	Cannot handle ambiguity well; rigid.	Difficult to interpret; “black-box” behavior.

Parametric vs Non-Parametric Models

Aspect	Parametric Models	Non-Parametric Models
Definition	Models with a fixed number of parameters.	Models where number of parameters grows with data.
Assumption	Assumes a specific functional form for data distribution.	Makes few or no assumptions about data distribution.
Learning	Learns a fixed set of parameters from training data.	Learns data patterns directly from training data.
Example	Linear regression, Logistic regression.	k-Nearest Neighbors (k-NN), Decision Trees.
Strengths	Efficient, simpler, easier to interpret.	Flexible, can model complex distributions.
Limitations	Limited flexibility; may underfit if model is wrong.	Computationally expensive; may overfit with small data.

Deterministic vs Stochastic Environment

Aspect	Deterministic Environment	Stochastic Environment
Predictability	Outcomes are completely predictable.	Outcomes are uncertain and can vary even with the same initial conditions and actions.
Modeling	Models are simpler as they do not need to account for uncertainty.	Models must incorporate uncertainty, often making them more complex.
Techniques	Uses algorithmic approaches like depth-first search, breadth-first search, A* algorithm.	Employs probabilistic reasoning, Bayesian networks, Markov decision processes, and reinforcement learning.
Examples	Chess, checkers, puzzle solving.	Autonomous driving, stock market analysis, weather forecasting.
Control and Planning	Planning and control are straightforward due to the lack of randomness.	Planning must consider multiple potential outcomes and adapt dynamically.
Testing and Validation	Easier to test and validate because scenarios can be exactly reproduced.	Testing is challenging due to inherent randomness; scenarios cannot be exactly reproduced.