Artificial Intelligence is evolving rapidly, and new AI terms appear almost every day. If you are a software developer starting your journey into AI, terms such as LLM, RAG, Embeddings, AI Agents, MCP, and Fine-Tuning can initially be confusing.
This guide explains the most commonly used AI terms in simple language, with practical examples that will help developers understand the modern AI ecosystem.
1. AI – Artificial Intelligence
AI stands for Artificial Intelligence.
Artificial Intelligence refers to computer systems that can perform tasks that normally require human intelligence, such as understanding language, recognizing images, solving problems, making predictions, and generating content.
Example: ChatGPT answering a question is an example of an AI application.
2. ML – Machine Learning
ML stands for Machine Learning.
Machine Learning is a branch of AI where computers learn patterns from data instead of being explicitly programmed with rules for every possible situation.
Example: An email system can learn from previous emails to identify whether a new email is spam.
3. DL – Deep Learning
DL stands for Deep Learning.
Deep Learning is a type of Machine Learning that uses neural networks with multiple layers to learn complex patterns from large amounts of data.
Deep Learning is widely used in image recognition, speech recognition, recommendation systems, and modern AI models.
4. GenAI – Generative AI
GenAI stands for Generative Artificial Intelligence.
Generative AI is AI that can create new content instead of only analyzing existing information.
- Text
- Images
- Audio
- Video
- Computer code
- Documents
Example: An AI tool generating a C# class from a natural-language description is an example of Generative AI.
5. LLM – Large Language Model
LLM stands for Large Language Model.
An LLM is an AI model trained on a very large amount of data to understand and generate human language.
LLMs can perform tasks such as:
- Answering questions
- Writing content
- Summarizing documents
- Translating languages
- Generating computer code
- Analyzing text
- Solving reasoning problems
Examples of LLM families include GPT, Llama, Gemini, Claude, and Qwen.
6. SLM – Small Language Model
SLM stands for Small Language Model.
An SLM is a smaller language model designed to use fewer computing resources than large language models.
SLMs are useful for:
- Local AI applications
- Mobile applications
- Edge devices
- Private applications
- Low-latency applications
7. NLP – Natural Language Processing
NLP stands for Natural Language Processing.
NLP is the field of AI that focuses on enabling computers to understand, process, analyze, and generate human language.
Examples: Translation, sentiment analysis, chatbots, text summarization, and speech processing.
8. Transformer
A Transformer is a neural network architecture that became the foundation of many modern AI language models.
Transformers use an attention mechanism that allows the model to determine which parts of the input are important when processing information.
Many modern LLMs are based on Transformer architecture.
9. Token
A token is a unit of text processed by an AI model.
A token can represent a complete word, part of a word, punctuation, or another piece of text.
Input: Artificial Intelligence is powerful. Possible tokens: Artificial | Intelligence | is | powerful | .
Token counts are important because AI model context limits and many AI API pricing models are based on tokens.
10. Context Window
The context window is the maximum amount of information an AI model can process or consider at one time.
The context can include:
- User prompts
- Previous conversation messages
- Documents
- Source code
- Tool results
A larger context window allows an AI model to work with larger amounts of information in a single request.
11. Parameters
Parameters are numerical values learned by a neural network during training.
They influence how the model processes information and generates output.
You may see models described as:
7B parameters 14B parameters 70B parameters
Here, B means billion.
12. Training
Training is the process of teaching an AI model using data.
During training, the model adjusts its parameters to learn patterns from the training data.
Training a large language model can require enormous amounts of data, computing power, and time.
13. Fine-Tuning
Fine-Tuning means taking an already trained AI model and training it further for a specific task, domain, or behavior.
For example, a general-purpose LLM could be fine-tuned for:
- Customer support
- Medical terminology
- Legal documents
- Programming
- Company-specific terminology
14. SFT – Supervised Fine-Tuning
SFT stands for Supervised Fine-Tuning.
In SFT, a model is trained using examples where the expected output is provided.
Question: What is dependency injection? Expected Answer: Dependency injection is a design pattern used to provide dependencies to a class...
The model learns from these examples and becomes better at producing the desired type of output.
15. RLHF – Reinforcement Learning from Human Feedback
RLHF stands for Reinforcement Learning from Human Feedback.
RLHF uses human feedback to help align an AI model with desired behaviors and preferences.
Human evaluators can compare different responses and indicate which responses are more useful, accurate, or appropriate.
16. Embeddings
An Embedding converts information such as text into a numerical vector that represents its meaning.
For example, the following sentence can be converted into a numerical representation:
"How do I reset my password?"
Text with similar meanings generally produces embeddings that are mathematically closer together.
Embeddings are widely used in semantic search, recommendation systems, and RAG applications.
17. Vector Database
A Vector Database is a database designed to store and search vector embeddings efficiently.
A typical AI search system can work like this:
Document ↓ Chunking ↓ Embedding ↓ Vector Database ↓ Similarity Search
Vector databases are commonly used in AI applications that need to search large collections of documents based on meaning.
18. Semantic Search
Semantic Search searches based on the meaning of a query rather than only matching exact keywords.
For example, a user might search for:
"How can I change my password?"
A document might contain:
"Procedure for resetting account credentials."
A semantic search system can recognize that these two statements have a similar meaning even though they use different words.
19. RAG – Retrieval-Augmented Generation
RAG stands for Retrieval-Augmented Generation.
RAG allows an AI application to retrieve relevant information from an external knowledge source before asking the LLM to generate an answer.
A simplified RAG architecture looks like this:
User Question
↓
Create Embedding
↓
Vector Search
↓
Retrieve Relevant Documents
↓
Send Context to LLM
↓
Generate Answer
RAG is useful when an AI application needs to work with private, company-specific, or frequently changing information.
20. Chunking
Chunking means breaking a large document into smaller sections before storing or processing it.
For example:
Large PDF ↓ Document Chunks ↓ Embeddings ↓ Vector Database
Good chunking is an important part of building an effective RAG system.
21. Hallucination
An AI Hallucination occurs when an AI generates information that appears convincing but is incorrect, unsupported, or completely fabricated.
For example, an AI coding assistant might generate an API method that does not actually exist.
Techniques such as RAG, grounding, tool calling, validation, and human review can help reduce the impact of hallucinations.
22. Grounding
Grounding means connecting an AI model's response to reliable external information.
Instead of relying only on information learned during training, an application can provide the model with current information from:
- Databases
- APIs
- Company documents
- Search results
- Business systems
23. AI Agent
An AI Agent is an AI system that can understand a goal, decide what actions to take, use tools, and perform multiple steps to accomplish a task.
Unlike a simple chatbot that mainly generates a response, an agent can interact with external systems.
User: Find production errors and create a report. AI Agent: ↓ Query logs ↓ Analyze errors ↓ Group similar issues ↓ Generate report ↓ Save report
24. Agentic AI
Agentic AI refers to AI systems designed to perform tasks with a greater degree of autonomy.
The system can determine the next action required to achieve a goal rather than simply answering one question.
Agentic AI is becoming increasingly important in software development, automation, customer service, research, and business workflows.
25. Tool Calling
Tool Calling allows an AI model to request the execution of external tools.
Tools can include:
- APIs
- Databases
- Search engines
- Calculators
- File systems
- Business applications
For example:
User ↓ LLM ↓ "Get customer information" ↓ Customer API ↓ Customer Data ↓ LLM ↓ Final Response
26. Function Calling
Function Calling is a structured mechanism that allows an LLM to request execution of a specific function.
For example:
getWeather("Chennai")
The application executes the function and sends the result back to the AI model.
Function calling is particularly useful when building AI applications with APIs and backend services.
27. MCP – Model Context Protocol
MCP stands for Model Context Protocol.
MCP provides a standardized way for AI applications to connect AI models with external tools, data sources, and resources.
A simplified architecture looks like:
AI Application
↓
MCP
↓
┌────┼────┐
↓ ↓ ↓
Files Database APIs
MCP is becoming an important concept for modern AI applications and agent-based architectures.
28. Multi-Agent System
A Multi-Agent System uses multiple AI agents that collaborate to complete a larger task.
For example:
Manager Agent
↓
┌────┼─────┐
↓ ↓ ↓
Code Test Research
Agent Agent Agent
Each agent can specialize in a specific responsibility.
29. Local LLM
A Local LLM is a language model that runs directly on your own computer or infrastructure instead of sending requests to a cloud AI service.
Advantages can include:
- Improved privacy
- Offline operation
- Greater control
- Reduced dependency on external APIs
The main limitation is that running larger models requires more powerful hardware.
30. On-Device AI
On-Device AI means AI processing happens directly on a device such as a smartphone, laptop, PC, or IoT device.
On-device AI can reduce latency and can provide better privacy because data does not always need to be sent to a remote server.
31. Edge AI
Edge AI means performing AI processing close to where the data is generated rather than sending all data to a centralized cloud system.
Examples include AI running on cameras, vehicles, industrial machines, and mobile devices.
32. Quantization
Quantization reduces the numerical precision used to represent model weights.
For example:
FP16 → 16-bit INT8 → 8-bit Q4 → approximately 4-bit
Quantization can significantly reduce the memory requirements of an AI model and make it easier to run large models locally.
33. GGUF
GGUF is a model file format commonly used for running quantized language models locally.
It is widely associated with the llama.cpp ecosystem and is supported by many local AI tools.
34. LoRA – Low-Rank Adaptation
LoRA stands for Low-Rank Adaptation.
LoRA is a parameter-efficient technique for fine-tuning AI models.
Instead of modifying the entire model, LoRA trains a much smaller set of additional parameters.
This can make fine-tuning significantly more resource-efficient.
35. QLoRA
QLoRA combines Quantization and LoRA.
It allows developers to fine-tune quantized models while keeping memory requirements relatively low.
36. Knowledge Distillation
Knowledge Distillation is a technique where a smaller AI model learns useful behavior from a larger model.
Large Model
↓
Teacher
↓
Knowledge
↓
Small Model
↓
Student
The goal is to create a smaller and faster model while retaining useful capabilities.
37. Prompt
A Prompt is the instruction or input given to an AI model.
For example:
Explain dependency injection in C# with a simple example.
The AI model processes the prompt and generates an appropriate response.
38. Prompt Engineering
Prompt Engineering is the practice of designing effective instructions for AI models.
A good prompt can specify:
- Role
- Task
- Context
- Constraints
- Expected output format
- Examples
39. System Prompt
A System Prompt contains high-priority instructions that define how an AI system should behave.
For example:
You are a C# coding assistant. Provide production-ready code. Explain important design decisions.
40. Zero-Shot
Zero-Shot means asking an AI model to perform a task without providing examples.
Classify the following sentence as Positive or Negative: "The application is very easy to use."
No examples are provided to the model.
41. Few-Shot
Few-Shot prompting provides a small number of examples before asking the model to perform a task.
Input: Great product → Positive Input: Terrible service → Negative Input: Excellent support → ?
The examples help the model understand the expected output.
42. Multimodal AI
Multimodal AI refers to AI systems that can work with multiple types of information.
For example:
Text + Image + Audio + Video
↓
AI Model
A multimodal AI system might analyze an image, understand spoken audio, and respond using text.
43. VLM – Vision-Language Model
VLM stands for Vision-Language Model.
A VLM can understand both images and text.
For example, a developer can provide a screenshot and ask:
"What is wrong with this user interface?"
The model can analyze the image and provide a textual response.
44. ASR – Automatic Speech Recognition
ASR stands for Automatic Speech Recognition.
ASR converts spoken language into text.
Voice ↓ ASR ↓ Text
Voice assistants, meeting transcription systems, and voice-based applications commonly use ASR.
45. TTS – Text-to-Speech
TTS stands for Text-to-Speech.
TTS converts written text into spoken audio.
Text ↓ TTS ↓ Voice
TTS is commonly used in voice assistants, accessibility applications, and AI-powered voice applications.
46. Inference
Inference is the process of using a trained AI model to generate an output.
For an LLM, the process can be represented as:
Prompt ↓ Model Inference ↓ Generated Tokens ↓ Response
47. Latency
Latency is the amount of time required for an AI system to produce a response.
Lower latency generally results in a faster and more responsive user experience.
48. Throughput
Throughput measures how much work an AI system can process within a given amount of time.
For LLMs, throughput is often measured using tokens per second.
49. Benchmark
A Benchmark is a standardized test used to evaluate or compare AI models.
Different benchmarks can measure different capabilities, including:
- Mathematics
- Reasoning
- Coding
- Language understanding
- General knowledge
50. MoE – Mixture of Experts
MoE stands for Mixture of Experts.
MoE is a model architecture where different parts of the model, called experts, can specialize in different types of input.
Instead of activating the entire model for every request, the system can route an input to selected experts.
This can allow models to have a very large total number of parameters while using only a portion of them for each individual request.
Quick AI Terminology Cheat Sheet
| Term | Full Form | Simple Meaning |
|---|---|---|
| AI | Artificial Intelligence | Machines performing tasks that require human-like intelligence. |
| ML | Machine Learning | Learning patterns from data. |
| DL | Deep Learning | Machine learning using deep neural networks. |
| GenAI | Generative AI | AI that creates new content. |
| LLM | Large Language Model | AI model designed to understand and generate language. |
| SLM | Small Language Model | Smaller language model requiring fewer resources. |
| NLP | Natural Language Processing | AI technology for processing human language. |
| RAG | Retrieval-Augmented Generation | Retrieving external information before generating an answer. |
| MCP | Model Context Protocol | Standardized connection between AI applications and external tools/data. |
| VLM | Vision-Language Model | AI that understands images and text. |
| ASR | Automatic Speech Recognition | Converts speech into text. |
| TTS | Text-to-Speech | Converts text into speech. |
| SFT | Supervised Fine-Tuning | Fine-tuning using examples with expected outputs. |
| RLHF | Reinforcement Learning from Human Feedback | Using human preferences to improve model behavior. |
| LoRA | Low-Rank Adaptation | Efficient method for fine-tuning models. |
| QLoRA | Quantized LoRA | Combines quantization with LoRA fine-tuning. |
| MoE | Mixture of Experts | Architecture using specialized model experts. |
| GGUF | GGUF Model Format | Model format commonly used for local LLMs. |
How These AI Technologies Fit Together
Understanding how these terms connect is more useful than simply memorizing their definitions.
AI APPLICATION
│
┌────────────┴────────────┐
│ │
LLM AI Agent
│ │
│ Tool Calling
│ │
│ MCP
│ │
└──────────┬──────────────┘
│
RAG
│
Semantic Search
│
Embeddings
│
Vector Database
A model can also go through a process such as:
Large Dataset
↓
Pre-training
↓
LLM
↓
Fine-Tuning
↓
LoRA / QLoRA
↓
Quantization
↓
Local LLM
Conclusion
Artificial Intelligence has introduced a large number of new concepts, but you don't need to learn everything at once.
If you are a software developer starting with AI, focus first on these concepts:
LLM ↓ Tokens ↓ Transformer ↓ Embeddings ↓ RAG ↓ Vector Database ↓ Prompt Engineering ↓ Tool Calling ↓ AI Agents ↓ MCP ↓ Fine-Tuning ↓ LoRA ↓ Quantization ↓ Local LLM
Once you understand these concepts, modern AI architectures become much easier to understand. The next step is to start building small applications that combine an LLM with your existing software development skills, APIs, databases, and business logic.
No comments:
Post a Comment