LLM, GPT, Generative AI… A One-Stop Guide to Confusing AI Terms
Ever frozen in a meeting when someone said, "Can't we just do that with RAG?" Work through this guide's 30 core AI terms, and in your next meeting you'll be the one explaining.
Why You Need an AI Glossary
In a B2B sales meeting, when a client says **"We're considering adopting generative AI,"** being able to explain the difference between an LLM and an SLM makes you 2x more credible.
In internal DX (digital transformation) meetings, using the right terms cuts communication errors by more than 60%.
Before: "AI… can't we just use something like ChatGPT?"
After: "For our use case, the most cost-effective approach is a RAG-based SLM, fine-tuned on our internal data."
Step 1: The Absolute Basics — Understand the 5-Term AI Hierarchy
In this step, you'll draw a map of the AI world. The 5 terms below are nested, from top to bottom: each one contains the next.
1. AI (Artificial Intelligence)
One-line definition: The umbrella term for any technology that mimics human intelligence.
How to use it in a meeting: "Which category of AI does this fall under?"
💡 AI is the biggest umbrella. Every other term below falls inside it.
2. ML (Machine Learning)
One-line definition: A subfield of AI that learns from data to find patterns.
Examples: Spam email filters, customer churn prediction models
Before/After: Rule-based filter (70% accuracy) → ML-based filter (95% accuracy)
3. DL (Deep Learning)
One-line definition: A machine learning method that trains artificial neural networks with many stacked layers.
Key point: "Deep" means the neural network has many layers.
4. Generative AI
One-line definition: AI that doesn't just analyze existing data but creates new content (text, images, code).
Examples: ChatGPT (text), Midjourney (images), Suno (music)
💡 It's easiest to explain by contrast with "analytical AI." Analytical AI says, "This customer has an 80% chance of churning"; generative AI says, "Draft an email to send to this customer."
5. Foundation Model
One-line definition: A general-purpose AI model pretrained on massive datasets, serving as a base model for a wide range of tasks.
Examples: GPT-4, Claude, Gemini, LLaMA
Hands-on prompt — try it in ChatGPT:
"Explain the relationship between AI, machine learning, deep learning, generative AI, and foundation models in a way an elementary school student could understand, using Russian matryoshka dolls as an analogy."
Step 2: Master 7 Core Language Model Terms
This step covers the terms that come up most often in real work.
6. LLM (Large Language Model)
One-line definition: An ultra-large language model trained on billions to trillions of words of text data.
Leading models: GPT-4 (OpenAI), Claude (Anthropic), Gemini (Google)
Meeting example: "We need to decide whether to use the LLM directly or integrate it through an API for this task."
7. SLM (Small Language Model)
One-line definition: A lightweight model with fewer parameters than an LLM, optimized for specific tasks.
Leading models: Phi-3 (Microsoft), Gemma (Google)
Before/After: LLM API costs of KRW 5 million (≈ US$3,740) per month → KRW 800,000 (≈ US$600) per month when self-hosting an SLM
💡 In B2B sales, when a customer worries about cost, offer an “SLM + fine-tuning” combination as an alternative.
8. GPT (Generative Pre-trained Transformer)
One-line definition: The name of OpenAI’s LLM series, and also a technical architecture meaning “Generative Pre-trained Transformer.”
Key point: GPT is both a proper noun (an OpenAI product) and a technical architecture. “GPT-family models” refers to the latter.
9. Transformer
One-line definition: A deep learning architecture announced by Google in 2017, and the engine behind nearly every LLM today.
Why it matters: It began with the paper “Attention Is All You Need” and changed the history of AI.
💡 A handy analogy: the Transformer is the car engine, and GPT is a specific car model built with that engine.
10. Parameter
One-line definition: The number of internal variables an AI model adjusts during training. It is commonly used to express a model’s “size.”
Getting a feel for it: GPT-3 has 175 billion parameters; GPT-4 is estimated at about 1.8 trillion.
In meetings: “This is a 70B (70 billion) parameter-class model, so we’ll need on-premises GPUs.”
11. Token
One-line definition: The smallest unit in which an LLM processes text. In Korean, one character is typically 1–2 tokens.
Cost calculation: With GPT-4o, input costs about $2.50 per 1 million tokens → one A4 page of Korean (about 1,500 characters) ≈ 2,000–3,000 tokens
💡 Since “tokens = cost,” writing concise prompts reduces your API costs.
12. Context Window
One-line definition: The maximum number of tokens a model can read and process at once.
Comparison: GPT-3.5 (4K tokens) → GPT-4o (128K tokens) → Claude 3.5 (200K tokens)
Practical meaning: The larger the context window, the longer the documents you can feed in and analyze at once.
Practice prompt:
“Read this sales proposal (text below) and summarize the 3 points the customer will care about most. [Paste the full proposal]”
Step 3: Putting It to Work — 8 terms on how to use models
13. Prompt
One-line definition: The question or instruction you give an AI model.
Before: “Organize the sales data” → Result: a vague table
After: “Break down Q3 2024 sales by department and make a table showing the change vs. the previous quarter in %” → Result: a table you can drop straight into a report
14. Prompt Engineering
One-line definition: The technique of systematically designing prompts to get the results you want.
3 core techniques:
Assign a role: “You are a B2B sales strategist with 10 years of experience”
Specify a structure: “As a table / in 3 points / within 500 characters”
Provide examples: “Write it like the example below”
15. Few-shot / Zero-shot / One-shot
Zero-shot: Asking directly with no examples. “Write an email draft”
One-shot: Giving 1 example. “Write an email like the example below [1 example]”
Few-shot: Giving 2–5 examples.
Before (Zero-shot): 60% consistency → After (Few-shot): 90% consistency
16. RAG (Retrieval-Augmented Generation)
One-line definition: A technique in which an LLM searches an external database for reference before generating its answer.
Why it matters: LLMs don’t know anything after their training cutoff. With RAG, they can answer based on your company’s latest internal documents.
B2B use: Connect 1,000 pages of a client’s manuals to RAG → new hires can instantly get expert-level answers
💡 “Fine-tuning vs. RAG” is a question that comes up a lot. Fine-tuning changes the model itself; RAG leaves the model as is and simply connects it to reference material. RAG is far lighter on cost and time.
17. Fine-tuning
One-line definition: Training a pre-trained model further on data from a specific domain.
Example: General-purpose GPT → fine-tuned on 100,000 legal documents → a legal-specialized AI
Cost estimate: Fine-tuning OpenAI GPT-4o mini costs $3.00 per 1 million training tokens
18. Hallucination
One-line definition: When an AI produces information that sounds plausible but isn’t true.
Business risk: Putting a statistic that doesn’t exist into a proposal destroys your credibility.
💡 How to prevent it: ① Add a source-request to your prompt ② Use RAG ③ Make a fact-check step mandatory
19. Embedding
One-line definition: Converting text or images into numeric vectors (coordinates), which lets an AI compare meaning.
Business example: Embed 10,000 customer inquiries → automatically group similar ones → response time down 40%
20. API (Application Programming Interface)
One-line definition: A connection point that lets external programs call and use an AI model’s capabilities.
In meetings: “If we connect the GPT-4o API to our in-house app, we don’t need to build a separate AI.”
Before: Employees type into a ChatGPT window by hand (3 min per case) → After: Automatic API integration (0.5 sec per case)
Practice prompt:
“I’m a B2B SaaS sales rep. When a customer asks, ‘Should we adopt RAG or fine-tuning?’, build me a comparison table on three criteria: cost, speed, and accuracy. Include concrete numeric examples for each item.”
Step 4: 8 AI Service & Business Terms
21. ChatGPT
One-line definition: The name of the service OpenAI built by putting a conversational interface on top of its GPT models.
💡 GPT is the engine; ChatGPT is the car. Don’t use them interchangeably.
22. Copilot
One-line definition: Microsoft’s AI assistant brand, built into Office 365, Windows, GitHub, and more.
Business example: In Excel, typing “Turn the trend in this data into a chart” into Copilot
23. AI Agent
One-line definition: An AI system that makes its own decisions and uses multiple tools to complete a task without human intervention.
Example: Automatically carrying out “collect competitor news every day → summarize → send to Slack”
Before: A staff member spends 1 hour every day → After: The agent runs it automatically; a person only reviews (10 min)
24. Multimodal
One-line definition: The ability to process multiple types of data at once, such as text, images, audio, and video.
Example: Upload a product photo to GPT-4o and ask, “Give me 3 sales points for this product”
25. On-premise vs. Cloud
On-premise: Installing and running the AI model on your own servers.
Cloud: Using AI on external servers such as AWS or Azure.
How to choose: Security-sensitive data (healthcare, finance) → on-premise / Fast adoption → cloud
Cost comparison: On-premise initial investment of roughly KRW 300 (≈ US$1)–500 million vs. cloud at KRW 5 (≈ US$1)–20 million per month
26. Open-Source Model
One-line definition: A model whose weights (the training results) are made public so anyone can download, modify, and distribute it.
Leading models: Meta’s LLaMA 3, Mistral, Falcon
💡 Open source doesn’t mean free. Infrastructure (GPU servers) to run it costs extra.
27. Token Economy (Pricing Structure)
One-line definition: A cost structure in which API usage is billed by the number of input and output tokens.
Comparison (as of late 2024):
GPT-4o: input $2.50 / output $10.00 (per 1 million tokens)
Claude 3.5 Sonnet: input $3.00 / output $15.00 (per 1 million tokens)
GPT-4o mini: input $0.15 / output $0.60 (per 1 million tokens)
28. Guardrail
One-line definition: Setting up safeguards so AI doesn’t give inappropriate or risky answers.
Why it matters in B2B: A customer-facing chatbot without guardrails creates brand risk.
Example settings: “No disparaging competitors,” “No medical or legal advice,” “No unverified figures”
Step 5: The Final 2 Trend & Strategy Terms
29. AGI (Artificial General Intelligence)
One-line definition: AI that can perform any intellectual task like a human, not just a specific one.
Current status: Not yet achieved. Today’s GPT-4o, Claude, and similar models are at the ANI (artificial narrow intelligence) stage.
In meetings: “Today’s AI is not AGI but AI specialized for specific tasks, so we need to define the scope of adoption clearly.”
30. AI Governance
One-line definition: A framework for managing policy, ethics, security, and regulatory compliance for AI use within an organization.
Essentials: Data privacy policy, AI output verification process, employee AI usage guidelines
Before: Everyone uses AI their own way → risk of data leaks → After: Governance in place → 0 security incidents
Step 6: All 30 Terms at a Glance — Cheat Sheet
No. | Term | Quick explanation | Category |
|---|---|---|---|
1 | AI | Umbrella term for technology that mimics human intelligence | Basics |
2 | ML | AI that learns patterns from data | Basics |
3 | DL | ML with deeply stacked neural networks | Basics |
4 | Generative AI | AI that creates new content | Basics |
5 | Foundation Model | Large-scale pretrained general-purpose model | Basics |
6 | LLM | Very large language model | Language Model |
7 | SLM | Lightweight, specialized language model | Language Model |
8 | GPT | OpenAI’s Generative Pre-trained Transformer | Language Model |
9 | Transformer | Core architecture of modern LLMs | Language Model |
10 | Parameters | Number of learned variables inside a model | Language Model |
11 | Token | Smallest unit of text an LLM processes | Language Model |
12 | Context window | Maximum number of tokens processed at once | Language Model |
13 | Prompt | Instructions you give to AI | Usage |
14 | Prompt Engineering | Designing prompts systematically | Usage |
15 | Few-shot/Zero-shot | Prompting with or without examples | Usage |
16 | RAG | Retrieves external data, then generates answers | Usage |
17 | Fine-tuning | Additional training on domain-specific data | Usage |
18 | Hallucination | When AI generates false information | Usage |
19 | Embedding | Converts text into numeric vectors | Usage |
20 | API | Connection channel for calling AI features | Usage |
21 | ChatGPT | OpenAI’s conversational AI service | Service |
22 | Copilot | Microsoft’s AI assistant brand | Service |
23 | AI agent | AI that performs tasks autonomously | Service |
24 | Multimodal | Processes text, images, and audio at once | Service |
25 | On-premises / Cloud | How AI infrastructure is run | Business |
26 | Open-source model | Model with publicly released weights | Business |
27 | Token economics | Token-based API pricing | Business |
28 | Guardrails | Setting up AI safety controls | Business |
29 | AGI | Human-level general intelligence (not yet achieved) | Trends |
30 | AI governance | Framework for managing AI within an organization | Trends |
💡 Screenshot or print this table and pin it next to your monitor. You can check a term in 3 seconds during a meeting.
Step 7: Putting it to work — 3 scenarios you can use today
Scenario A: Answering a customer’s question in a sales meeting
Customer: "Should we do fine-tuning or RAG?"
Suggested answer framework:
"The two serve different purposes. RAG leaves the model as is and connects your internal documents as a search source, so you can adopt it quickly, within 2–4 weeks. Fine-tuning changes the model itself, so accuracy is higher, but data preparation takes 2–3 months. We recommend building an MVP with RAG first, then upgrading to fine-tuning once you’ve seen results."
Scenario B: Estimating AI adoption costs for an internal report
Enter into ChatGPT:
"Assume we handle 10,000 customer inquiries a month with the GPT-4o mini API. If each inquiry averages 500 input tokens and 300 output tokens, calculate the monthly API cost. Then build a comparison table showing the cost if we used GPT-4o instead."
Scenario C: Creating team training materials
Enter into ChatGPT:
"Create an AI terminology quiz for our team (marketing department, 10 people). Write a four-choice question for each of 7 terms: LLM, RAG, fine-tuning, hallucination, prompt engineering, token, and multimodal. Include the correct answers and explanations."
Wrap-Up Checklist
When today’s lesson is done, check the following:
You can explain in your own words the hierarchy of AI → ML → DL → Generative AI → LLM
Explain the difference between LLMs and SLMs in terms of cost
Explain the difference between RAG and fine-tuning using analogies
Understand the concept of tokens and roughly estimate API costs
Describe the risks of hallucination and 3 ways to prevent it
I printed the cheat sheet and posted it next to my monitor
💡 Final tip: Knowing a term and using it naturally in a meeting are two different things. This week, consciously use at least 3 terms in your meetings. Once you say them out loud, they become part of your own vocabulary.