LLM

What it is

LLM (Large Language Model) is a type of artificial intelligence model trained on massive amounts of text to:

Understand natural language - Comprehend questions, instructions, and context
Generate coherent text - Responses, code, documentation
Reason - Follow logic, solve problems step by step
Translate and summarize - Between languages and formats

LLMs are the “brain” behind ChatGPT, Claude, Gemini, and Copilot.

What else it is called

“Large Language Model” in full
“Modelo de lenguaje grande” (in Spanish)
“Foundation model” (broader term)

How it works

┌─────────────────────────────────────────────────────────────┐
│                    LLM ARCHITECTURE                          │
├─────────────────────────────────────────────────────────────┤
│                                                              │
│   TRAINING (once, plus later fine-tuning)                   │
│   ┌──────────────────────────────────────────────────────┐  │
│   │  Internet, books, code, conversations                │  │
│   │  → Trillions of text tokens                          │  │
│   │  → Weeks/months on thousands of GPUs                 │  │
│   └──────────────────────────────────────────────────────┘  │
│                           │                                  │
│                           ▼                                  │
│   TRAINED MODEL                                             │
│   ┌──────────────────────────────────────────────────────┐  │
│   │  Parameters: from billions to hundreds of billions   │  │
│   │  (or more)                                           │  │
│   │  Knowledge "frozen" in weights                       │  │
│   └──────────────────────────────────────────────────────┘  │
│                           │                                  │
│                           ▼                                  │
│   INFERENCE (each use)                                      │
│   ┌──────────────────────────────────────────────────────┐  │
│   │  User: "Explain what Kubernetes is"                  │  │
│   │  LLM: Predicts next token → next → ...               │  │
│   │  Response: "Kubernetes is a platform..."             │  │
│   └──────────────────────────────────────────────────────┘  │
│                                                              │
└─────────────────────────────────────────────────────────────┘

Major LLM families

Family	Company	Weights	Strength
GPT	OpenAI	Closed (API)	General reasoning
Claude	Anthropic	Closed (API)	Code, safety
Gemini	Google	Closed (API)	Multimodal, search
Llama	Meta	Open (up to 405B in Llama 3.1)	Self-hosting
Mistral	Mistral AI	Open and closed	Efficiency, European company

Versions change every few months and most closed-model vendors don’t disclose parameter counts. Check each vendor’s documentation for the current model.

LLM vs SLM

Aspect	LLM (Large)	SLM (Small)
Parameters	Tens or hundreds of billions	From hundreds of millions up to ~10-15B
Hardware	Powerful GPUs, cloud	Laptop, phone, modest GPU
Cost	Higher (pay per token via API)	Lower, or just hardware when local
Capability	Complex, varied tasks	Specific tasks
Examples	GPT, Claude, Gemini	Phi-3, Llama 3.2 1B/3B

Practical Example: LLM in Development

Before (without LLM)

Developer: Search Stack Overflow
           Read documentation
           Copy/adapt code
           Debug manually

Now (with LLM)

Developer: "Implement JWT authentication in Express"

LLM: Generates a draft of code + tests + documentation
     Explains each part
     Suggests security improvements

Developer: Reviews, tests, and adjusts before merging

Time savings vary widely depending on the task and the team’s experience; human review remains essential.

Limitations

What LLMs can NOT do well

Limitation	Description
Hallucinations	Make up information that sounds real
Limited knowledge	Training cutoff date
Complex math	Errors in long calculations (without tools)
Private data	Cannot access your database on their own
Real-time	Don’t know the current date or events without tools

Mitigations

┌────────────────────────────────────────┐
│   HOW TO IMPROVE LLMs                   │
├────────────────────────────────────────┤
│                                         │
│   + RAG (Retrieval)                    │
│     → Access to current documents      │
│                                         │
│   + Tools/Function Calling             │
│     → Execute code, APIs               │
│                                         │
│   + Fine-tuning                        │
│     → Specialize in your domain        │
│                                         │
│   + Agents                             │
│     → Plan and execute tasks           │
│                                         │
└────────────────────────────────────────┘

Agentic AI - LLMs that can act autonomously
SLM - Small models for local devices
MCP - Protocol to connect LLMs to tools
Prompt Injection - Security attacks against LLMs

Remember: LLMs are powerful tools but not perfect. Always verify critical information and use your professional judgment.

Pronunciation