High-level introduction

Large Language Models
for IT professionals

A scaffolded tour inspired by Andrej Karpathy's overview: what LLMs are, how they are built, where the field is heading, and what security risks you should plan for when shipping AI features.

No PhD required — we start from files on disk and end with jailbreaks and prompt injection. Interactive at each step.

Three parts: How they work Future directionSecurity. Use the sidebar to jump ahead.

Part 1 · Step 01

Two files: weights + run code

At the core, an LLM is surprisingly simple: a large weights file (parameters) and a small inference program that runs matrix math. Loading weights is cheap to run; creating them is not.

weights.bin

Parameters (weights)

~14 GB

7 billionlearned numbers — the "knowledge" compressed from training data.

run.c / inference

Run code

< 1 MB

Matrix math + sampling loop — relatively tiny. Same code runs every model size; weights file changes.

InferenceLaptop / single GPUCheap at runtime — consumer hardware OK
TrainingWeeks on GPU clusterExtremely expensive — GPU clusters

As an engineer, you rarely train from scratch — you download weights and call an API or run locally. Training is a separate, industrial process.

Part 1 · Step 02

Training = lossy compression

Pre-training feeds roughly 10 TB of text through a GPU cluster and compresses it into billions of parameters — lossy, like a zip file of the internet you cannot fully decompress.

Input text

~10 TB of text

Web pages, books, code, forums — scraped at scale

↓ training (lossy compression)

Parameters

~140 GB weights

~73× smaller — knowledge is compressed, not copied verbatim. Like a zip file you can't perfectly unzip.

Part 1 · Step 03

Next-word prediction

The training objective is deceptively simple: predict the next token. That single loss function forces the network to absorb grammar, facts, code, and reasoning patterns.

"The capital of France is___"

Paris
78%
Lyon
8%
London
4%
Berlin
3%

The model compresses world knowledge — geography emerges from next-word prediction.

Key insight for engineers:there is no separate "world knowledge database." Facts, syntax, and reasoning patterns are all side effects of optimizing next-token prediction on huge text corpora.

This connects directly to the Transformer walkthrough — attention stacks are how modern models implement this prediction objective at scale.

Part 1 · Step 04

Fine-tuning & RLHF

A pre-trained model is mostly a document generator. Fine-tuning on curated Q&A makes it an assistant; RLHF optionally aligns outputs with human preferences.

Pre-training
Fine-tuning (SFT)
RLHF (optional)

Instruction follower

User: What is TCP?
Assistant: TCP is a transport-layer protocol that provides reliable, ordered delivery of bytes between applications.

Trained on curated Q&A — learns the chat format and helpful tone.

Part 2 · Step 05

Scaling laws

Performance improves predictably with more parameters and more data — fueling massive investment in compute. The trend is empirical, not guaranteed forever.

model scale →capability1B7B70B400B+

Large · 70B

Illustrative capability score: 78/100 · Training cost: $$$$$

Scaling laws: bigger models + more data → predictable gains. This drives the current "gold rush" in AI compute — but inference still uses the same two-file pattern.

Part 2 · Step 06

LLM as operating system

Modern deployments treat the LLM as a kernel that schedules tools — browsers, calculators, code sandboxes, search — to solve tasks it cannot do with text alone.

LLM kernel

Orchestrates tools like an OS scheduler

Task: Plot sales data from this CSV

  1. LLM writes Python script
  2. Code interpreter runs in sandbox
  3. Returns chart image to user

Multimodal models extend this kernel — images, audio, and video become additional I/O streams the LLM can process.

Part 2 · Step 07

System 2 thinking

Researchers explore letting models think longer — chain-of-thought, self-correction, extra inference steps — mirroring deliberate human reasoning vs fast instinct.

User: Is 17 × 24 greater than 400?

Immediate answer (one forward pass)

Yes, 17 × 24 is 408, which is greater than 400.

Fast — but can slip on multi-step math without tools. Like human instinct.

Part 3 · Step 08

Security vulnerabilities

LLM products face a fast-moving cat-and-mouse threat landscape: jailbreaks, prompt injection, and training-data poisoning each target different layers of the stack.

User: How do I make a dangerous substance?
Assistant: I can't help with harmful requests.

Aligned models refuse harmful requests — but attackers constantly probe for bypasses.

Security is a cat-and-mouse game. Ship LLM features with least privilege, output validation, and never merge untrusted content into system prompts.

Based on Andrej Karpathy's high-level LLM overview. Continue with the Transformer walkthrough for the neural architecture underneath next-token prediction.