Learning
Where to build the fundamentals, understand how AI really works, and practise offensive security, hands-on.
Foundations
How computers and code actually work, from logic gates up.
-
Build a working computer from first principles. You start with a single NAND gate and work upward through logic chips, an ALU and CPU, then write an assembler, a virtual machine, a compiler and a small operating system, until the machine you built can run Tetris.
Nothing explains what's under an exploit better: after this course, memory, registers and the call stack stop being abstractions.
-
CS50's Introduction to Programming with Python. Lectures and problem sets covering functions, conditionals, loops, exceptions, libraries, unit tests, file I/O, regular expressions and object-oriented programming, finishing with a project of your own.
Python is the glue of security work, from quick exploit scripts to parsing logs and automating recon.
AI
How the models at the top of the AI stack actually work, and where to learn to build one yourself.
How it works · Part 4 How a language model works Twelve steps from a sentence to the next token: tokens and embeddings, attention, the layers where the facts live, pretraining on trillions of words, the post-training that turns a model into an assistant, why inference is memory-bound, and what still goes wrong. With two things to try: watch a word decide what to look at, and generate text one token at a time with a temperature dial.The short version
It predicts the next piece of text
ChatGPT, Claude and Gemini are trained to answer one question, over and over: given this text, what comes next? Everything else is built on that.
Text becomes numbers
Words are split into tokens from a fixed vocabulary, and each token becomes a long list of numbers the model can do arithmetic on.
The transformer reads it in context
Dozens of layers let every token look back at the ones before it (attention) and pull in what's relevant, then the model scores every token in its vocabulary.
Pretraining
Trillions of tokens of text, each weight nudged towards the right next token, across tens of thousands of GPUs for months. Grammar, facts and reasoning patterns emerge.
Post-training
Example conversations, then reinforcement learning on answers people prefer and on problems that can be checked, which is where reasoning models come from.
Inference, one token at a time
Each new token is one pass through every layer, reading all the weights from memory, which is why the chips on this site are built around memory bandwidth.
-
The clearest explanations anywhere of how neural networks and language models work, from someone who has built them at the frontier. His videos come in two kinds: long talks with no code, for anyone who wants to understand what's happening inside ChatGPT, and hands-on lectures where he builds everything from scratch in Python, explaining every line as he types it.
Start with Deep Dive into LLMs like ChatGPT, three and a half hours that cover pretraining, tokenization, post-training, hallucinations and reinforcement learning without any code. Then Neural Networks: Zero to Hero, eight lectures that begin with backpropagation in about a hundred lines of Python (micrograd) and end with you building GPT and its tokenizer yourself.
Zero to Hero is the natural next step after CS50P: it assumes solid Python and school-level calculus. And knowing how a model reads its input is the foundation for AI security work, from prompt injection and jailbreaks to training-data leakage.
Offensive security
Hands-on labs, from single bugs to full networks.
-
A free, hands-on cybersecurity curriculum built at ASU. Challenges are grouped into dojos and run in a real Linux environment straight from the browser, starting with the command line and climbing through assembly, shellcode, reverse engineering, memory corruption and kernel exploitation. You earn belts as you clear modules.
-
Web application hacking built on real-world bugs and CVEs, with a strong focus on security code review. Exercises are grouped into badges by topic, such as JWT, deserialization, SSRF and authentication flaws, so you learn how a class of vulnerability works and then exploit it in a live lab.
-
HTB Academy is the structured side: guided modules and job-role paths such as penetration tester and bug bounty hunter, each with in-browser labs, leading to certifications like CPTS and CBBH.
HTB Labs is the open arena: vulnerable machines and challenges ranked by difficulty, seasonal competitions, and Pro Labs that simulate whole enterprise networks to work through end to end.
-
An offensive security academy with affordable courses and realistic, multi-machine labs you launch in the browser, including full Active Directory networks. It runs regular CTFs and is backed by a free Discord community focused on helping people break into the field.