How LLMs Work From N-grams to Modern Architectures
Date: 26 September 2026 | Posted by: voska89
How LLMs Work From N-grams to Modern Architectures
Published 9/2026
MP4 | Video: h264, 1920x1080 | Audio: AAC, 44.1 KHz, 2 Ch
Language: English | Duration: 1h 30m | Size: 406.94 MB
No math or coding needed. Understand ChatGPT by following what each model solved, from N-grams to Gated DeltaNet
What you'll learn
Explain in your own words that a language model is a machine that outputs the probability of the next word, and that text generation is just repeating that game
Trace N-grams, RNNs, LSTMs, Attention, Transformers, Mamba, and Gated DeltaNet by explaining which problem each new technique solved.
Compute N-gram probabilities and Self-Attention (dot product, softmax, weighted average) by hand with small numbers
Explain the meaning of terms such as embedding, gate, KV cache, state space model, MoE, and attention sink by tying each to an analogy
Compare the strengths and weaknesses of Transformer and Mamba (precise search versus cheap long-term summarization) and explain why hybrid architectures emerged
Read and interpret the idea behind the hybrid architecture of Qwen3-Next / 3.5 / 3.8, which combines three Gated DeltaNet layers with one Gated Attention layer
Requirements
No prior knowledge of math or programming is required. Addition, multiplication, and a feel for proportions are enough to follow to the end
Having used a conversational AI such as ChatGPT at least once will make the course examples easier to picture
If you want to try the hands-on exercises introduced in the course, a browser environment that can open Google Colab is handy (not required)
Description
What happens inside ChatGPT, told through 100 years of history
What exactly is a large language model (LLM) like ChatGPT doing inside? This course aims to let you explain the basic principles in your own words, even if you have no background in programming or math.
The approach is a little unusual. Instead of jumping straight to the latest models, we walk through roughly 100 years of language model history in order, from Markov chains in 1913 to the hybrid architectures of 2026. Every technique was born to solve a problem the previous technique had. When you follow each problem together with its solution, even intimidating cutting-edge techniques fall into place naturally.
Figures, analogies, and a minimum of math
Throughout the course we use a single analogy: "how a reader remembers a book." Looking only at the last few words, summarizing into a single memo, looking back at sticky notes, running a full-text search every time, and returning to a smart summary memo. We understand each model as a different way of remembering.
Math is kept to a minimum, and whenever it is truly necessary we always explain in the order figure, analogy, then equation. The course is built so that the story still holds together if you skip the equations. The only arithmetic you need is addition, multiplication, and a feel for proportions. For Self-Attention, we trace a single hand calculation with small numbers so the mechanism sinks in. Each section ends with a 3-line summary and a quiz, and once you can answer it you are ready to move on.
What you will learn
The starting point: a language model is a "next-word guessing game"
N-grams: the era of counting to predict, and its two limits
Embeddings, RNNs, and LSTM gates: turning words into numbers and remembering as you read
seq2seq and Attention: the idea of looking back at sticky notes with a weighted average
Transformer: query, key, value, Self-Attention, positional encoding, multi-head, FFN, and the weakness called the KV cache
Linear Attention, S4, Mamba: handling long context with a fixed-size state
Gated DeltaNet and Gated Attention: the hybrid of three summary-memo layers and one full-text-search layer, attention sinks, MoE, and a reading of Qwen3-Next / 3.5 / 3.8
By the end of the course, terms like Attention, KV cache, state space model, and MoE that show up in news and paper titles will no longer be intimidating, and when a new model appears you will be able to read it with the question "what did this solve compared to the one before?" The recommended pace is one section per day.
Who this course is for
People who use AI tools such as ChatGPT but do not know what is happening inside
People who feel uneasy about terms like Attention, Transformer, and KV cache that appear in LLM articles and news
Beginners who want to understand the basic principles of AI systematically without a math or programming background
People who want to be able to read the specs and paper titles of the latest models in the context of how the technology evolved
Homepage
https://www.udemy.com/course/how-llms-work-from-n-grams-to-qwen38-en/DOWNLOAD LINK
No Password - Links are Interchangeable
See full article