Intro to LLMs in AI & Artificial Intelligence

Welcome to the age of Generative AI. Large Language Models (LLMs) like GPT-4 are the culmination of decades of deep learning research, trained on nearly all human-written knowledge.

These models are massive neural networks based on the Transformer architecture. They are called 'Large' because they have billions (or even trillions) of parameters—the internal mathematical 'knobs' that define their intelligence. This extreme scale is what allows them to demonstrate emergent reasoning capabilities.

Despite their apparent intelligence, LLMs are fundamentally just incredibly advanced probability engines. Their only real job is to look at a sequence of text and predict the most likely next 'Token'.

When you ask an LLM a question, it doesn't 'think' like a human. It calculates the statistical probability for every single possible next token in its vocabulary, picks the best one, adds it to the sequence, and repeats the process. A token is typically a word or a sub-word unit.

Creating an LLM requires two massive stages. The first is 'Pre-training'.

During pre-training, the model is fed essentially the entire internet—Wikipedia, Reddit, GitHub, books, and articles. It learns grammar, facts, coding syntax, and human logic. However, a purely pre-trained model isn't very useful; it just aggressively autocompletes text and can easily spout toxic or unhinged content.

The second stage is what turns the chaotic text generator into a helpful assistant. This is called 'Alignment', often achieved through RLHF (Reinforcement Learning from Human Feedback).

Humans rate the model's responses, teaching it to favor helpful, honest, and harmless answers. This is why ChatGPT refuses to tell you how to pick a lock, and why it writes in a polite, conversational tone. Alignment is what makes the raw intelligence actually usable.

Tokens are the fundamental currency of LLMs. Every model is constrained by a 'Context Window'—the maximum number of tokens it can hold in its short-term memory at any given time.

If you paste a 100-page document into a model with a small context window, it will literally 'forget' the first 50 pages because they get pushed out of its memory buffer. Managing this context window is the most critical skill for AI developers, especially when building advanced systems like RAG.

Welcome to the age of Generative AI. Large Language Models (LLMs) like GPT-4 are the culmination of decades of deep learning research, trained on nearly all human-written knowledge.

LLMs are massive Transformers. They are called 'Large' because they have billions (or trillions) of parameters—the internal 'knobs' that define their intelligence.

LLMs are probability engines. They look at a sequence of text and predict the most likely next 'Token'. A token is a word or a sub-word unit.

Checkpoint: At its most fundamental level, what is the core task an LLM is performing when it generates a response?

• →Actual conscious reasoning
• →Predicting the most likely next token in a sequence

Training happens in stages. 'Pre-training' teaches the model general knowledge. 'RLHF' (Alignment) teaches the model to be a helpful, safe assistant.

Tokens are the currency of LLMs. Most models have a 'Context Window'—the maximum number of tokens they can 'think' about at one time.

Checkpoint: If a model has a context window of 8,000 tokens and you provide a 10,000 token document, what will likely happen?

• →The model will get faster
• →The model will 'forget' or truncate the earliest 2,000 tokens

Intelligence scaled! You've successfully navigated the foundations of the world's most powerful reasoning engines. You're ready to master the art of Prompting.

Apply Real Temperature Scaling. Finish applying temperature scaling to a logit, controlling how sharp or flat the output distribution becomes.