A Very Gentle Introduction to Large Language Models without the Hype

Purpose: The article aims to provide insights into how chatbots like ChatGPT function without requiring a technical background.
Coverage: It will explain core concepts related to Large Language Models (LLMs) and use metaphors to clarify these concepts.
Approach: Begin with basic definitions and gradually build up to complex ideas, avoiding technical jargon where possible.

Definition: A computer system performing behaviours perceived as intelligent by humans.
- Example: AI in video games often uses simple code logic to keep players engaged.
Clarification: The term 'intelligent' is problematic because there is no consensus on its definition, yet it suffices in practice.
- Key Insight: AI is not magical; it is explainable.

Definition: A process where behaviour is created by using data to form a model and executing this model.
Reason for Use: Traditional methods (such as manually entering if-then-else statements) are often insufficient for complex phenomena, such as human language.
Model: A simplification of a phenomenon, analogous to a model car representing a real car without entirely replacing it.

Description: A learning model based on associations similar to how neurons interact in the human brain.
- History: Concepts for neural networks date back to the 1940s, but large-scale applications began only in 2017 with sufficient computing power.
Metaphor: Comparing neural networks to electrical circuitry, where signals (electricity) flow between components (neurons).
- Example: In a self-driving car, sensors relay information to actuators (such as brakes and steering), just as signals travel through a neural network.
- Issues in Direct Wiring: All sensors initially connected to all actuators lead to chaotic responses.
- An adjustment process begins, utilising resistors to control flow and gates to enhance decision-making.

Definition: A subset of machine learning that allows more complex calculations (adding and multiplying) within circuits.

Goal: For example, in the sequence “Once upon a ___,” the model must guess what word fits (here, it predicts ‘time’).
Probability: Language models assess the likelihood of a word appearing based on its preceding words, which form the basis of their outputs.
- Notation: Example represented as the probability $P( ext{word} ext{n} | ext{word}_{1: ext{n-1}})$ indicates what will follow a sequence of words.
Typographical comparison: Visualise as a typewriter, where inputs relate to corresponding outputs through weighted connections.

Function: Compress input data from numerous sensors (e.g., words) into a smaller output set (e.g., 256 values).
Example cases: If 'king' reduces to [0.1, 0.0, …, 0.4] and 'queen' reduces similarly, this reflects their relationship.

Role: Receives the encoded input (256 values) and translates it back into a larger word output from its understanding, activating the most probable word output from sensors.
- Process shown: One encoding leads to the highest energy input to a singular output.

Definition: Allows language models to learn appropriate outputs based on provided inputs without requiring separate datasets for evaluation.
Masked Language Models: A type of model that assumes some words in a sentence are missing and must be filled.
Auto-regressive nature of models: Generative models predict the next word based on prior words, with several implications for context and coherence based on previously generated words.

Purpose: Introduced by the paper Attention is All You Need in 2017, it revolutionised LLMs by improving the encoding process through a mechanism called self-attention.
- Aspects of Transformers: Generative, pre-trained, and designed to model language through rich interaction with context.
- Basic Operations: Ensuring the interrelatedness of words helps in predicting missing words during processing.

Function: Enables the model to discern which words in a sentence relate to one another, enriching prediction accuracy.
Mathematical operations: Uses dot product similarity which governs how closely words relate based on numerical encoding.
- Self-attention matrices record and reflect these relationships for further processing.

Functionality: LLMs predict logical word sequences based on context and training examples, enabling human-like text generation.
Training Source: Massive datasets scraped from the internet expose LLMs to diverse writing styles and topics, enabling them to develop a broad understanding.

Result: The tendency for LLMs to provide average responses from observed training data, which harnesses large-scale interactions across contexts.
Caution: It is critical to understand that LLMs emulate learned patterns rather than possess innate intelligence or awareness.

Training Content: Models learn from all internet data, including biases and misinformation.
Responses: LLMs generate responses based on the consistency of patterns regardless of moral or factual accuracy, so verification is essential.
Error Margins: Accumulation of errors in auto-regressive models can lead to compounding inaccuracies in outputs.
Response Quality: A well-crafted query leads to improved specificity in output.
Memory Concept: LLMs don't retain conversation history; each input is treated independently, despite appearing conversational.

Reinforcement Learning with Human Feedback (RLHF): This added method refines outputs for better alignment with human expectations.
Instruction Tuning: This process corrects models when they misinterpret prompts, enhancing performance.
Overall Insight: By understanding LLMs’ workings and limitations, one can effectively utilise their capabilities for generating helpful responses without misattributing human-like qualities to their behaviour.