Overview
Large language models become substantially more useful once they are understood not as omniscient databases, but as probabilistic token predictors surrounded by optional tools. The tutorial begins with ChatGPT’s core mechanics: users and assistants jointly build a context window, while pre-training supplies compressed, potentially outdated knowledge and post-training supplies the assistant persona. It then turns this mental model into practical selection rules. Routine writing and familiar, stable facts may suit a fast non-reasoning model; difficult math, code, and logic can justify a slower reasoning model; recent or obscure questions require web search; and precise calculations or data analysis call for a code interpreter. Uploaded documents place concrete source material into working context, enabling collaborative reading, research, and question answering. Multimodal systems extend the same interaction pattern to speech, images, and video, while specialized interfaces such as Claude Artifacts, Cursor, and NotebookLM turn model capabilities into apps, coding workflows, diagrams, or generated podcasts. Across every use case, the central warning remains unchanged: models can hallucinate facts, silently introduce assumptions, misuse tools, or report results inconsistently. Their output should therefore be treated as a strong first draft whose evidence, transcription, calculations, and code must be inspected in proportion to the stakes.
Sections
Core Concepts
Terms used to understand how LLM applications process information and extend model capabilities.
- Token: a small chunk used to represent text and, in multimodal systems, encoded audio or image content.
- Context window: the ordered token sequence serving as the model’s immediate working memory for the current interaction.
- Pre-training: the expensive stage that learns probabilistic patterns and broad knowledge by predicting tokens across internet-scale documents.
- Post-training: the stage that uses conversational examples to shape the pre-trained model into an assistant.
- Thinking model: a model additionally optimized through reinforcement learning to spend more tokens working through difficult reasoning problems.
- Tool use: an application-level mechanism through which special model outputs trigger external capabilities such as search or code execution and return results to context.
- Deep research: an extended workflow combining repeated internet search and reasoning to produce a source-linked report.
- Few-shot prompting: teaching a requested behavior by supplying several concrete input-output examples in addition to verbal instructions.
Capability and Workflow Tradeoffs
Explicit distinctions that determine which model, interface, or tool best fits a task.
- Non-thinking models respond quickly and suit routine tasks; thinking models take longer but can improve difficult math, coding, and logic results.
- Parameter knowledge is broad but hazy and dated; search supplies current external documents that the model can reference directly.
- Ordinary voice input converts speech to text before inference; advanced voice processes audio more natively and can respond with expressive audio.
- Browser chat is suitable for isolated prompts and lightweight artifacts; coding assistants such as Cursor retain repository context and can edit or execute across files.
- ChatGPT is presented as the broad default, Perplexity as a strong search interface, Claude Artifacts as useful for interactive apps and diagrams, and Grok voice as less restrictive entertainment.
Recommended Operating Practice
Concrete actions for using LLM applications more effectively and safely.
- Start a new chat whenever the topic changes and previous context is no longer relevant.
- Confirm the selected model and subscription tier before judging an application’s capability.
- Use a fast non-reasoning model first, then retry difficult math, code, or logic with a reasoning model.
- Enable web search explicitly for recent, changing, or obscure information and open the cited primary sources.
- Use a code interpreter for exact arithmetic, plotting, or data analysis; inspect generated code, inputs, assumptions, units, and displayed values.
- When uploading an image or screenshot, request transcription first and verify it before asking for analysis.
- Treat deep-research reports, medical interpretations, and product comparisons as first drafts for source checking and professional follow-up.
- Package frequently repeated tasks as custom instructions or few-shot custom GPTs.
- Use dictation for routine prompts and reserve typing for names, libraries, or technical terms that speech recognition may transcribe incorrectly.
Technical Mechanics and Examples
Implementation-level details described in the tutorial.
- The sample user request was described as 15 tokens, its haiku response as 19 tokens, and the formatted conversation as 42 tokens after special message markers were included.
- A model was illustrated as roughly a one-terabyte artifact containing approximately one trillion neural-network parameters, explicitly as a conceptual estimate rather than a universal specification.
- Search integration pauses token generation, executes a query, retrieves page text, inserts that material into the context window, and resumes generation over the augmented context.
- Interpreter integration lets the model emit a program plus control tokens, pauses model execution, runs the program, and returns its output as text for the model to explain.
- Claude Artifacts can render generated React components locally in the browser, enabling tools such as flashcard apps without a conventional database-backed application.
- Mermaid syntax was used within an Artifact to define and render a conceptual graph of an Adam Smith chapter.
- Cursor was described as supplying repository files to a remotely hosted model, allowing an agent-like composer to edit multiple files and execute commands.
- Audio and images can be encoded as quantized token sequences; the Transformer then models their token patterns while modality-specific encoders and decoders handle conversion.
- ChatGPT memory was described as a separate database of short user-related text entries that is prepended to later conversations and can be reviewed or edited.