# 103 Things Everyone Should Know About LLMs

Source: https://www.youtube.com/watch?v=Ym-iMJ-sds0
Recap page: https://rapidrecap.app/video/Ym-iMJ-sds0
Generated: 2025-11-11T19:07:42.697+00:00

---
## Quick Overview

The video explains how Large Language Models (LLMs) like ChatGPT and Gemini process context through their context window, utilize specialized tools like Web Search, Image Generation, and Canvas, and how features like System Prompts, User Memory, and Projects help maintain focus and consistency across complex, long-term tasks, ultimately enabling more accurate and tailored outputs.

**Key Points:**
- LLMs rely on a context window, which includes the current conversation and potentially stored memory, to maintain focus, but this window has a finite limit (e.g., 400,000 tokens for GPT-5).
- Advanced models integrate specialized tools like Web Search (for up-to-date data), Image Generation (like DALL-E 3 or Gemini's Imagen), and Canvas for visual prototyping.
- System Prompts allow users to define a specific role, output format, and instructions for the model to follow consistently across a project, such as creating a 'personalized Japan travel planner AI'.
- User Memory allows ChatGPT to retain crucial personal details (like occupation or preferences) across different chats, making responses more personalized without constant repetition of instructions.
- The 'Projects' feature organizes related chats, files, and custom instructions into dedicated workspaces, preventing context loss for long-term work.
- The Chain of Thought/Reason and Act process forces the model to break down complex problems into logical steps (Input -> Analyze -> Infer -> Decide -> Output), leading to more robust reasoning and accurate solutions.
- The video demonstrates superior performance in complex tasks (like MRR modeling or logic puzzles) when using advanced reasoning features compared to standard, direct prompting.

![Screenshot at 00:24: 00:The core concept of understanding LLMs is broken down into two areas: understanding 'Under the hood' mechanics and utilizing the model to its 'Full Potential'.](https://ss.rapidrecap.app/screens/Ym-iMJ-sds0/00-00-24.png)

**Context:** This video serves as an advanced guide detailing various features and techniques for maximizing the utility and performance of modern Large Language Models (LLMs) like ChatGPT (specifically referencing GPT-5 capabilities) and Google's Gemini. It moves beyond basic prompting to cover structural tools (Projects, System Prompts), memory management, specialized features (Web Search, Image Generation, Canvas), and advanced reasoning techniques (Chain of Thought).

## Detailed Analysis

The video comprehensively details techniques for leveraging advanced LLM capabilities beyond simple questioning. It starts by explaining that LLM intelligence hinges on the 'Ability to Reason and Think' (02:35), which is facilitated by breaking down problems using the Input-Analyze-Infer-Decide-Output process, known as Chain of Thought or Reason and Act (03:02). The context window limits how much information the model retains, causing older information to 'fade' (05:27), necessitating Project organization for long-term focus (09:40). Users can tailor responses using Personalization settings, including selecting specific 'ChatGPT personality' styles (e.g., Nerd, Listener) and setting 'Custom Instructions' that apply across all future interactions in that context (08:35). The models are inherently multi-modal (06:28), supporting file uploads (like PDFs for analysis, 06:34) and image generation (04:17). Specific tools like Web Search (12:11) and Canvas (13:48) extend functionality by providing real-time data access or a collaborative virtual workspace, respectively. Furthermore, the introduction of OpenAI's Codex (15:08) highlights specialized code-generation capabilities, and Gemini's CLI shows integration with terminal execution, requiring explicit user confirmation for security (15:17). The overall message is that mastering LLMs involves understanding their context limitations and actively using features like Projects, Memory, and specialized tools to guide the model toward high-quality, consistent results.

### LLM Context Management

- Context window limits information retention
- Conversation history fades as it grows longer
- User Memory stores personal preferences across chats
- Projects organize related files and instructions for long-term tasks

### Advanced Reasoning Techniques

- Chain of Thought (Reason and Act) breaks down problems logically (Input -> Output)
- Essential for complex math and logic problems, exemplified by the Knights and Knaves puzzle (03:24)

### Tool Integration and Multimodality

- Tools like Web Search provide real-time data
- Image Generation (DALL-E 3, Gemini's Imagen) create visuals from prompts
- Canvas offers a virtual workspace for collaborative editing
- File analysis processes uploaded documents (e.g., PDF summary, 06:38)

### Model Personalization

- Custom Instructions define desired style, tone, and format (e.g., Gen Z style, 08:55)
- ChatGPT Personalities (e.g., Robot, Listener) set default response styles
- System Prompts within Projects provide project-specific constraints (09:59)

### Code Generation and Execution

- Codex provides code support
- Gemini CLI allows secure, sandboxed execution of code (e.g., npm install) after user confirmation (15:17)
- Canvas supports prototyping UI code (React example, 13:51)

![Screenshot at 00:01: 17:Progression of OpenAI models from GPT-2 through GPT-5 indicated by evolving logo designs.](https://ss.rapidrecap.app/screens/Ym-iMJ-sds0/00-00-01.png)
![Screenshot at 00:04: 40:Demonstration of the 'Think' mode in the prompt bar \(Google Gemini interface\), indicating the model will process steps before giving a final answer.](https://ss.rapidrecap.app/screens/Ym-iMJ-sds0/00-00-04.png)
![Screenshot at 00:49: 50:Visual diagram illustrating the concept of context window limitations, where older information falls out of the model's active memory as the conversation progresses.](https://ss.rapidrecap.app/screens/Ym-iMJ-sds0/00-00-49.png)
![Screenshot at 01:32: 15:Comparison slide highlighting the strengths of Claude \(coding, structured reasoning\), ChatGPT \(creative ideation, ecosystem\), and Gemini \(research, data access, long-form content\).](https://ss.rapidrecap.app/screens/Ym-iMJ-sds0/00-01-32.png)
![Screenshot at 02:40: 00:Flowchart explaining the Chain of Thought process: Input -\> Analyze -\> Infer -\> Decide -\> Output, mirroring human reasoning.](https://ss.rapidrecap.app/screens/Ym-iMJ-sds0/00-02-40.png)
![Screenshot at 08:17: 00:Demonstration of setting User Memory in ChatGPT settings to retain facts \(like preferred coding language\) across future chats.](https://ss.rapidrecap.app/screens/Ym-iMJ-sds0/00-08-17.png)
![Screenshot at 10:13: 00:Example of a detailed System Prompt being created within a Project for a 'personalized Japan travel planner AI', specifying tone, focus, and goal.](https://ss.rapidrecap.app/screens/Ym-iMJ-sds0/00-10-13.png)
![Screenshot at 11:56: 00:Visual explanation of how Web Search works by navigating search results, sub-links, and extracting key details, contrasting with simpler models that only scratch the surface.](https://ss.rapidrecap.app/screens/Ym-iMJ-sds0/00-11-56.png)
![Screenshot at 13:51: 00:Canvas feature in action, showing a generated React login page code snippet with options to preview or edit.](https://ss.rapidrecap.app/screens/Ym-iMJ-sds0/00-13-51.png)