Home > Videos

How I use LLMs

By Andrej Karpathy

🤖 AI Summary - Exploring Large Language Models (LLMs) and AI Tools

Introduction to Large Language Models (LLMs)

• The video continues a general audience series on large language models (LLMs), particularly ChatGPT.
• Previous videos covered how LLMs work under the hood; this one focuses on practical applications.

Evolution of LLMs

• ChatGPT’s Launch (2022): First major LLM deployment by OpenAI, allowing users to interact with AI through a text-based interface.
• 2025 AI Ecosystem: The space has evolved, with competitors like:
  • Google Gemini
  • Meta’s AI assistant
  • Microsoft Copilot
  • Anthropic’s Claude
  • Elon Musk’s Grok
  • Chinese (DeepSeek) and French (Mistral) AI models.

Where to Track AI Model Performance

• Leaderboard Websites:
  • Chatbot Arena (for ranking AI models).
  • Scale AI’s Eval Leaderboard (comparison of model performances on different tasks).

How LLMs Work

• Inputs and outputs are broken into tokens.
• LLMs predict the next token based on patterns learned during training.
• Limitations & Knowledge Cutoff – LLMs rely on pre-existing data and may lack real-time updates.

AI-Powered Internet Search

• AI can visit multiple web pages, extract relevant information, and summarize results.
• Search Token Mechanism: Some AI models use a search token to trigger live web searches.
• Comparison of AI Search Capabilities – Different AI models handle search differently (e.g., ChatGPT vs. Claude).

Deep Research Feature in AI

• A new feature allowing AI to spend more time on complex research.
• Combines internet search and reasoning to generate detailed reports.

LLMs and File Processing

• Uploading Documents for Analysis – AI can analyze PDFs, scientific papers, and books.
• Interactive Learning: AI can provide summaries, historical context, and explanations.

AI for Data Analysis and Code Generation

• LLMs and Coding Assistance – AI can generate code but must be verified.
• AI for Advanced Data Analysis (ADA) – Creating graphs, trend predictions, and statistical modeling.
• Claude’s “Artifacts” Feature – Interactive coding with real-time JavaScript and React execution.

AI-Assisted Coding in VS Code & Cursor

• Cursor as an AI Coding Assistant – AI-driven project modifications in VS Code.
• ”Vibe Coding” Approach – AI automates repetitive coding tasks.

AI and Multimodal Capabilities

• LLMs now process images, audio, and video.
• Example: Analyzing a nutrition label from a supplement mix and categorizing ingredients.

AI for Medical and Scientific Analysis

• Blood Test Interpretation: AI can transcribe blood test results, provide insights, and highlight concerns.
• AI in Mathematics and Science: Recognizing and solving complex equations from an image.

AI for Everyday Object Analysis

• Example: Uploading a toothpaste ingredient list to analyze which chemicals are essential vs. unnecessary.

AI Image and Video Generation

• DALL·E 3 – Generates high-quality images based on text prompts.
• AI-generated videos – Tools like Runway, Sora, and Pika compete in video generation.

Custom GPTs and AI Assistants

• Users can build custom GPTs to automate tasks.
• Examples:
  • Korean Vocabulary Extractor – Turns sentences into flashcards.
  • AI Subtitle Reader – Extracts and translates subtitles from screenshots.

Timestamps for Key Sections

Introduction to Large Language Models (LLMs)

• 00:00:00 - Introduction – Overview of LLMs and ChatGPT’s evolution.
• 00:05:30 - AI Ecosystem in 2025 – Google Gemini, Meta AI, Claude, Grok, etc.
• 00:12:45 - How LLMs Work – Tokenization, prediction, and training phases.

AI-Powered Search and Research

• 00:32:10 - AI for Web Search – How AI fetches real-time information.
• 00:40:00 - Deep Research Tools – AI performing in-depth searches and summarization.
• 00:50:20 - File Analysis – Uploading PDFs for AI-powered summarization.

AI for Data Science & Coding

• 01:06:00 - AI-Assisted Coding – Using AI for software development.
• 01:14:45 - Claude’s Artifacts Feature – Generating code and flashcards.
• 01:23:30 - Vibe Coding in VS Code – AI-assisted programming and auto-coding.

AI for Multimodal Understanding

• 01:41:30 - AI Image Processing – Uploading images for AI analysis.
• 01:45:00 - AI for Blood Test Interpretation – Uploading medical results for AI insights.
• 01:50:00 - AI for Meme Explanation – Decoding humor and social media memes.

AI-Generated Content and Personalization

• 02:00:00 - AI Image Generation – Creating visuals using DALL·E.
• 02:05:30 - ChatGPT Memory Feature – How AI remembers user preferences.
• 02:09:30 - Custom GPTs for Language Learning – Creating AI-powered language tools.