74 lines
No EOL
3.6 KiB
Markdown
74 lines
No EOL
3.6 KiB
Markdown
---
|
|
type: youtube
|
|
source_url: https://www.youtube.com/watch?v=EWvNQjAaOHw
|
|
retrieved: 2026-06-26
|
|
channel: "Andrej Karpathy"
|
|
title: "How I use LLMs"
|
|
duration_sec: 7871
|
|
has_transcript: false
|
|
tags: [karpathy, llm, tutorial, llm-fundamentals, tool-integration, multimodal, voice-interaction, custom-gpts, context-window, thinking-models, learning-resource]
|
|
---
|
|
|
|
# How I use LLMs
|
|
|
|
**Channel:** Andrej Karpathy
|
|
**URL:** https://www.youtube.com/watch?v=EWvNQjAaOHw
|
|
**Duration:** 2:11:11 (7871 sec)
|
|
**Views:** ~2.5M (as of 2026-06-26)
|
|
**Published:** ~1 year before retrieval (mid-2025)
|
|
**Language:** English
|
|
|
|
## Context
|
|
|
|
Shared by k9ert in OME-Gruppe Topic 1345 ("Ich, der DAU und sein Bot LOGBUCH") on 2026-06-26. Recommended as a tutorial/learning resource for understanding AI/LLMs, especially for people wanting to learn about AI.
|
|
|
|
## Summary
|
|
|
|
*Reconstructed from YouTube metadata and web search summaries — no transcript available.*
|
|
|
|
Andrej Karpathy's "How I use LLMs" is a practical, example-driven exploration of how to leverage Large Language Models in daily life and work. The video covers:
|
|
|
|
### LLM Fundamentals
|
|
- LLMs as AI systems trained on vast datasets
|
|
- Pre-training for knowledge acquisition vs. post-training (fine-tuning) for alignment and specific tasks
|
|
- The concept of a "knowledge cutoff" — models are only aware of information up to their training date
|
|
- "Thinking models" that use reinforcement learning for improved accuracy
|
|
|
|
### Core Interactions
|
|
- Text-based interaction: prompts and AI-generated responses
|
|
- Managing the "context window" (LLM's working memory)
|
|
- Starting fresh chats for new topics to maintain accuracy and responsiveness
|
|
|
|
### Tool Integration
|
|
- Internet search to access up-to-date information beyond training data
|
|
- Python interpreter for code execution and calculations
|
|
- ChatGPT's Advanced Data Analysis for figures and plots
|
|
- File uploads (documents, images — e.g., deciphering a blood test report)
|
|
|
|
### Multimodal Capabilities
|
|
- Audio interaction: speech-to-text input and text-to-speech output
|
|
- Karpathy notes over half of his LLM interactions are voice-based for efficiency
|
|
- Distinction between "fake audio" (TTS) and "true audio" (direct audio processing)
|
|
- Image input including OCR and visual data analysis
|
|
- Video input — querying models about real-world objects via camera
|
|
|
|
### Personalization and Efficiency
|
|
- "Custom instructions" to fine-tune LLM behavior and tone
|
|
- "Custom GPTs" — specialized mini-LLMs for specific tasks (language learning, coding)
|
|
- LLM memory — models remembering previous conversations within a chat
|
|
|
|
## Why This Video Matters
|
|
|
|
Karpathy is one of the most respected voices in AI education. His ability to explain complex topics clearly makes this video an excellent starting point for anyone wanting to understand LLMs practically — not just theoretically. The video bridges the gap between "what is an LLM" and "how do I actually use one effectively."
|
|
|
|
## Related Wiki Pages
|
|
|
|
- `[[wiki/concepts/llm/llm-knowledge-base.md]]` — Karpathy's wiki pattern (this repo's foundation)
|
|
- `[[wiki/tutorials/karpathy-how-i-use-llms.md]]` — Tutorial wiki page for this video
|
|
- `[[wiki/concepts/llm/llm-model-catalog.md]]` — Catalog of LLMs mentioned across the wiki
|
|
- `[[wiki/concepts/llm/llm-sycophancy-confabulation.md]]` — LLM behavior pitfalls (relevant for practical usage)
|
|
|
|
## Notes
|
|
|
|
- `has_transcript: false` — No transcript was fetched for this ingest. Content summary reconstructed from YouTube metadata and web search results.
|
|
- Karpathy also has a longer companion video "Deep Dive into LLMs like ChatGPT" (3:31:24, 7.7M views) at https://www.youtube.com/watch?v=7xTGNNLPyMI — referenced as related content. |