[1hr Talk] Intro to Large Language Models
Hi everyone. So recently I gave a 30-minute talk on **large language models**.
Lumi / Features
AI Video Summary, Timestamped Q&A & Cross-Video Knowledge Library
Map out hour-long talks with timestamped Timeline Chapters, review structured Video Articles, and ask questions across a single video or your personal Knowledge Library — with clickable timestamp citations that take you straight to the source.
Examples & specifications
| Item | Details |
|---|---|
| Timeline Chapters | Structured chapter titles, concise takeaways, and clickable jump timestamps. |
| Video Article | Reorganizes spoken transcripts into a cohesive, skimmable reading guide for review and note-taking. |
| Knowledge Scope Snapshot | Freezes selected videos, caption tracks, text content, search indexes, and revision versions at the start of each Q&A turn. |
| Evidence Handle | Short-lived, opaque, permission- and version-bound credential for reading cited transcript context — never a public URL. |
| Time | Chapter summary |
|---|---|
| 0:21 | Definition of LLM as two files A Large Language Model is fundamentally defined as just two files: the model parameters and the run code. |
| 6:50 | Neural network predicts next word The core function of an LLM is to predict the next word in a sequence, a simple objective that leads to complex capabilities. |
| 16:37 | Fine-tuning on Q&A data transforms a base model into an assistant. Fine-tuning a base model on high-quality question-and-answer data transforms it from a text predictor into a helpful assistant. |
Andrej Karpathy · Intro to Large Language Models · 2026-06-27 · summary-v3 / article-v4
Workflow
- Open a video with usable captions and save it to your personal Knowledge Library when you want to keep it.
- Generate Timeline Chapters for a quick structural overview or a Video Article for continuous reading.
- Select either the current video or a multi-video collection scope and ask your question.
- Click any cited timestamp or transcript excerpt in the answer to jump the player to that exact moment.
Questions & answers
- When should I use Timeline Chapters, a Video Article, or Video Q&A?
- Each view fits a different stage of learning: Timeline Chapters give you a bird's-eye outline with timestamps so you can spot and jump to key sections in seconds; a Video Article turns spoken delivery into well-structured prose ideal for post-watch review and note-taking; Video Q&A answers targeted questions by retrieving passages from the active transcript or your saved Knowledge Library (rather than running an ungrounded open-web search). Output depth scales with transcript coverage and your chosen model, and the sample on this page illustrates real production output structure.
- Why does cross-video Q&A freeze a Knowledge Scope Snapshot on every turn?
- When querying across multiple videos, changing a caption track or editing your collection mid-stream could cause retrieved citations to drift. A Knowledge Scope Snapshot locks the exact set of selected videos, transcript choices, text bodies, search indexes, and revision versions the moment you submit a question, guaranteeing consistent retrieval and citation resolution throughout that turn. Once a newly saved video has its transcript and search index ready, it will automatically be included in your next question's snapshot.
- Do clickable timestamp citations guarantee that an AI answer is 100% error-free?
- Timestamps are built to make verification effortless, not to replace critical reading. Under the hood, an Evidence Handle is a short-lived, permission- and version-bound token that lets Lumi load the surrounding transcript lines for a cited passage (it is not a publicly shareable link). Because any LLM can still compress nuance or misread a speaker's caveat, you should click the timestamp citation to inspect the surrounding captions or replay the original clip whenever numbers, causal claims, or critical conclusions matter.
- Are my personal Knowledge Library and Q&A conversations shared with other users?
- Never. Your saved video collections, personal notes, and Q&A threads belong strictly to your private account scope. The Shared AI Artifact pool only caches standardized, non-personal outputs and explicitly excludes video chat and library Q&A. Every citation lookup is gated by an authenticated, version-checked Evidence Handle so clients cannot enumerate or guess storage addresses. Multi-video retrieval runs solely across the indexed videos you explicitly select in your own library.