How long would it take to watch every Andrej Karpathy lecture vs. ask one question across all of them?

Watching every lecture Andrej Karpathy has published on YouTube would take roughly 40 hours — a full working week of continuous, attentive viewing. Asking one question across all of that material, once it is indexed in a SageTube Expert, takes about ten seconds, and the answer arrives with a timestamp pointing at the exact moment in the exact lecture where he addresses it. That is the whole trade in one sentence: a week of your time versus ten seconds, for the common case where you don’t need the entire course — you need the three minutes of it that answer today’s question.

What does 40 hours actually consist of?#

Karpathy’s freely available lecture material breaks down into three main bodies of work (runtimes are approximate):

• Neural Networks: Zero to Hero — the from-scratch series on his own channel: micrograd (~2.5 h), the five makemore installments (~8 h combined), “Let’s build GPT” (~2 h), the GPT tokenizer deep-dive (~2 h), and the four-hour GPT-2 reproduction. Together: roughly 18–19 hours across nine videos.
• Stanford CS231n, Winter 2016 — the convolutional neural networks course he co-taught, with lectures mostly in the 70–80 minute range: roughly 17 hours.
• The standalone LLM talks — “Intro to Large Language Models” (~1 h), “Deep Dive into LLMs like ChatGPT” (~3.5 h), “How I use LLMs” (~2 h): a bit over 6 hours.

Add conference talks, Tesla AI Day segments, and multi-hour podcast appearances, and the realistic total for “everything Karpathy has said on camera” clears 50 hours. Nobody watches that end-to-end and retains it. People watch some of it, remember a fraction, and then — months later — need one specific thing he said, somewhere, in one of those videos.

Why is re-finding one moment so hard?#

Because the length of the material is the obstacle, a problem we cover in depth in the time-scarcity pillar. Say you remember Karpathy explaining why a particular weight-initialization choice tames exploding activations. Was that makemore part 3? A CS231n lecture? Twenty minutes into the GPT-2 video? YouTube’s search ranks whole videos by title and metadata — it cannot take you to a sentence spoken at minute 43. Your options today are scrubbing timelines, opening transcript panels one video at a time, or re-watching. All three fail at 40 hours of scale.

Our own index data shows how concentrated this problem is in exactly this kind of content. As of July 2026, SageTube’s production database tracks 46,486 videos, of which 21,469 are transcribed — about 4,994 hours of searchable speech. Long-form videos (20 minutes and up, the lecture-and-podcast format) are only about 20% of those transcribed videos, but they account for roughly 70% of the hours — 3,512 of the 4,994. The videos that are hardest to re-search by hand are precisely the ones holding most of the recorded knowledge.

How do you put all the lectures behind one question?#

You build a SageTube Expert from them. An Expert is a knowledge base that attaches any number of YouTube videos — individually, or whole channels at once — through a many-to-many relation in our data model (app/Models/Rag.php:102 for videos, :82 for channels). One Expert can hold the Zero to Hero series, the CS231n uploads, and the standalone talks side by side. (To be clear: this is something you assemble yourself in a few minutes — SageTube doesn’t ship a pre-built Karpathy Expert.)

From there, three things happen to every video:

1. Transcription, captions first. SageTube fetches the video’s existing captions when they’re available — the fast path, on the order of seconds per video (app/Jobs/ProcessTranscript.php:79). Karpathy’s lectures have good captions, so a full index of them builds quickly. Videos without usable captions fall back to audio extraction and speech-to-text transcription instead of being skipped.
2. Chunking that keeps time. Each transcript is split into passages, and every passage retains its timestamp_start and timestamp_end mapped back to the video timeline (app/Services/ChunkManager.php:288). The index doesn’t just know that something was said — it knows when.
3. Answers with receipts. When you ask a question, the relevant passages are retrieved from across every video in the Expert, and the evidence the answer actually used is resolved into full citations — video, channel, timestamp — attached to the response (app/Services/CitationHydrator.php:18).

So the question “what did Karpathy say about initializing the final layer to make early losses sane?” stops being a 40-hour archaeology project. It becomes one query whose answer tells you it came from makemore part 3 at a specific minute — and the timestamp is clickable, so verifying the answer costs you thirty seconds of watching the actual source, not blind trust in a chatbot’s memory.

Isn’t watching the lectures the point?#

For a first pass — absolutely. Zero to Hero is built to be watched and coded along with; no retrieval system replaces that. The ten-second query wins on every pass after the first: reviewing before an interview, checking a half-remembered detail while debugging your own training run, or settling whether he actually said that thing you’re about to quote. Watching is for learning the material; asking is for using it later. The 40 hours and the ten seconds are not competitors — they’re the before and after of the same investment.

You can try question-answering over public video knowledge bases on our Explore page, or build your own Expert from the channels you keep going back to.

Every product claim in this post is tied to the SageTube codebase as of July 2026: multi-video and multi-channel Experts (app/Models/Rag.php:102, :82), the captions-first transcript pipeline with audio fallback (app/Jobs/ProcessTranscript.php:79), timestamp-preserving chunking (app/Services/ChunkManager.php:288), and per-answer citation hydration (app/Services/CitationHydrator.php:18). Index statistics were queried from the production database on 2026-07-19. Lecture runtimes are approximations based on the published playlists.