Andrej Karpathy — Intro to LLMsY Combinator — How to get startup ideas
Your YouTube video
Paste a YouTube link. Nothing to upload.
←YouTube video
Make a research report
Turn YouTube videos into a research report
Paste several links. One question, the evidence for it, and the places the speakers disagree.
No link handy? Try one:
Free to try, no card. Your YouTube video opens in Ancher.
Caption track with timingsKey frames from the videoDescription and linked sources
Kept in parts
Every one of these stays separately addressable instead of collapsing into one block of text.
The questionKey findingsEvidenceCounterpoints
Research report
A structured answer to a question, with the evidence and the counter-evidence attached.
One video is one person's argument, so a report built from a single source is a summary wearing a report's clothes. Ancher works across every video you add, groups claims by the question they answer, keeps the disagreements between speakers rather than averaging them, and marks which findings rest on one voice alone.
The honest version
A report built from one video is just that video
The value of a report is that it survives its sources being wrong, which requires more than one of them. Ancher is built to work across a collection, so the report can say how many speakers support a finding, which one dissents, and where the field is actually unsettled — none of which is visible from inside a single video.
One second of one video, at 31:20, read two ways.Illustrative. The figure is invented; the gap it falls into is not.
...most people cite the older figure here, but the updated number is actually and it changes the conclusion.
Cited on screen, contradicts two other talks — not readThe specific never reaches the draft
Research report
Cited on screen, contradicts two other talks
1.8x, not 3x
from 31:20
Lands in the draft, with its timecode
- Claims are grouped by the question they answer, so two speakers addressing the same point land next to each other.
- Disagreements are preserved and named, because a report that resolves every conflict has deleted the useful part.
- Any finding resting on a single speaker is flagged as such rather than presented with the same weight as a consensus.
What comes out
Research report, section by section
A structured answer to a question, with the evidence and the counter-evidence attached.
The question
One question the set of videos actually addresses, stated narrowly enough that a finding can answer it.
Key findings
Claims grouped by question, each showing how many videos support it and which ones.
Evidence
The specific figures and examples behind each finding, with the video and timecode they came from.
Counterpoints
Direct contradictions between speakers, named on both sides rather than smoothed into a middle position.
What to check next
What the set does not answer, plus the sources the speakers themselves cite as the place to look.
Multi-video research report prompt
The link on its own gets a summary. This is the instruction that produces the 5 sections above, in that order, with the rules that keep them honest. Paste it with your YouTube videos — in Ancher, or in whatever assistant you already use.
Across every video in this collection, produce a research report. (1) State the single question the set addresses. (2) Group claims by question; for each finding, say how many videos support it and name them. (3) Give the specific figures behind each finding with video and timecode. (4) List every direct contradiction between speakers, naming both sides — do not resolve them. (5) Mark any finding resting on one speaker alone. (6) End with what the set does not answer. Never state a claim without a source.
The detail, if you want it
How each part of a YouTube video becomes a section
| From the YouTube video | Into | Why |
|---|---|---|
| Caption track with timings | Key findings | Claims are matched across videos by what they assert, so the same point made by three speakers is counted once and attributed three times. |
| Key frames from the video | Evidence | The figure supporting a claim is usually on a slide, so the evidence column is only fillable if frames were read. |
| Caption track with timings | Counterpoints | Contradictions surface where two speakers make incompatible assertions about the same quantity, which needs both transcripts side by side. |
| Description and linked sources | What to check next | Where several creators cite the same source, that source is almost certainly worth reading before you go further. |
| Title, channel, duration, date | The question | Publication dates bound the report — a finding from a 2022 talk contradicting a 2025 one may just be out of date rather than wrong. |
Everything Ancher keeps from a YouTube video
Caption track with timings
Published and auto-generated captions both come through, and every line keeps the second it was spoken.
Key frames from the video
Frames are sampled and scored, so a number shown on a slide survives even though it was never said out loud.
Description and linked sources
The creator's own cited links come along, so a claim can be traced past the video.
Likes and comment counts
Reception is a weak signal about correctness and a strong one about which claims got attention.
Title, channel, duration, date
Basic provenance, kept so the finished work can attribute the video properly.
The four steps
- Add the YouTube video. Paste a YouTube URL. Ancher pulls the caption track with its timings, samples key frames from the video itself, and keeps the description links and engagement counts on the same source.
- Nothing gets flattened. Caption track with timings, key frames from the video, description and linked sources, likes and comment counts, title, channel, duration, date stay separately addressable instead of collapsing into one block of text.
- Ask for a research report. Ancher fills the question, key findings, evidence, counterpoints, what to check next using the mapping above, not a generic summary pass.
- Check it before you use it. Every specific keeps a link back to where it came from in the YouTube video, so a claim can be verified without reopening the source.
Where this goes wrong
- A report from a single video is a summary. If the findings section cannot say how many sources support a claim, the collection was too small.
- Videos on a topic tend to cite each other, which produces the appearance of consensus from one original source. Check whether independent speakers are actually independent.
- Speakers are selected for being watchable, not for being representative. A report built from YouTube reflects who makes videos about a subject, which is not the same as who works on it.
Questions people ask
How many videos before this is worth doing?
Below about five, reading them yourself is faster and you will retain more. The value comes from cross-source structure — agreement counts, contradictions, gaps — and structure needs enough sources for a pattern to be real rather than coincidental.
Can I mix videos with articles and PDFs?
Yes, and you should. A claim that appears in three videos and one peer-reviewed paper is a different kind of finding from one that appears in three videos alone, and the report can only make that distinction if both are in the collection.
Will it invent citations?
Every claim is bound to a source and a timecode inside your own collection, so a claim with no source has nothing to attach to. This is the specific reason to synthesize from material you added rather than from open-web recall.
Bring your own YouTube videos
Everything you save lives in one workspace, so the research report is built from your sources — not from a model's memory of the internet.
Open Ancher →Not a YouTube video? The same research report also comes from PDF, X thread, LinkedIn post.