Learn 80% of NotebookLM in Under 13 Minutes!
Video by Jeff Su on YouTube · 13 min · uploaded December 2024
Notes and analysis by BrainTube, written from the video. Not affiliated with Jeff Su. Watch on YouTube
In short
Jeff Su's tutorial teaches NotebookLM through real notebooks: health reports, equipment manuals, tax, recruiting, projects and earnings analysis. The practical tips are accurate and useful, especially selecting sources deliberately and saving answers as notes so they are not lost. Two things need care: his size comparison with ChatGPT and Claude mixes up retrieval and context windows, and a cited answer is only as reliable as the source it cites.
The argument
The video argues NotebookLM beats general chatbots when three things are true: you have low tolerance for hallucination, your material is scattered across formats (documents, slides, video, audio), and you need to turn it into something coherent quickly. Because answers are grounded in the sources you add, with inline citations, it is the right tool for questions about your own documents. It is less creative than Gemini, so he uses it to gather and check, then hands the output to Gemini or Claude for the final draft. Its weaknesses are mostly workflow: chat answers vanish unless saved, and every answer draws on every selected source.
Key ideas, with timestamps
- 1:03
A mixed-source notebook
His demo notebook holds three annual health checkup reports, a PDF of Tools of Titans, a YouTube video on uric acid and an Andrew Huberman video on fasting. The source guide gives a short summary of each dense report, and clicking a key topic expands it using all selected sources.
- 2:15
Every selected source is in play
Everything in chat considers all selected sources, even when a key topic came from just one. To keep a source out of an answer you must deselect it. He returns to this at 7:57 for recruiting: select only the current candidate's documents, or details from other candidates leak in.
- 3:02
Trends, then a citation check
Asked for the top 10 health trends across three reports, it flags elevated uric acid over three years. He then asks what to do about it and clicks the inline citation, which opens the video transcript at the line where the doctor recommends potassium citrate.
- 3:46
Save to note or lose it
He got burned by this: chat answers disappear when you reload the notebook unless you click save to note. He relays Google's explanation that NotebookLM does not train on uploads or conversations, so unsaved chat is not kept.
- 4:56
It says when sources are silent
Asked about deload weeks, which none of his sources cover, it says so and mentions related topics. He uses this to explain the trade-off with Gemini: NotebookLM is tuned to hallucinate less and is less creative as a result.
- 5:35
Notes become sources
You can select several saved notes, or all of them, and convert them into a standalone source. He suggests opening that source to copy the combined text for use elsewhere.
- 5:59
Four notebook patterns
Focused retrieval over equipment manuals found by searching the product name plus "user manual" filetype PDF; a tax notebook with tax codes and audit reports; a recruiting notebook with rubrics and interview notes; and one notebook per work project with meeting transcripts and old recaps. Sites that block NotebookLM can be added by pasting their text.
- 9:45
Earnings analysis and audio
His earnings notebook combines company reports with analyst articles saved as Google Docs, which can be resynced when they change. He customises an audio overview to focus on how one company's earnings affect competitors, for a listener with no technical background.
- 11:36
Limits and source quality
He cites about 25 million words per notebook against roughly 500,000 for Gemini, 100,000 for Claude and 64,000 for ChatGPT, and a cap of 20 sources per notebook that you can work around by merging files. His last point is that output quality depends on source quality.
What it gets right
The source-selection tip is the most important one in the video and it is often missed. Grounded tools still answer from everything in scope, so a notebook with two candidates' CVs will blend them. Treating selection as part of the prompt is the right habit.
Save to note is also a real trap for new users, and the notes-to-source move is a smart workaround: it turns scattered good answers into one document the notebook can build on.
The working split he ends with is sensible. Use NotebookLM to collect and check facts from your sources, then use a general model for the creative final draft. That plays to each tool's design instead of fighting it.
What it leaves out
- The size comparison is not like for like. The figures for Gemini, Claude and ChatGPT describe how much text a model holds in one conversation. NotebookLM's figure describes how much material it can search across, retrieving relevant parts per question. Bigger storage does not mean it reads everything at once.
- The limits look inconsistent. 25 million words with a cap of 20 sources implies over a million words per source. The arithmetic fits 50 sources at 500,000 words each better. Either way, limits change, so check current numbers rather than the video's December 2024 ones.
- A citation is not verification. Clicking through to the transcript shows that the doctor in the video said potassium citrate works. It does not show the claim is right. The same goes for being told it is fine to keep fasting 36 hours a week based on blood work: grounded answers inherit the quality of the sources, which he notes for blog posts but not for health advice.
- Notebooks are silos. Each notebook only knows its own sources. Knowledge in your health notebook is invisible from your project notebook, and none of it is reachable from ChatGPT or Claude without copying it over. His notes-to-source tip exists partly to work around this.
- Features have moved. The Notebook guide layout he shows has changed since the upload. The ideas transfer; the clicks may not.
How this applies to memory across AI tools
NotebookLM is a good model of source-grounded answers: you add the material, it answers from it with citations. Its limit is that the material stays inside each notebook and inside one app. If you also work in Claude, ChatGPT or Cursor, you end up uploading the same videos and PDFs in several places.
The alternative is one store of sources that any assistant can search. BrainTube works this way: it saves YouTube videos with their transcripts, articles, PDFs and podcasts into one compiled knowledge base and exposes it as an MCP server, so Claude, ChatGPT, Cursor and other MCP clients can search it. For other options, see NotebookLM alternatives for YouTube, and for the general idea, chatting with YouTube videos using AI.
Who should watch it
Watch it if you are new to NotebookLM or have tried it once and not stuck with it. The demo up to 5:45 covers the core habits: source selection, citations and saving notes. The use cases from 5:59 to 11:32 are the best part for knowledge workers, especially the project notebook pattern at 8:08. Skip it if you already use NotebookLM daily, since most tips will be familiar, and keep in mind the interface has changed since December 2024. There are short plugs for his toolkit and newsletter around 5:52 and 9:36.
Frequently asked
- Why do my NotebookLM chat answers disappear?
- Chat answers are not kept when you reload a notebook unless you save them. Click save to note on any answer you want to keep. You can later combine several saved notes into a new source for the notebook.
- How do I make NotebookLM answer from only one source?
- Deselect the other sources in the source list before you ask. NotebookLM answers from every selected source, so leaving extra sources selected can mix in unrelated information, for example details from another job candidate.
- Is NotebookLM better than ChatGPT for my own documents?
- For questions about a set of your own documents, NotebookLM is designed to answer from those sources with citations and to say when they do not cover a topic. It is less suited to creative writing, so many people draft the final output in a general chatbot.
Keep videos like this one searchable
BrainTube saves YouTube videos with their transcripts into one compiled knowledge base that Claude, ChatGPT, Cursor and other MCP clients can search.
Go deeper
More video notes
- Give Me 13 Minutes. I'll Teach You 80% of NotebookLM · Linking Your Thinking with Nick Milo
- AI Agent Memory Architecture: Structured SQL, Semantic Vector Search & Memory Bank · Google Cloud Tech
- Build an Obsidian BACKEND Not a Second Brain! · Eric Michaud
- Complete Guide to building an AI Second Brain with Claude · Ray Fu
Are you the creator and want this page changed or removed? Contact us and we will act within 48 hours.