Recruiters don’t always get me. So I did what anyone else would do -I generated talking heads to speak for me. I used Google’s NotebookLM to review my career and generate a 12 minute podcast. It was pretty good. Then I took a brief excerpt and split the audio into individual tracks and used generative AI to create faces to say it all for me. Here’s the how and why.
This article covers:
What is NotebookLM?
NotebookLM was released a few days ago as an AI research assistant, known within Google’s Gemini family as Project Tailwind. It’s a lot more than just another player among the growing marketplace of RAG tech, and actually makes me feel (for the first time in years) genuinely excited to be trying out and using a Google product.
While the industry’s buzz continues to be about OpenAI, Anthropic and Meta, Google’s Gemini, like some plucky billion dollar start up has been popping up now and then. Sometimes over-promising, but often being considered not the first choice but instead an inevitable one. The one upon which Google is betting a large acreage of the farm on, and which will in some form be as ubiquitous as Gmail. I say ‘in some form’ because I don’t anticipate a future where there’s going to be a good-at-everything product.
NotebookLM is a niche product which quietly showcases some amazing tech but is going to find the most traction among writers.
It’s an experimental AI-powered notebook designed to enhance note-taking and research. It allows users to upload documents like Google Docs, PDFs, text files copy-pasted notes and even URLs – grounding the AI in these sources for more accurate responses. A nice way of saying that it uses this information as its foundation to reduce the chances of hallucination. It’s not perfect, and in my experiments I found it was still prone to errors, but like Perplexity, each response includes inline citations, linking back to the original sources, making verification easier if something smells wrong.
So while acknowledging that it’s not always going to be right (and anecdotal reports in Reddit forums suggest some pretty wild mistakes) they also claim that this approach creates a more personalized and reliable experience for users.
Key features include generating summaries, answering questions, and brainstorming ideas based on uploaded content. In the first indication that this isn’t your standard Google play, it emphasizes privacy, assuring users that uploaded data isn’t used for training models. It also supports audio overviews of documents and aims to assist students and professionals in synthesizing information more effectively.
You can effectively use NotebookLM to organize your brainstorming sessions through the following steps:
-
- Capture Ideas: As you read sources, quickly pin interesting ideas as notes on the noteboard.
- Use the Chatbox: Ask questions related to your notes, and NotebookLM will summarize relevant information, providing citations for context.
- Suggest Related Ideas: After jotting down initial thoughts, select your notes and use the “suggest related ideas” feature to generate new concepts based on your input.
- Create Structured Documents: Select collected notes and ask NotebookLM to format them into outlines or study guides with quizzes, or briefing notes to streamline the organization process.
- Export Easily: Once organized, you can export your notes to Google Docs for further editing or sharing
- The Feature that’s going to get the most attention: and probably brought you here. I’ll get to that later.
It’s the ‘Suggest Related Ideas’ functionality which I think people are going to find the most useful, by grounding its suggestions in user-uploaded sources, unlike many AI tools like ChatGPT that draw from a broader, less specific dataset. When you write a paragraph and select it, NotebookLM generates a collection of related ideas based on that text, drawing from the sources you’ve provided.
This helps users make new connections and enhances creative thinking by suggesting themes or topics to explore further, ultimately streamlining the brainstorming process and fostering deeper understanding of the material. It allows NotebookLM to provide tailored, contextually relevant ideas based on the user’s own research materials, enhancing creativity and idea generation.
And by offering clickable citations for verification, it’s easier for users to fact-check and explore suggested concepts further. This approach promotes deeper engagement with the material compared to more generalized brainstorming tools available in other AI applications, where for example ChatGPT or Claude might lean into their perception of the outside world and go off into a hallucination. I suppose that could take you in an unexpected but helpful direction, but in my experience when a platform goes off the reservation it just turns into nonsense. Worse, it’s nonsense which is delivered as a confident mansplainer. Up until now, for research-with-citations I’ve usually turned to Perplexity and used the whichever AI is closest to hand to discuss the ideas. NotebookLM felt more like genuine brainstorming. A writing Copilot companion, if you will.
Talk to your PDFs, and then have them talk about you
Like I say, it’s not perfect for everything. I’m not going to ask it to help me code or generate UI ideas. Many platforms already allow you to ‘talk to your files’ and it does this well enough. It does not support direct analysis or summarization of Excel spreadsheets. Its capabilities are primarily focused on document formats like Google Docs, PDFs, and Google Slides. However, it can analyse data if the information from an Excel sheet is copied and pasted into a compatible format or if the data is included in a supported document type. For spreadsheet-specific summarization, other tools like GPT Excel are available.
The feature that’s getting the most attention, the hook that probably brought you to this article, is its ability to ingest all that material and generate an audio conversation between two apparently engaged, podcasters. In this ‘audio overview’ they’ll talk excitedly about the ideas they’ve discovered and essentially summarise the material, giving their own insights as to what they found to be the most interesting. The speed and neatness with which they do this is breath-taking, and is what prompted me to write this article, and in turn prompted me to take things further and create some AI generated video by putting faces to the voices.
I’m not usually a vain man, but as I wanted to check the platform’s abilities against a topic on which I’m a subject matter expert, I created a notebook and told the AI to generate an audio overview of this website’s homepage:
Not only was it so flattering that it made me want to go to the home of humblebragging and tell people about it, it also inspired me to build video content on top of it. Not so much because it was saying such nitce things, but what’s not getting enough attention is the ability to delivery summaries in a form which so closely mimics human interactions.
But it’s still a parlour trick
Because as a use case, it’s pretty narrow. I’m still going to use NotebookLM to gather my thoughts and extend my research, but the ability to conjure up a couple of humans to discuss it is not a necessity. The pauses, the intonation, the breathing are way ahead of what you’d get from an on-demand text to speech offering from, well, Google, and underlines the relevance of Gemini and Google Labs. It creates a compulsively sharable gimmick. I’ve no doubt that this narrow implementation (two hosts, podcast style intro, discussion of key points, outro) can be expanded but right now it’s an excellent showcase. And a powerful hook.
To illustrate this point, if you’ve listened to 12 minutes on the highly employable marvellousness that is me, you might see a formula emerging when hearing 3 minutes about the very different but oft compared:
In an interview on HardFork just a couple of days ago, Steven Johnson (author and New York Times writer) described his involvement in the project as Editorial Director at Google Labs. As well as this backstory, what I also found interesting was the interviewer’s observation that NotebookLM does not ‘feel’ like a Google product. Johnson put that down to the luxury of being an experiment at Google Labs, where they were given free reign over decisions like UI. And that worried me for a second. Not that their web scraping apparently bypasses Cloudflare’s recent filtering of AI bots to prevent them from hoovering sites -not mentioned in the interview, but I’m guessing this is due to the platform not visiting your site, but accessing the information Google already knows about it – but that as an experiment, even as a wildly impressive one, their niche might still leave them vulnerable to a common affliction among Google projects.
Take it further: how to put AI-generated faces to NotebookLM’s AI-generated voices
Ok, it’s another parlour trick, but it pulled together another distraction I’ve been meaning to play with and I’m hoping creates a better hook to drive traffic here.
-
- Take the audio and chop it down to less than a minute (I used the free open source editor, Audacity).
- Split that audio into two channels, one for each host and export as two files.
- Use Hedra to generate talking heads (I didn’t want to use my own so generated them both from my own prompts using Flux) and then applied their voices -those two audio files.
- Combine these two video files using Kapwing and add subtitles and branding.
I realise that I was a bit brief on the howto, but I’m trying not to get too distracted from other projects. If you’d like some specifics, please don’t heistate to contact me. Perhaps I’ll even do a separate post and video!
