An Oracle Speaks

ChatGPT Was Just the Beginning

Last year, I was at a dinner in New York where Steven Johnson talked up his new project with VCs and product people. Given his resume, you'd expect Johnson was there to introduce an innovation story destined for the bestseller list. He's written thirteen books to high acclaim, including Where Good Ideas Come From, How We Got to Now, and a favorite, Everything Bad Is Good for You. An intellectual alchemist, he draws insights from science, technology, and history, revealing how they converge to drive progress.

This time, Johnson wasn't talking about his next page-turner. He'd been cooking up NotebookLM, an AI research assistant in partnership with Google. Unlike products such as ChatGPT, which train on a massive dataset, this one processes only the personal data you share with it.

After a limited release in July 2023, NotebookLM became widely available for free last December. Its new true-to-life voice feature now has people raving about it.

How does it work?

Before we get to the buzz about its voice, a bit more about the platform. As Johnson describes, NotebookLM is distinct from other models in this way:

"The core idea behind the current version of the product is what we call 'source-grounded AI': you define a set of documents that are important to your work—called 'sources' in the NotebookLM parlance—and from that point on, you can have an open-ended conversation with the language model where its answers will be 'grounded' in the information you've selected."

Sources are files you feed it—like PDFs, text docs, and web pages—that make the model an expert on the material. Once digested, it returns insights from the content, including citations and quotes.

NotebookLM has a whole bag of tricks: FAQs, study guides, tables of contents, timelines, and briefing docs. It also gives you a few suggested questions to get the juices flowing. Google swears your data stays private, so for now, you don’t need to worry about the model training on it.

Say you're prepping for a big presentation. You've got piles of business reports, competitor analysis, and cultural intel. So, instead of spending days or weeks sifting through it all, you feed it to Notebook LM. In the time it takes to pour a cup of coffee, the AI creates documents with key takeaways and talking points tailored to what you’re looking for. The starter docs are a portal into chat, where you prompt for more specific information needed for your pitch.

I've been using it for months, and it's remarkable. It's turned years of research—my books, highlights, quotes, and notes—into a living, personalized knowledge base. I uploaded a PDF of my book to demonstrate what it can do. To see what people are buzzing about, check out the upper right corner of the window.

The audio overview is where it gets pretty weird. It's a podcast with two male and female voices riffing on the book's content. They're downright enthusiastic about Perspective Agents (which is nice). They hit the main points, bounce ideas off each other, and even crack a few jokes about how crazy media has become. In conversation, they interrupt, finish each other's sentences, and laugh at the absurdity of it all.

And the speed? Mind-boggling. Two-hundred-sixty pages uploaded in nineteen seconds flat. An executive summary and document options popped up like magic. Three minutes and twenty-nine seconds later, boom, a full-fledged podcast. Out comes eight minutes of back-and-forth banter. You can give the cut a thumbs up or down. The first take was a bit rough, so I hit the 'redo' button a couple of times. The duo eventually did an astonishing job of describing the book. Give it a listen here.

Now, any document—a book, a slide show, even a resume—can become conversation fodder for a podcast. By now, you know the twist. These aren't humans; you have AIs talking as if they were. They're fairly convincing replicas of real people.

An audio sludge pile?

Steven Levy, a veteran technology journalist and author of Hackers and In the Plex, has covered digital immersion into our lives for decades. In a Wired review last December, he cautioned that NotebookLM might be too good. He said about the risk:

"Users of NotebookLM, who simply want to get a good job done quickly, might not take the time to do that hard work of thinking. They might not even bother to pore through the research materials themselves. Why take the time when your AI buddy has gone through the material much more closely than you and has already reached some nifty conclusions about it?"

Now, add AI podcasts to how we might think or depict thought. Synthetic dialogue is a double-edged sword. On one side, it can democratize content creation, offer education tailored to every learning style, and unravel complex ideas with lightning speed. On the other hand, it can create a tidal wave of audio sludge, a vehicle for disinformation, and a medium where human voices get drowned out in a sea of fabricated chatter.

For now, early adopters are electrified. It has a feel of ChatGPT’s early days, the sense that something big is going on here.

As a podcasting side hustle, NotebookLM has potential beyond the research assistance Johnson and Google designed it for. Imagine AI pundits battling it out over the latest political issue, theoretical physicists explaining the mysteries of the universe like they're chatting over a beer, or your half-baked ideas transformed into a coherent audio drama.

While this may not be achievable today, the platform will likely evolve rapidly. Soon, it could allow for completely tailored productions, including selecting the voice, tone of conversation, gender, number of participants, and edits. (For reference, note OpenAI’s voice updates this week.)

What’s available now from the Johnson/Google collaboration is excellent for its intended use. In the case of my book, Perspective Agents in NotebookLM is a far better knowledge partner than my OpenAI agent that trained on it. The audio aid alone prompts authors to rethink how their work will be consumed and reviewed. All in all, it feels like a breakthrough or breakdown, depending on your view.

For kicks, I created a notebook trained on this 1,102-word essay. You can listen to the six-minute podcast made from it here. The result made me wonder if my time was well spent writing this and yours reading it.

You might grasp the story better from two AIs talking about it than from the words I put on the page.

---

I'm happy to share a link if you want access to the Perspective Agents on NotebookLM. Hit me up direct, and I'll send it along.