Like nearly everybody, we had been impressed by the power of NotebookLM to generate podcasts: Two digital folks holding a dialogue. You may give it some hyperlinks, and it’ll generate a podcast primarily based on the hyperlinks. The podcasts had been attention-grabbing and fascinating. However additionally they had some limitations.
The issue with NotebookLM is that, when you may give it a immediate, it largely does what it’s going to do. It generates a podcast with two voices—one male, one feminine—and provides you little management over the outcome. There’s an non-compulsory immediate to customise the dialog, however that single immediate doesn’t can help you do a lot. Particularly, you’ll be able to’t inform it which subjects to debate or in what order to debate them. You may strive, nevertheless it gained’t hear. It additionally isn’t conversational, which is one thing of a shock now that we’ve all gotten used to chatting with AIs. You may’t inform it to iterate by saying “That was good, however please generate a brand new model altering these particulars” like you’ll be able to with ChatGPT or Gemini.
Study quicker. Dig deeper. See farther.
Can we do higher? Can we combine our information of books and know-how with AI’s means to summarize? We’ve argued (and can proceed to argue) that merely studying methods to use AI isn’t sufficient; that you must discover ways to do one thing with AI that’s higher than what the AI may do by itself. It’s essential combine synthetic intelligence with human intelligence. To see what that may seem like in apply, we constructed our personal toolchain that provides us way more management over the outcomes. It’s a multistage pipeline:
- We use AI to generate a abstract for every chapter of a e book, ensuring that every one the necessary subjects are coated.
- We use AI to assemble the chapter summaries right into a single abstract. This step basically offers us an prolonged define.
- We use AI to generate a two-person dialogue that turns into the podcast script.
- We edit the script by hand, once more ensuring that the summaries cowl the fitting subjects in the fitting order. That is additionally a chance to right errors and hallucinations.
- We use Google’s speech-to-text multispeaker API (nonetheless in preview) to generate a abstract podcast with two individuals.
Why are we specializing in summaries? Summaries curiosity us for a number of causes. First, let’s face it: Having two nonexistent folks talk about one thing you wrote is fascinating—particularly since they sound genuinely and excited. Listening to the voices of nonexistent cyberpeople talk about your work makes you’re feeling such as you’re dwelling in a sci-fi fantasy. Extra virtually: Generative AI is certainly good at summarization. There are few errors and virtually no outright hallucinations. Lastly, our customers need summarization. On O’Reilly Solutions, our clients incessantly ask for summaries: summarize this e book, summarize this chapter. They wish to discover the knowledge they want. They wish to discover out whether or not they really want to learn the e book—and in that case, what components. A abstract helps them do this whereas saving time. It lets them uncover shortly whether or not the e book might be useful, and does so higher than the again cowl copy or a blurb on Amazon.
With that in thoughts, we needed to assume via what probably the most helpful abstract can be for our members. Ought to there be a single speaker or two? When a single synthesized voice summarized the e book, my eyes (ears?) glazed over shortly. It was a lot simpler to take heed to a podcast-style abstract the place the digital individuals had been excited and enthusiastic, like those on NotebookLM, than to a lecture. The give and take of a dialogue, even when simulated, gave the podcasts vitality {that a} single speaker didn’t have.
How lengthy ought to the abstract be? That’s an necessary query. Sooner or later, the listener loses curiosity. We may feed a e book’s whole textual content right into a speech synthesis mannequin and get an audio model—we could but do this; it’s a product some folks need. However on the entire, we anticipate summaries to be minutes lengthy fairly than hours. I’d hear for 10 minutes, possibly 30 if it’s a subject or a speaker that I discover fascinating. However I’m notably impatient once I take heed to podcasts, and I don’t have a commute or different downtime for listening. Your preferences and your scenario could also be a lot totally different.
What precisely do listeners anticipate from these podcasts? Do customers anticipate to study, or do they solely wish to discover out whether or not the e book has what they’re in search of? That depends upon the subject. I can’t see somebody studying Go from a abstract—possibly extra to the purpose, I don’t see somebody who’s fluent in Go studying methods to program with AI. Summaries are helpful for presenting the important thing concepts introduced within the e book: For instance, the summaries of Cloud Native Go gave overview of how Go could possibly be used to handle the problems confronted by folks writing software program that runs within the cloud. However actually studying this materials requires examples, writing code, and practising—one thing that’s out of bounds in a medium that’s restricted to audio. I’ve heard AIs learn out supply code listings in Python; it’s terrible and ineffective. Studying is extra possible with a e book like Facilitating Software program Structure, which is extra about ideas and concepts than code. Somebody may come away from the dialogue with some helpful concepts and probably put them into apply. However once more, the podcast abstract is barely an outline. To get all the worth and element, you want the e book. In a latest article, Ethan Mollick writes, “Asking for a abstract will not be the identical as studying for your self. Asking AI to resolve an issue for you will not be an efficient method to study, even when it feels prefer it must be. To study one thing new, you will must do the studying and considering your self.”
One other distinction between the NotebookLM podcasts and ours could also be extra necessary. The podcasts we generated from our toolchain are all about six minutes lengthy. The podcasts generated by NotebookLM are within the 10- to 25-minute vary. The longer size may permit the NotebookLM podcasts to be extra detailed, however in actuality that’s not what occurs. Moderately than discussing the e book itself, NotebookLM tends to make use of the e book as a leaping off level for a broader dialogue. The O’Reilly-generated podcasts are extra directed. They comply with the e book’s construction as a result of we offered a plan, a top level view, for the AI to comply with. The digital podcasters nonetheless specific enthusiasm, nonetheless herald concepts from different sources, however they’re headed in a course. The longer NotebookLM podcasts, in distinction, can appear aimless, looping again round to select up concepts they’ve already coated. To me, at the least, that seems like an necessary level. Granted, utilizing the e book because the jumping-off level for a broader dialogue can be helpful, and there’s a steadiness that must be maintained. You don’t need it to really feel such as you’re listening to the desk of contents. However you additionally don’t need it to really feel unfocused. And if you would like a dialogue of a e book, you must get a dialogue of the e book.
None of those AI-generated podcasts are with out limitations. An AI-generated abstract isn’t good at detecting and reflecting on nuances within the authentic writing. With NotebookLM, that clearly wasn’t underneath our management. With our personal toolchain, we may actually edit the script to mirror no matter we wished, however the voices themselves weren’t underneath our management and wouldn’t essentially comply with the textual content’s lead. (It’s debatable that reflecting the nuances of a 250-page e book in a six-minute podcast is a dropping proposition.) Bias—a sort of implied nuance—is a much bigger problem. Our first experiments with NotebookLM tended to have the feminine voice asking the questions, with the male voice offering the solutions, although that appeared to enhance over time. Our toolchain gave us management, as a result of we offered the script. We gained’t declare that we had been unbiased—no one ought to make claims like that—however at the least we managed how our digital folks introduced themselves.
Our experiments are completed; it’s time to indicate you what we created. We’ve taken 5 books, generated quick podcasts summarizing every with each NotebookLM and our toolchain, and posted each units on oreilly.com. We’ll be including extra books in 2025. Hearken to them—see what works for you. And please tell us what you assume!
