Google NotebookLM made a whole podcast from my notes in just some clicks

Gemini, Google’s basic function AI assistant (and household of multimodal AI fashions) may get all the eye, however NotebookLM, an AI-powered productiveness instrument Google first launched in 2023, is by far the corporate’s most attention-grabbing utility of generative AI. Not as a result of it is amazingly succesful, however as a result of it is intentionally restricted, centered even, not on an internet-sized dataset, however as an alternative sources you present. It could actually create summaries and reply questions on no matter materials you convey, and as of Sept. 2024, additionally create podcasts primarily based on these sources too.

Google designed the instrument to assist customers work with massive portions of knowledge by making their analysis and notes conversational — you may ask NotebookLM particular questions concerning the sources you present — and letting them generate new materials from the stuff you’ve uploaded. NotebookLM can create issues like summaries, FAQs, research guides, and now, structured audio clips Google calls Audio Overviews.

We have coated a few of the potential makes use of of NotebookLM earlier than, however Audio Overviews go a step additional than simply being the audio model of textual content summaries by injecting human-like pauses and digressions into an AI-generated dialog between two synthetic voices. The entire characteristic works surprisingly effectively, but additionally can not help however illustrate AI’s capacity to sound good, however not essentially be it.

NotebookLM can flip internet pages, PDFs, and docs into a brief podcast

The instrument can generate textual content and audio primarily based on quite a lot of sources

In case you’ve by no means used NotebookLM earlier than, the very first thing the instrument prompts you to do is add sources, whether or not it is a Google Doc, Google Slide, PDF, YouTube video, or unpaywalled internet web page.

From there, NotebookLM creates a abstract of every of your sources, and a click on away within the Pocket book Information, an general abstract of your total pocket book. You’ll be able to add your personal notes in a bit within the heart of the web page and, after all, ask questions on what you’ve got uploaded to NotebookLM within the chat field, too. Audio Overviews, Google’s latest characteristic, can be found within the high proper nook of the Pocket book Information.

At the moment, Audio Overviews can solely work with English textual content, and solely produce English audio. The format can be fairly set: it all the time generates a podcast with two hosts, and is all the time a giant image overview of the topic your sources cowl. These podcasts are comparatively quick, shifting in size primarily based in your sources (5 sources obtained me an Overview that was over 10 minutes lengthy), and take a while to generate. In case you’ve gotten used to the moment responses of Gemini Liveor ChatGPT Advanced Voice Mode, get used to ready. It took me round 4 minutes to generate an Audio Overview, however it can seemingly differ relying on the variety of sources NotebookLM has to synthesize.

Sooner or later, according to Wired, Google plans on making these Audio Overviews extra customizable, like letting you modify “the podcast’s size or subject of focus.” Like different Google AI merchandise, new languages may even seemingly be added over time. It is comparatively experimental for now, very similar to the remainder of NotebookLM.

Pocket book LM’s Audio Overviews make you’re feeling such as you perceive

AI is shockingly good at mimicking the rhythms of a podcast

I made a decision to make my pattern pocket book on the computational concept of thoughts or CTM, the philosophical concept that the human thoughts works like a pc or primarily is a pc by a special identify. It is one of many core beliefs that animates AI analysis — if the human thoughts already works like a pc, then it must be doable to create a man-made one with a pc. It additionally appeared greater than just a little ironic to make AI hosts focus on it on a pretend podcast.

As my sources, I added internet pages from Stanford College’s philosophy division, a abstract of a bibliography from Oxford, and the Wikipedia web page for my chosen topic. I additionally wrote a word outlining my intentions for this challenge: I needed to discover CTM, but additionally illuminate what appeared to me to be fairly apparent shortcomings. I am undecided if Audio Overviews even takes notes into consideration, however I believed if I might skew the content material in a sure route, this is able to be the best way.

Three to 4 minutes later, I obtained a 13-minute Audio Overview that coated the important thing beliefs of anybody who follows CTM, criticisms of the speculation, and different areas of analysis which have grown out of it. The podcast NotebookLM created was surprisingly lifelike. The pretend hosts pause, chortle, and share anecdotes (one notable one a couple of canine neither host owns opened the “episode”) similar to you may hear on a traditional chat podcast.

The ultimate product additionally had quite a lot of unusual mispronunciations of phrases. One of many hosts mentioned “I” and “S” moderately than the phrase “is” and a minimum of as soon as a bunch mentioned “nervous laughter” as an alternative of simply…laughing. These errors appear minor within the grand scheme of issues, however they undoubtedly break the phantasm that that is someway an actual podcast hosted by actual individuals.

NotebookLM might nonetheless be helpful, however Audio Overviews really feel like a bauble

Podcasts aren’t the most effective studying instrument as is

I used to be undeniably impressed with what NotebookLM generates, however I do not discover it to be a very compelling characteristic by itself. Nothing concerning the podcast I heard felt extra detailed than the summaries or different textual content paperwork that NotebookLM might already create. A few of the extra informal language the pretend podcast hosts made the subjects really feel extra approachable, nevertheless it largely appeared like a lot ado about nothing.

There are downsides inherent within the present implementation of generative AI, however one of many extra insidious ones is the tacit acceptance that sounding like an individual is nice sufficient. There’s an assumption that talking casually, and profusely, as a human can, is sufficient to paper over any AI inaccuracy. That an AI speaking like a human is a couple of steps away from being human. It is onerous to not see the Audio Overviews and voice interfaces like ChatGPT’s Superior Voice Mode, because the extra convincing proof for that very same argument.

That is additional sophisticated by the truth that podcasts are solely actually pretty much as good as their analysis and writing, and NotebookLM would not make for the most effective podcast producer. I can take heed to The Day by day, The New York Occasions’ every day information podcast, and declare that I am knowledgeable, and I most likely can be compared to not listening to it in any respect, however I do not assume that is the identical factor as gaining deep data of the subjects mentioned. If something, it will most likely give me false confidence that I do know one thing, once I actually do not. I believe NotebookLM’s Audio Overviews may very well be responsible of the identical factor.

NotebookLM

Google’s NotebookLM is an experimental product that makes an attempt to synthesize massive portions of knowledge into research guides, chatbots, and even AI-generated podcasts, which might be simpler to work with.

It is good that Google is providing choices to customers — some individuals will be taught higher by listening to a podcast than studying paragraphs of textual content — and that NotebookLM is structured extra as a instrument the place issues like Audio Overviews are a place to begin for producing data and never the ultimate product. However till they’re extra customizable, NotebookLM’s different options appear extra helpful.