Generative AI
Particularly if all it takes is 1 click on to show any content material into podcast
For many who might not know, NotebookLM is a personalised AI analysis assistant powered by Gemini 1.5 Professional, designed to make sense of complicated data. Along with answering questions based mostly in your uploaded sources (paperwork, slides, charts, and so on.), it will possibly additionally create personalised examine supplies by routinely producing issues like a desk of contents, examine guides, briefing paperwork, FAQs, and extra. Whereas it formulates solutions based mostly on the uploaded sources, it additionally gives inline citations, highlighting the particular textual content blocks within the supply paperwork used to generate the response.
The uploaded content material can vary from analysis papers and assembly transcripts to quotes from attention-grabbing books, chapters of a novel you’re writing, company paperwork, and extra. These sources can embrace Google Docs, Slides, PDFs, textual content recordsdata, copied textual content, and even internet pages.
Now, to the principle cause for this text: Final month, NotebookLM introduced a brand new characteristic — Audio Overviews — which has been making headlines. This characteristic presents a brand new technique to work together together with your supply paperwork. With only one click on, it generates partaking “deep dive” discussions that summarize the important thing subjects in your sources.
What’s much more spectacular is the way it transforms any piece of content material, irrespective of how dry, by producing two AI hosts (one male and one feminine) who talk about the doc’s contents in a podcast-style format.
For those who’re questioning what “podcast-style format” means, think about the pleasant banter, the little jokes, the back-and-forth conversations, the laughs, interruptions, “umms,” and “you is aware of’”— basically all of the hallmarks of an awesome podcast listening expertise.
These podcast-style conversations create pure connections and segues out of your textual content, leading to a fascinating dialogue.
To check it out, I made a decision to repurpose one in every of my previous Medium articles and create a podcast from it to cater to a extra audio-loving viewers.
The arrange for a similar was fairly simple.
- Go to NotebookLM. You’ll must sign up together with your Google ID should you aren’t already. If it’s your first go to, you’ll see a number of pattern notebooks and you may create a brand new one with the “Create” button.
- Subsequent, add content material to your pocket book. I used the web site supply to feed in my Medium article. Alternatively, you possibly can paste textual content or fetch from Google Drive.
- Lastly, click on the “Generate” button contained in the Pocket book information (see picture under) to create the audio. And go seize a ☕️ as it would take a couple of minutes relying on the content material size.
P.S. It took round 4 minutes to generate a 13 minute audio from my 1100-word article. You’ll be able to play and hear right here.
P.S. I ended up attempting Audio Overview with varied sources, equivalent to podcast transcripts, analysis papers, and knowledge science blogs. The next takeaways are an amalgamation of my experiences throughout all these sources.
Let’s begin with the great things:
- It’s outstanding that we are able to shortly create a podcast episode in simply minutes, permitting many people to have a facet gig as podcasters (do you have to select to). It is a smart way for writers to repurpose their content material and for others to interact with comparatively complicated subjects in a enjoyable and accessible method.
- Using analogies all through the audio is actually outstanding and fascinating. Within the case of my Medium article, it was in a position to take a comparatively area of interest (learn:boring) subject (scaling challenges with Gen AI may not enchantment to everybody exterior the instant subject) and make connections to on a regular basis issues.
As an example, at one level the hosts talk about Gen AI token prices and supply a way more relatable instance, evaluating how these prices can add as much as micro-transactions in a cellular recreation. Equally, they clarify immediate engineering with an instance of offering an entire recipe with measurements, reasonably than merely saying “make me a scrumptious meal”. In addition they use the analogy of a automobile remembering a typical route to elucidate LLM caching. - The best way the 2 hosts construct on one another’s sentences feels very pure, and the segues move seamlessly. For instance, utilizing phrases like “talking of…” to introduce a brand new subject feels natural and never pressured in any respect.
- Emphasis on sure phrases at simply the appropriate moments helps maintain the listeners’ consideration. Expressions like “oh wow”, “oops”, and “aah” convey real shock at what the opposite host simply mentioned. Pure pauses to consider the appropriate phrase make the dialog really feel spontaneous reasonably than rehearsed.
- After testing this on a number of deep studying papers, I can confidently say will probably be a recreation changer for explaining complicated analysis that advantages from analogies and “clarify like I’m 5” (ELI5) examples. In reality, the rules in one in every of their pre-prepared instance notebooks, titled Introduction to NotebookLM, state that it’s designed for researchers, journalists, college students, and enterprise professionals.
Having seemed on the key benefits, there are additionally a number of limitations to think about:
- Typically, the dialog between the 2 hosts doesn’t really feel actual. Fairly often, they end one another’s sentences, even when the primary host has simply requested the second host to elucidate a brand new idea and some seconds later, Host 1 finally ends up answering their very own query.
- Not all enter sources generate audio of equal high quality. As a part of stress testing, I attempted inputting the transcript from one other podcast, and the hosts appeared extra inclined to make humorous noises at one another — ‘yayaya,’ ‘oh yeah,’ ‘hmm,’ ‘uh-huh,’ ‘proper,’ ‘gotcha,’ and so on.!
- The one draw back to having a number of analogies whereas discussing a subject is that typically the AI can get the analogies fallacious. As an example, whereas discussing a weblog on forecasting metrics, it used the analogy of “identical to in colleges a decrease rating is usually higher, it means your forecast is nearer to actuality”.
Such hallucinations are frequent throughout totally different generative AI fashions and have been included as a disclaimer of their software as effectively. These may be extra pronounced if we offer a really area of interest, extremely specialised subject, such because the position of microRNAs in gene regulation (the subject that received the Nobel Prize in 2024 this week). In such instances, it might begin hallucinating with analogies used attributable to a scarcity of related inherent data🤷♀. - For very giant texts, the podcast can usually finish abruptly. This means that there could also be a cutoff level for the coaching knowledge, past which the audio can’t adapt to offer a clean, pure ending.
- (Very minor however) A number of the phrases, largely abbreviations, are garbled within the audio. For some cause RAG is pronounced as ArrrR-G as a substitute of particular person alphabets like R-A-G.
- At instances, hosts overly agree with each other, utilizing filler phrases like ‘proper’ and ‘precisely’ whereas the opposite host remains to be speaking. This will really feel like pressured responses; I imply, let the poor man end!
Now that we’ve coated the nice and the unhealthy, let’s transfer on to the million-dollar query: is that this new tech sufficient to provide podcasters a severe competitors?
My easy reply is — not but. The explanation? All of the aforementioned points we’ve mentioned. And I do know a few of you would possibly disagree and say these issues are minor, and also you’d be proper. For those who hearken to only one podcast, you might not even discover them, however should you repeatedly hearken to a number of episodes, particularly on a every day or weekly foundation, the sheer variety of analogies and “exactlys” can grow to be overwhelming. For these causes, maybe Google by no means positioned it as a podcasting software of their preliminary launch.
That mentioned, it is going to undoubtedly decrease the barrier to entry for a lot of who wish to discover this subject however might not wish to use their very own voice for varied causes. Extra importantly, I see its use as a technique to eat complicated subjects in digestible codecs.
