Generative AI
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 info. Along with answering questions based mostly in your uploaded sources (paperwork, slides, charts, and many others.), it may possibly additionally create customized research supplies by robotically producing issues like a desk of contents, research guides, briefing paperwork, FAQs, and extra. Whereas it formulates solutions based mostly on the uploaded sources, it additionally offers 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 fascinating 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 announced a new feature — Audio Overviews — which has been making headlines. This function affords a brand new technique to work together along 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.
When you’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 important podcast listening expertise.
These podcast-style conversations create pure connections and segues out of your textual content, leading to an enticing dialogue.
To try it out, I made a decision to repurpose one in all my old 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 register along with your Google ID in case you aren’t already. If it’s your first go to, you’ll see a number of pattern notebooks and you’ll 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’ll be able to paste textual content or fetch from Google Drive.
- Lastly, click on the “Generate” button contained in the Pocket book information (see picture beneath) 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 possibly can play and pay attention right here.
P.S. I ended up attempting Audio Overview with numerous sources, akin to podcast transcripts, analysis papers, and information 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 will rapidly create a podcast episode in simply minutes, permitting many people to have a aspect gig as podcasters (must you select to). It is a wonderful means for writers to repurpose their content material and for others to have interaction with comparatively complicated subjects in a enjoyable and accessible method.
- The usage of analogies all through the audio is really outstanding and fascinating. Within the case of my Medium article, it was in a position to take a comparatively area of interest (learn:boring) matter (scaling challenges with Gen AI may not attraction to everybody exterior the speedy discipline) and make connections to on a regular basis issues.
As an illustration, 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 sport. Equally, they clarify immediate engineering with an instance of offering an entire recipe with measurements, relatively than merely saying “make me a scrumptious meal”. In addition they use the analogy of a automotive 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 circulate seamlessly. For instance, utilizing phrases like “talking of…” to introduce a brand new matter feels natural and never compelled 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 stated. Pure pauses to think about the appropriate phrase make the dialog really feel spontaneous relatively than rehearsed.
- After testing this on a number of deep studying papers, I can confidently say it will likely be a sport changer for explaining complicated analysis that advantages from analogies and “clarify like I’m 5” (ELI5) examples. In reality, the rules in one in all their pre-prepared instance notebooks, titled Introduction to NotebookLM, state that it’s designed for researchers, journalists, college students, and enterprise professionals.
Having appeared on the key benefits, there are additionally a number of limitations to think about:
- Generally, 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 many others.!
- The one draw back to having lots of analogies whereas discussing a subject is that typically the AI can get the analogies flawed. As an illustration, 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 widespread throughout completely different generative AI fashions and have been included as a disclaimer of their device as effectively. These could be extra pronounced if we offer a really area of interest, extremely specialised matter, such because the position of microRNAs in gene regulation (the subject that received the Nobel Prize in 2024 this week). In such circumstances, it could begin hallucinating with analogies used as a consequence of an absence of related inherent data🤷♀. - For very massive texts, the podcast can typically finish abruptly. This implies that there could also be a cutoff level for the coaching information, past which the audio can’t adapt to supply a easy, pure ending.
- (Very minor however) A few of the phrases, largely abbreviations, are garbled within the audio. For some cause RAG is pronounced as ArrrR-G as an alternative of particular person alphabets like R-A-G.
- At occasions, hosts overly agree with each other, utilizing filler phrases like ‘proper’ and ‘precisely’ whereas the opposite host remains to be speaking. This could really feel like compelled 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 rationale? 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. When you hearken to only one podcast, you might not even discover them, however in case you repeatedly hearken to a number of episodes, particularly on a every day or weekly foundation, the sheer variety of analogies and “exactlys” can turn out to be overwhelming. For these causes, maybe Google by no means positioned it as a podcasting device of their preliminary launch.
That stated, it’ll positively decrease the barrier to entry for a lot of who need to discover this discipline however might not need to use their very own voice for numerous causes. Extra importantly, I see its use as a technique to devour complicated subjects in digestible codecs.