It’s sufficient to say that NotebookLM has transformed how we interact with a vast amount of information. But as anyone who deals with extensive datasets knows, the real test of such a tool often comes when you push it beyond its comfortable limits. I was curious to see just how robust NotebookLM truly is, so I deliberately overloaded it with massive documents and long YouTube videos, far larger than typical use cases, and see how it handled (or didn’t).

Let me share my journey into the depths of NotebookLM’s processing capabilities and reveal what happened when I tried to break it with overwhelming information.

According to NotebookLM’s FAQs, a user can get 100 notebooks, with up to 50 sources (300 in the Pro plan), and 500,000 words each.

Understanding Tesla’s sustainable energy impact report

A mixed bag

When I decided to truly stress-test NotebookLM, I didn’t just pick any document. I chose a 42-page file titled ‘Tesla’s Sustainable Energy Impact: 2024 report’. This wasn’t some simple text file; it was a dense, carefully crafted report.

The report was packed with text, charts, images, and graphs illustrating energy consumption and production trends. It was a comprehensive, multi-faceted document that demanded not just textual understanding but also visual comprehension from the AI.

As expected, the upload process was smooth, and NotebookLM generated a summary in no time. Here’s a glimpse into the kind of questions I asked.

  • What are the main areas of focus for Tesla’s sustainable energy initiatives in the coming years?
  • What were the primary sources of greenhouse gas emissions mentioned in Tesla's operations?

In both cases, NotebookLM got me relevant answers. Now, I wanted to extract specific data, so I asked the question below.

  • How many metric tons of CO2e did Tesla customers avoid releasing into the Earth's atmosphere in 2024?

Here, the correct answer was 32M (a 60% increase from 2023), but NotebookLM failed to highlight that.

Reading Apple’s earnings call via NotebookLM

Unleashing NotebookLM’s true potential

After being impressed with how NotebookLM handled the Tesla report, I decided to push its multi-document capabilities. I gathered transcripts from Apple’s last three quarterly earnings calls. These aren’t short, casual conversations; each call is a lengthy, detailed exchange between Apple executives and financial analysts about performance metrics, market trends, and future outlooks.

My goal was clear: I wanted to see if NotebookLM could stitch together insights from these separate, yet related, documents.

I created a new notebook, added these PDF files, and once I was confident NotebookLM had ‘read’ them, I started asking questions. I began with broad questions like What were the key revenue figures and growth drivers for Apple in the past three quarters?, and it gave me key figures about relevant categories with sources.

I decided to push it further and asked How does Apple describe its strategy for ‘Apple Intelligence’ across these calls? and received a detailed answer regarding their phased and expanding rollout, deep hardware integration, and key features.

The best part is, I can always click the source and refer to the exact paragraph in the uploaded PDF document.

I even asked What’s the outlook for the next quarter? and NotebookLM pulled relevant information from the previous quarter's report and shared key findings.

Learning self-hosting with educational YouTube videos

The results will surprise you

I often come across detailed, long-form educational content on YouTube, but watching a 3-hour or 4-hour video, even on a topic I’m interested in, is a huge time commitment. For instance, I’m already familiar with Kubernetes, but I knew these couple of videos likely contained specific architectural details, troubleshooting tips, and best practices.

So, I found two Kubernetes videos on YouTube – one clocking in at around four hours and another at approximately three hours. My mission was to bypass the traditional viewing experience and go straight to the knowledge.

I simply copied and pasted the YouTube links directly into NotebookLM and had them listed as active sources.

First, I asked Can you explain the core differences between a Deployment and a StatefulSet? and it pulled relevant information in no time.

I decided to push it further and asked, What are the recommended best practices for securing a Kubernetes cluster?, and to my surprise, it pulled relevant information from both videos, stitched it together, and displayed the answer.

Overall, I found that NotebookLM works at its best with text-heavy PDF files and YouTube videos. As soon as you upload large PDF files with a lot of images and graphs, the accuracy can be hit or miss. After all, there is a reason why Google asks to double-check NotebookLM’s responses.

Confessions of a NotebookLM abuser

After pushing NotebookLM to its very limits with large documents, my experiment has been eye-opening. There were several hiccups, but in most cases, NotebookLM did the job just fine.

Of course, like any tool, NotebookLM has its limitations with the number of sources, but it’s a generous one, and you will have a hard time crossing that for your workflow. You can now go ahead and create an ‘Everything’ notebook in NotebookLM to handle information overload without breaking a sweat.