I mean, you ask an AI a highly complex question. You hope for profound clarity, right? Yeah, you want to answer that actually connects the dots. Exactly. But instead, you just get this massive, impenetrable wall of text beat, and then the real work begins. Oh, totally. You have to manually extract that data line by line. You build the presentation yourself. You rebuild the spreadsheet yourself. It takes hours of your time. And honestly, it's incredibly frustrating. Two secs silence.
But I want you to imagine a totally different scenario today. OK, I'm listening. You ask that exact same question, but instead of text, a fully editable PowerPoint file just appears. A populated Excel spreadsheet materializes right there in the chat. I mean, that changes everything about how we work. You stop fighting the interface entirely, and you start actually collaborating. Welcome to the deep dive. I'm very glad you're here with us. Our primary mission today is exploring
that exact shift. We are looking really closely at Google's notebook, LM 2 .0, because the source material reveals a fascinating transition in what this technology is actually doing. Right. This tool was previously just a simple document reader. It would summarize what you gave it. Yeah, but now it operates much closer to a junior data analyst. To really trust an AI, to build our spreadsheets and parse our data, we first have to understand how it suddenly learned to
do real math. And that requires looking closely at the engine powering this whole operation. Right, because to understand how it creates these tangible files, we need context. We kind of have to look at the brain first. Notebook LM has officially moved to the Gemini 3 .5 engine, and that brings some heavy implications for you. The benchmarks in the research are pretty staggering. It's showing a 78 .2 % win rate in web research. And when it comes to analyzing massive, complex documents,
it is hitting nearly 70%. But, you know, raw numbers only tell part of the story. Yeah. The fascinating part is why it is suddenly so much more capable. Exactly. Historically, large language models are terrible at math. They don't actually calculate numbers. No, they just predict the next most likely word in a sequence based on their training data. Right. It is basically highly educated guessing. Which is why you could never truly cross them with your accounting or your
precise data sets. But the architecture here has fundamentally changed. Every single notebook now gets its own sandbox. Right. Which is a safe, isolated space for running code. Right. And that is the actual game changer here. Before, the AI just estimated math from text. Now... It writes and runs its own code. It relies on over 100 curated software skills to compute absolute answers. Whoa. I mean, imagine it just spinning up its own secure cloud computer to calculate your invoices.
That is wild. You upload three years of complex billing records. Instead of summarizing what might be in them, it takes physical action. It writes a unique script. It runs it in that isolated sandbox. And it hands back a mathematically perfect finished report. Beat. Okay, let's untack this. It stops being a guessing machine and becomes a computation engine. But wait, if it is writing code to achieve this, am I expected to know how to prompt in Python or manually trigger these
specific skills? Not at all. That is the beauty of the system design. The AI automatically evaluates your prompt, realizes it needs hard math instead of text, and silently writes and executes the necessary scripts in the background. So it silently picks the right tools without you lifting a finger. Precisely. It completely abstracts the coding layer away from you. But an engine that powerful is useless without the right fuel. So how do
we actually feed this new system? The fundamental rule of data analysis has not changed at all. The quality of your sources still absolutely matters. Yeah, the old adage applies perfectly here. Barbage in, garbage out. Notebook LM simply amplifies the foundation you handed. Let's make this concrete for you. Imagine you're building a new course on freelance copywriting. OK, you need a very solid, multifaceted knowledge base for that. Exactly. So you upload competitor course
pages and their pricing tiers. You drop in lengthy Reddit threads where real freelancers are asking questions. You add your own scattered notes from past client work. You include broad industry reports on freelance income trends. It actively connects all of those disparate pieces together. It treats all of it. as one connected mind. Yeah, rather than separate files sitting in different folders. But, you know, human research is rarely ever perfect. There are almost always missing
pieces in our logic. Right, and this is where the self -directed research feature comes in. You can simply ask the chat what is missing from your own brain trust. It reviews your uploaded sources alongside your conversation. For our copywriting course example, it easily spots two clear structural gaps. First, it identifies completely missing financial data. Right. You lack customer acquisition cost numbers and you lack average conversion rates for sales pages. It finds the
financial weaknesses in your business plan. It really does. Then it finds a blind spot in your focus. Your research perfectly explains how students will acquire freelance clients. But it entirely fails to explain how you, the creator, will acquire those students in the first place. I genuinely love that level of critical pushback from a tool. It's not just agreeing with me. No, it's finding the holes in your thinking. And you can tell it to actively chase down one of those gaps.
Let's say you pick customer acquisition costs. It searches the live web for those exact missing metrics. It comes back with a tailored report and suggested external sources to fill the hole. Two -sec silence. I have to admit something vulnerable here. I still wrestle with the fear of AI hallucinating data, so approving sources manually is huge for me. If it's searching the open web to fill these gaps, how do I know it's not quietly slipping
unverified junk in my course outline? Because of the hard gatekeeping built into the interface, it will suggest those web sources, yes, but it physically cannot incorporate that data into your Notebook's core knowledge base until you explicitly click approve. You remain the strict gatekeeper for every single piece of data. You do. Every answer stays completely grounded in material you have personally reviewed. This deep dive is completely supported by our listener
community. Your backing keeps this conversation independent, ad -free, and deeply curious. We appreciate you being part of this journey with us. And we are back. Let's keep exploring this architecture. We now have a perfectly curated and vetted brain of sources. Right. But how do we actually extract the work from it? Right. downloadable files. What's fascinating here is the physical reality of the files it generates. This all happens within the studio panel on the
right side of your screen. Let's look at the first major example. Say you want an Excel spreadsheet. Okay. You need a 12 -month revenue projection broken down by specific marketing campaigns. You just select the approved sources holding your figures and you give the instructions in the chat. It pulls those scattered numbers into one unified Excel sheet. And importantly, the formulas and the numbers are real data. They automatically update when you edit them later
in Excel. Yeah, it performs the tedious structural formatting work for you. Saves hours of cell linking. And then there's the PowerPoint generation, which is equally impressive. You need a 10 slide pitch deck for investors. You tell it to focus heavily on the business model. You instruct it to skip the company history section entirely because investors don't care. A real formatted file shows up a few minutes later. But there is a significant cache we must discuss here.
It is a known limitation that many users eventually run into. There is. It involves how the AI handles the PowerPoint text boxes. They often generate as flat images packed inside a PowerPoint frame rather than editable text fields. It's like stacking Lego blocks of data. You can't just mold the plastic, you have to swap the blog entirely. Because large language models struggle with spatial bounding boxes, they often just render the text
as a picture. Exactly. You cannot simply click and type directly on the slide to fix a typo. You must use the revise button located in the chat interface. You ask the chat to change the specific slide. And it regenerates the whole block. The revise button becomes your primary tool for edits. It rewrites the slide instead of you typing on it. If I am staring at a newly generated deck right before a meeting, how do I quickly figure out if I am looking at editable
text or just a baked -in image block? It is a very simple manual test. You download the file and just click any line of text. If your cursor does not let you type directly into the paragraph, it is an image block. So a quick click... instantly reveals if it is an image block. Exactly right. That need for structural revision is a good segue into how it handles truly dense contradictory information. Because sometimes even a perfectly formatted spreadsheet is still too dense. Right.
We need to turn that density into a cohesive story we can understand at a single glance. Let's look at a very specific data comparison example from the research. You upload two distinct files into your notebook. First, your monthly Facebook ad spend broken down by campaign. Second, your daily course enrollments exported from your hosting platform. Two completely different formats of data. Right. And you ask it to compare those two specific data sets. The resulting analysis
is incredibly sharp. It calculates the cost per enrollment for each month by synthesizing the two files. It discovers the cost drops from roughly $79 .59 in January, all the way down to $31 .91 by June. The advertising is clearly getting much more efficient over time. But here is the truly critical moment in this interaction. It's the moment that proves how the architecture has changed. You ask the AI which specific campaign drove the most signups. And the system admits outright
that it cannot answer that question. The uploaded data only shows total daily signups. It does not explicitly link those individual signups to specific ad campaigns. Beat, the AI does not try to guess. Yeah, that algorithmic honesty is absolutely vital. Historically, AI models are people pleasers. They are designed to give you an answer, even if they have to hallucinate to do it. But this system is programmed to hit a wall when the uploaded data stops. doesn't
invent a bridge to cross it. Exactly. Which is why I'm so used to these models trying to impress me with an answer no matter what. Why is the AI refusing to answer the campaign question actually a massive feature, not a bug? Because its primary directive is grounded analysis, not conversation. It is designed to build profound trust. You know it will not invent convenient answers when your data falls short. honesty about its limits proves
it will never invent fake data. Exactly. And you can visualize this data density easily, too. You just ask it to turn this comparison into a visual chart. The chart appears within a minute, ready to download. And it shows the ad spend climbing from $3 ,900 to $9 ,000. but the enrollments rocket from 49 to 282. That is a massive 475 % increase. One glance proves the fundamental efficiency of the business. You see both lines rise, but the enrollments grow significantly
faster. Yeah. You can run other incredibly valuable comparisons too. Weekly website traffic versus newsletter signups. Customer support ticket volume against average response times. Or department budgets against actual quarterly spending. It turns a dense written analytical answer into actionable clarity. Trust and clarity are great in theory. But we need to ground this entirely in practical reality before we wrap up. Who can
actually use this system right now? And what is the best mental framework to approach it with? There is a paywall we need to acknowledge up front. This specific update is not available to everyone just yet. It requires Google AI Ultra. Or you need to be a workspace business customer with AI Ultra or Expanded Access. Broad access is planned, but it's currently limited. Let's quickly recap the ideal workflow you can copy today if you have access. Say you run a small
online store. you want to decide whether to expand your product line. You start a notebook with your existing documents, supplier catalogs, historical sales reports, customer reviews, competitor pricing matrices. Then you actively ask the tool what strategic gaps exist. You carefully review and approve any new web sources before adding them to the brain. Right. Then you generate a slide deck for a supplier pitch. You build a data comparison table comparing your past sales against ad spend.
Finally, you turn that comparison into a visual chart for your team. But you must remember the absolute golden rule here. You must treat everything this system produces as a powerful first draft. You cannot treat this as a perfectly finished, infallible product. You always verify the Excel math against your raw sources. You always test those PowerPoint editing capabilities before walking into a presentation. Treating it as a strong first draft keeps the entire workflow
trustworthy. It saves immense time, but it doesn't replace careful human verification. With all this capability, from coding in sandboxes to building pitch decks, does a tool like this effectively replace a human data analyst on a small team? Not at all. It just closes the massive gap between having a business problem and holding a workable draft solution. The human still completely owns the final strategic decisions. It does the grunt work, but you make the final strategic decisions.
Perfectly said. The critical thinking remains entirely yours. Let's take a measured calm moment here at the end. Let's reflect on the overarching theme of this deep dive. We have witnessed a profound architectural shift in consumer technology today. We really have. We have officially moved from AI as a simple conversational partner, a glorified chatbot, to AI as an active synthesizer and real file creator. It uses isolated secure computation to do real verifiable math. It spots
the logical gaps in our own human research. And it outputs highly tangible tools like bitch decks and complex charts. to sex silence. But all of this extraordinary capability comes with one fundamental condition. We must give it a solid factual foundation to stand on. I want to leave you with a final thought to ponder as you go about your day. If an AI can now perfectly connect the dots between your scattered notes and turn them into a beautifully polished investor pitch
in seconds, it raises a profound question. Yeah. Will the true future skill be less about knowing how to format a presentation and entirely about how deeply you understand the original raw ideas you feed into the machine? Keep questioning your assumptions. Keep exploring these new frontiers. Thank you for joining us on this deep dive.