Latest / AI Fire Daily / #503 Neil: NotebookLM 2.0 Now Creates PowerPoint And Excel Files
Transcript
- 0:00I mean, you ask an AI a highly complex question.
- 0:03You hope for profound clarity, right? Yeah, you
- 0:06want to answer that actually connects the dots.
- 0:09Exactly. But instead, you just get this massive,
- 0:11impenetrable wall of text beat, and then the
- 0:15real work begins. Oh, totally. You have to manually
- 0:17extract that data line by line. You build the
- 0:20presentation yourself. You rebuild the spreadsheet
- 0:22yourself. It takes hours of your time. And honestly,
- 0:25it's incredibly frustrating. Two secs silence.
- 0:28But I want you to imagine a totally different
- 0:30scenario today. OK, I'm listening. You ask that
- 0:33exact same question, but instead of text, a fully
- 0:36editable PowerPoint file just appears. A populated
- 0:41Excel spreadsheet materializes right there in
- 0:44the chat. I mean, that changes everything about
- 0:45how we work. You stop fighting the interface
- 0:47entirely, and you start actually collaborating.
- 0:50Welcome to the deep dive. I'm very glad you're
- 0:52here with us. Our primary mission today is exploring
- 0:55that exact shift. We are looking really closely
- 0:57at Google's notebook, LM 2 .0, because the source
- 1:01material reveals a fascinating transition in
- 1:03what this technology is actually doing. Right.
- 1:06This tool was previously just a simple document
- 1:08reader. It would summarize what you gave it.
- 1:10Yeah, but now it operates much closer to a junior
- 1:13data analyst. To really trust an AI, to build
- 1:15our spreadsheets and parse our data, we first
- 1:17have to understand how it suddenly learned to
- 1:19do real math. And that requires looking closely
- 1:22at the engine powering this whole operation.
- 1:24Right, because to understand how it creates these
- 1:27tangible files, we need context. We kind of have
- 1:30to look at the brain first. Notebook LM has officially
- 1:33moved to the Gemini 3 .5 engine, and that brings
- 1:37some heavy implications for you. The benchmarks
- 1:40in the research are pretty staggering. It's showing
- 1:42a 78 .2 % win rate in web research. And when
- 1:45it comes to analyzing massive, complex documents,
- 1:49it is hitting nearly 70%. But, you know, raw
- 1:53numbers only tell part of the story. Yeah. The
- 1:55fascinating part is why it is suddenly so much
- 1:58more capable. Exactly. Historically, large language
- 2:01models are terrible at math. They don't actually
- 2:03calculate numbers. No, they just predict the
- 2:05next most likely word in a sequence based on
- 2:08their training data. Right. It is basically highly
- 2:10educated guessing. Which is why you could never
- 2:12truly cross them with your accounting or your
- 2:16precise data sets. But the architecture here
- 2:18has fundamentally changed. Every single notebook
- 2:21now gets its own sandbox. Right. Which is a safe,
- 2:24isolated space for running code. Right. And that
- 2:26is the actual game changer here. Before, the
- 2:30AI just estimated math from text. Now... It writes
- 2:33and runs its own code. It relies on over 100
- 2:36curated software skills to compute absolute answers.
- 2:40Whoa. I mean, imagine it just spinning up its
- 2:42own secure cloud computer to calculate your invoices.
- 2:45That is wild. You upload three years of complex
- 2:48billing records. Instead of summarizing what
- 2:50might be in them, it takes physical action. It
- 2:53writes a unique script. It runs it in that isolated
- 2:56sandbox. And it hands back a mathematically perfect
- 2:58finished report. Beat. Okay, let's untack this.
- 3:01It stops being a guessing machine and becomes
- 3:03a computation engine. But wait, if it is writing
- 3:06code to achieve this, am I expected to know how
- 3:08to prompt in Python or manually trigger these
- 3:10specific skills? Not at all. That is the beauty
- 3:13of the system design. The AI automatically evaluates
- 3:16your prompt, realizes it needs hard math instead
- 3:18of text, and silently writes and executes the
- 3:21necessary scripts in the background. So it silently
- 3:24picks the right tools without you lifting a finger.
- 3:26Precisely. It completely abstracts the coding
- 3:29layer away from you. But an engine that powerful
- 3:32is useless without the right fuel. So how do
- 3:35we actually feed this new system? The fundamental
- 3:38rule of data analysis has not changed at all.
- 3:41The quality of your sources still absolutely
- 3:43matters. Yeah, the old adage applies perfectly
- 3:45here. Barbage in, garbage out. Notebook LM simply
- 3:49amplifies the foundation you handed. Let's make
- 3:51this concrete for you. Imagine you're building
- 3:53a new course on freelance copywriting. OK, you
- 3:56need a very solid, multifaceted knowledge base
- 3:59for that. Exactly. So you upload competitor course
- 4:02pages and their pricing tiers. You drop in lengthy
- 4:05Reddit threads where real freelancers are asking
- 4:07questions. You add your own scattered notes from
- 4:09past client work. You include broad industry
- 4:11reports on freelance income trends. It actively
- 4:14connects all of those disparate pieces together.
- 4:17It treats all of it. as one connected mind. Yeah,
- 4:20rather than separate files sitting in different
- 4:22folders. But, you know, human research is rarely
- 4:26ever perfect. There are almost always missing
- 4:28pieces in our logic. Right, and this is where
- 4:31the self -directed research feature comes in.
- 4:33You can simply ask the chat what is missing from
- 4:36your own brain trust. It reviews your uploaded
- 4:39sources alongside your conversation. For our
- 4:42copywriting course example, it easily spots two
- 4:45clear structural gaps. First, it identifies completely
- 4:49missing financial data. Right. You lack customer
- 4:51acquisition cost numbers and you lack average
- 4:54conversion rates for sales pages. It finds the
- 4:57financial weaknesses in your business plan. It
- 4:59really does. Then it finds a blind spot in your
- 5:01focus. Your research perfectly explains how students
- 5:04will acquire freelance clients. But it entirely
- 5:07fails to explain how you, the creator, will acquire
- 5:10those students in the first place. I genuinely
- 5:12love that level of critical pushback from a tool.
- 5:15It's not just agreeing with me. No, it's finding
- 5:17the holes in your thinking. And you can tell
- 5:19it to actively chase down one of those gaps.
- 5:22Let's say you pick customer acquisition costs.
- 5:24It searches the live web for those exact missing
- 5:28metrics. It comes back with a tailored report
- 5:31and suggested external sources to fill the hole.
- 5:34Two -sec silence. I have to admit something vulnerable
- 5:37here. I still wrestle with the fear of AI hallucinating
- 5:41data, so approving sources manually is huge for
- 5:44me. If it's searching the open web to fill these
- 5:47gaps, how do I know it's not quietly slipping
- 5:49unverified junk in my course outline? Because
- 5:52of the hard gatekeeping built into the interface,
- 5:54it will suggest those web sources, yes, but it
- 5:57physically cannot incorporate that data into
- 5:59your Notebook's core knowledge base until you
- 6:01explicitly click approve. You remain the strict
- 6:04gatekeeper for every single piece of data. You
- 6:06do. Every answer stays completely grounded in
- 6:09material you have personally reviewed. This deep
- 6:11dive is completely supported by our listener
- 6:13community. Your backing keeps this conversation
- 6:15independent, ad -free, and deeply curious. We
- 6:19appreciate you being part of this journey with
- 6:21us. And we are back. Let's keep exploring this
- 6:24architecture. We now have a perfectly curated
- 6:27and vetted brain of sources. Right. But how do
- 6:30we actually extract the work from it? Right.
- 6:36downloadable files. What's fascinating here is
- 6:38the physical reality of the files it generates.
- 6:41This all happens within the studio panel on the
- 6:43right side of your screen. Let's look at the
- 6:45first major example. Say you want an Excel spreadsheet.
- 6:48Okay. You need a 12 -month revenue projection
- 6:51broken down by specific marketing campaigns.
- 6:54You just select the approved sources holding
- 6:56your figures and you give the instructions in
- 6:58the chat. It pulls those scattered numbers into
- 7:00one unified Excel sheet. And importantly, the
- 7:03formulas and the numbers are real data. They
- 7:06automatically update when you edit them later
- 7:08in Excel. Yeah, it performs the tedious structural
- 7:10formatting work for you. Saves hours of cell
- 7:13linking. And then there's the PowerPoint generation,
- 7:15which is equally impressive. You need a 10 slide
- 7:18pitch deck for investors. You tell it to focus
- 7:20heavily on the business model. You instruct it
- 7:23to skip the company history section entirely
- 7:24because investors don't care. A real formatted
- 7:27file shows up a few minutes later. But there
- 7:29is a significant cache we must discuss here.
- 7:31It is a known limitation that many users eventually
- 7:35run into. There is. It involves how the AI handles
- 7:38the PowerPoint text boxes. They often generate
- 7:42as flat images packed inside a PowerPoint frame
- 7:45rather than editable text fields. It's like stacking
- 7:48Lego blocks of data. You can't just mold the
- 7:50plastic, you have to swap the blog entirely.
- 7:53Because large language models struggle with spatial
- 7:54bounding boxes, they often just render the text
- 7:57as a picture. Exactly. You cannot simply click
- 8:00and type directly on the slide to fix a typo.
- 8:02You must use the revise button located in the
- 8:05chat interface. You ask the chat to change the
- 8:07specific slide. And it regenerates the whole
- 8:09block. The revise button becomes your primary
- 8:12tool for edits. It rewrites the slide instead
- 8:14of you typing on it. If I am staring at a newly
- 8:17generated deck right before a meeting, how do
- 8:20I quickly figure out if I am looking at editable
- 8:22text or just a baked -in image block? It is a
- 8:25very simple manual test. You download the file
- 8:28and just click any line of text. If your cursor
- 8:31does not let you type directly into the paragraph,
- 8:33it is an image block. So a quick click... instantly
- 8:36reveals if it is an image block. Exactly right.
- 8:39That need for structural revision is a good segue
- 8:42into how it handles truly dense contradictory
- 8:45information. Because sometimes even a perfectly
- 8:48formatted spreadsheet is still too dense. Right.
- 8:51We need to turn that density into a cohesive
- 8:53story we can understand at a single glance. Let's
- 8:56look at a very specific data comparison example
- 8:59from the research. You upload two distinct files
- 9:02into your notebook. First, your monthly Facebook
- 9:05ad spend broken down by campaign. Second, your
- 9:08daily course enrollments exported from your hosting
- 9:10platform. Two completely different formats of
- 9:13data. Right. And you ask it to compare those
- 9:15two specific data sets. The resulting analysis
- 9:18is incredibly sharp. It calculates the cost per
- 9:21enrollment for each month by synthesizing the
- 9:23two files. It discovers the cost drops from roughly
- 9:26$79 .59 in January, all the way down to $31 .91
- 9:31by June. The advertising is clearly getting much
- 9:35more efficient over time. But here is the truly
- 9:40critical moment in this interaction. It's the
- 9:42moment that proves how the architecture has changed.
- 9:45You ask the AI which specific campaign drove
- 9:47the most signups. And the system admits outright
- 9:50that it cannot answer that question. The uploaded
- 9:53data only shows total daily signups. It does
- 9:56not explicitly link those individual signups
- 9:59to specific ad campaigns. Beat, the AI does not
- 10:03try to guess. Yeah, that algorithmic honesty
- 10:05is absolutely vital. Historically, AI models
- 10:07are people pleasers. They are designed to give
- 10:09you an answer, even if they have to hallucinate
- 10:11to do it. But this system is programmed to hit
- 10:13a wall when the uploaded data stops. doesn't
- 10:16invent a bridge to cross it. Exactly. Which is
- 10:18why I'm so used to these models trying to impress
- 10:20me with an answer no matter what. Why is the
- 10:22AI refusing to answer the campaign question actually
- 10:25a massive feature, not a bug? Because its primary
- 10:28directive is grounded analysis, not conversation.
- 10:31It is designed to build profound trust. You know
- 10:35it will not invent convenient answers when your
- 10:37data falls short. honesty about its limits proves
- 10:39it will never invent fake data. Exactly. And
- 10:42you can visualize this data density easily, too.
- 10:45You just ask it to turn this comparison into
- 10:46a visual chart. The chart appears within a minute,
- 10:48ready to download. And it shows the ad spend
- 10:50climbing from $3 ,900 to $9 ,000. but the enrollments
- 10:55rocket from 49 to 282. That is a massive 475
- 11:00% increase. One glance proves the fundamental
- 11:03efficiency of the business. You see both lines
- 11:06rise, but the enrollments grow significantly
- 11:08faster. Yeah. You can run other incredibly valuable
- 11:11comparisons too. Weekly website traffic versus
- 11:13newsletter signups. Customer support ticket volume
- 11:15against average response times. Or department
- 11:18budgets against actual quarterly spending. It
- 11:20turns a dense written analytical answer into
- 11:23actionable clarity. Trust and clarity are great
- 11:26in theory. But we need to ground this entirely
- 11:29in practical reality before we wrap up. Who can
- 11:32actually use this system right now? And what
- 11:35is the best mental framework to approach it with?
- 11:37There is a paywall we need to acknowledge up
- 11:40front. This specific update is not available
- 11:42to everyone just yet. It requires Google AI Ultra.
- 11:46Or you need to be a workspace business customer
- 11:49with AI Ultra or Expanded Access. Broad access
- 11:52is planned, but it's currently limited. Let's
- 11:54quickly recap the ideal workflow you can copy
- 11:56today if you have access. Say you run a small
- 11:59online store. you want to decide whether to expand
- 12:01your product line. You start a notebook with
- 12:03your existing documents, supplier catalogs, historical
- 12:06sales reports, customer reviews, competitor pricing
- 12:09matrices. Then you actively ask the tool what
- 12:12strategic gaps exist. You carefully review and
- 12:14approve any new web sources before adding them
- 12:16to the brain. Right. Then you generate a slide
- 12:19deck for a supplier pitch. You build a data comparison
- 12:22table comparing your past sales against ad spend.
- 12:24Finally, you turn that comparison into a visual
- 12:27chart for your team. But you must remember the
- 12:30absolute golden rule here. You must treat everything
- 12:34this system produces as a powerful first draft.
- 12:37You cannot treat this as a perfectly finished,
- 12:40infallible product. You always verify the Excel
- 12:43math against your raw sources. You always test
- 12:46those PowerPoint editing capabilities before
- 12:48walking into a presentation. Treating it as a
- 12:51strong first draft keeps the entire workflow
- 12:53trustworthy. It saves immense time, but it doesn't
- 12:57replace careful human verification. With all
- 13:00this capability, from coding in sandboxes to
- 13:02building pitch decks, does a tool like this effectively
- 13:05replace a human data analyst on a small team?
- 13:07Not at all. It just closes the massive gap between
- 13:10having a business problem and holding a workable
- 13:12draft solution. The human still completely owns
- 13:15the final strategic decisions. It does the grunt
- 13:18work, but you make the final strategic decisions.
- 13:20Perfectly said. The critical thinking remains
- 13:22entirely yours. Let's take a measured calm moment
- 13:25here at the end. Let's reflect on the overarching
- 13:27theme of this deep dive. We have witnessed a
- 13:30profound architectural shift in consumer technology
- 13:33today. We really have. We have officially moved
- 13:35from AI as a simple conversational partner, a
- 13:39glorified chatbot, to AI as an active synthesizer
- 13:42and real file creator. It uses isolated secure
- 13:45computation to do real verifiable math. It spots
- 13:48the logical gaps in our own human research. And
- 13:51it outputs highly tangible tools like bitch decks
- 13:54and complex charts. to sex silence. But all of
- 13:57this extraordinary capability comes with one
- 14:00fundamental condition. We must give it a solid
- 14:03factual foundation to stand on. I want to leave
- 14:06you with a final thought to ponder as you go
- 14:08about your day. If an AI can now perfectly connect
- 14:10the dots between your scattered notes and turn
- 14:12them into a beautifully polished investor pitch
- 14:15in seconds, it raises a profound question. Yeah.
- 14:18Will the true future skill be less about knowing
- 14:20how to format a presentation and entirely about
- 14:23how deeply you understand the original raw ideas
- 14:25you feed into the machine? Keep questioning your
- 14:28assumptions. Keep exploring these new frontiers.
- 14:31Thank you for joining us on this deep dive.