Imagine for a second that you're absolutely starving. You've got two options for dinner. Option A is this trendy five -star restaurant. You walk in. The service is instant. The food appears like magic. It's delicious. How good so far? But you have to admit, the kitchen is a total black box. You have no idea what's in the sauce. Right. And if you asked to see the recipe, they'd probably kick you out. Exactly. That's most AI tools today, like notebook LM, fast, polished. but opaque.
Now option B, you're in your own kitchen, it's hot, it's messy, maybe you burn the toast, but you bought every single ingredient yourself. You know exactly what's going into that meal. And that, right there, is the trade -off Surf Sense is asking us to make. Welcome back to the Deep Dive. Today, we aren't just looking at another productivity app. We're looking at a tool that is trying to fundamentally shift the power dynamic between you and your data. We're talking about
SurfSense. And this is so much bigger than just organizing files. The core theme here is really digital sovereignty. That's a heavy term. It is. But it boils down to a simple question. Do you own your thoughts? Or are you just renting them from a server somewhere in California? SurfSense is making a very bold pitch that you should be the chef in your own kitchen. even if that means you have to do the dishes. Even if it means doing the dishes yourself. I love that framing. So
here's the roadmap. We're going to break down how this whole local AI thing works in practice. We'll look at the setup, the ability to actually edit your source documents, which is a feature I didn't know I needed, and the reality of running this offline. And we have to be honest about the fiction points because, you know, spoiler alert. Cooking at home takes longer than ordering
takeout. It really does. We're going to get into the slow processing, the loud laptop fans, and why the audio features might be a total letdown. But let's start at the beginning. When you first open SurfSense, it feels different. It's dark. I mean, you open ChatGPT, and the UI is trying to be your best friend immediately. How can I help you today? Right. It's very eager. SurfSense just opens and it's quiet. It's empty. It felt a little like moving into a new house before
you've unpacked any boxes. Just... white walls. That emptiness is intentional, though. It forces a mental shift. Most of us treat AI like a junk drawer. We just dump everything in and hope for the best. And SurfSense says no. SurfSense says no, you need to curate this. The whole interface is built on two concepts, spaces and sources. OK, let's talk about spaces. To me, that just sounds like a fancy word for folders. Is there a real difference? There is. And it's actually
crucial. A space isn't just a folder. It's a A context container. A context container. Yeah. And your normal cloud drive, you might have a folder called work stuff with random PDFs. Do that in Surf sense. And the AI is going to get confused. It might hallucinate. Because it doesn't know what lens to look through. Exactly. It needs boundaries. The source we look at had this great example about naming. You can't just call a space test. Right. If you do that, the AI has no semantic
anchor. It doesn't know if bank means a river bank or a financial institution. The example in the guide was Remote work tips, productivity, and health, which is a mouthful. It is, but look what it does. You're basically prompt engineering before you even ask a question. You're telling the system, hey, limit your scope to just this. You're priming the engine. So you're whispering the intent to the AI before you even upload a file. That's interesting. It makes you be more
organized. It does. It's not for the lazy. It asks you to think about why you're saving something. So we have our box. We've named it with intention. Now we have to fill it. And this is where I think a lot of people are going to hit their first speed bump. Yeah, you're going to hit the processing wall. The dreaded progress bar. When you upload a PDF, it's not just storing it. It's indexing it locally on your machine. It's turning every word into vectors, which is just math the AI
can read. And that takes horsepower. And there's a file limit, right? Something like 300 pages? Right. So you can't just upload a massive textbook in one go. The system will choke. You've got to split it up. But once that file is in there, something happens that I have not seen in other tools. And I think this is the aha moment. On the Editor feature. The document isn't static. No, it's a game changer for anyone who deals with messy data. You can actually type inside
the source document. You can. Think about why that matters. A PDF is usually locked, right? It's like a photograph of text. But say you upload a scan from 1990. the OCR might mistake an S for a five. So you get FIFE offense instead of SurfSense. Exactly. And if you feed that garbage into a normal AI, it might get confused. It's just noise. In SurfSense, you can literally open the text layer, see that error, and just fix it. So you are scrubbing the data clean before
the AI even touches it. It's the garbage in garbage out principle. You're cleaning the vegetables before you cook. You're ensuring the answer is actually accurate. I can see that being huge for researchers. But, you know, we don't just live in a world of PDFs. Half my reading is just watching YouTube videos. And SurfSense knows that. It calls itself a digital binder. It's trying to flatten the information hierarchy. It doesn't care if the data came from a hardcover
book or a YouTuber. It's all just data. Tell me about the video workflow. Because my watch later playlist is basically a graveyard where content goes to die. We all have that graveyard. Yeah. SurfSense changes that. You take the URL for, say, a 45 -minute video on home office economics, and you just paste it in. And it doesn't just save a link. No. It rips out the transcript. It pulls all the spoken words and turns them
into text. So now, in your remote workspace, you've got a PDF guys, a blog post, and the full text from that video. And the AI sees them all as equal. That's the moment of wonder. You can ask a question, and it can pull a quote from page 10 of a PDF and combine it with a tip from the 20 -minute mark of a YouTube video. It synthesizes across formats. That is powerful. It's like having a research assistant who watches the videos for you. But, okay, having the data is one thing.
Getting good answers out, that's where I usually struggle. I'll admit it, I'm a lazy prompter. To summarize this prompt, we all are. But with a tool like this, Precision really pays off. It has this mentions feature, kind of like on social media. It's critical for local AI. You have to remember this is running on your laptop, not a supercomputer. If you ask a generic question, it has to scan everything. It's inefficient. Very. But with mentions, you point the flashlight.
You type the at at symbol. And I can say, at YouTube video, what did it say about chair height? Exactly. You're telling the AI, ignore the PDFs, focus only on this. It's constraints -based prompting. It gets you bitter answers faster. The guide had some great examples. One was asking for a comparison table. Yes. Create a table comparing the advice at PDF and at video. And the result, was it good? The content was accurate, but visually it was a little rough. Squished was the word
I saw. Very squished, yeah. You had to scroll sideways to read it. It doesn't have that polished design you'd get from Google. But the information was spot on. It correctly identified that the PDF focused on time blocking while the video was all about disk ergonomics. That's a high level synthesis that would take me 20 minutes to do by hand. And it did it in seconds. Another great one was the quiz generation. Oh, for students this must be gold. It is. Create 10 multiple
choice questions based on at textbook. and it generates a test with an answer key. That's amazing for active recall. But the thing that really builds trust for me is the citations. The little numbers. The little numbers. When it gives you an answer, there's a little number at the end of the sentence. And when you click it, it doesn't just take you to the document. It highlights the exact sentence in the source text where it got that information. That's the accountability.
That's checking the chef's work. Precisely. It makes it really easy to spot hallucinations. If the AI says the moon is made of cheese, and cites page four, you can click and immediately see it actually says the moon is made of rock. Okay, so we've got our spaces, our data is clean, we're prompting with the at symbol, but does this tool live on an island? Or does it connect to anything else? It's surprisingly social for
a privacy tool. It integrates with Gmail, Google Drive, Discord, and This is the big one, Notion. Notion users are a passionate bunch. I say that as one of them. It's a whole lifestyle. Yeah. And SurfSense taps right into it. The guide detailed a use case with messy meeting notes that I think will resonate. Walk us through that because getting notes out of Notion is usually a pain. Well, normally you'd have to export to PDF. download
it, re -upload it, it's friction. With SurfSense, you just connect it directly to a specific Notion page. And then you can prompt it, like, look at meeting notes and list my to -dos. And it spits out a clean list with due dates. It turns that mess into an action plan. Yeah. And it's a live connection, so if you update Notion, SurfSense sees the changes. That is incredibly useful. But now we have to talk about the elephant in the room. or the server in the room. The local
mode. This is what separates SurfSense from everything else, isn't it? It's the whole point. It is. I mean, with chat GPT, your data gets sent to a massive server farm. Your tax returns, your diary, it all leaves your house. Which creates a vulnerability, even if you trust the company. Exactly. SurfSense says you don't need to trust them. You can download the brain, the actual AI model, like Llama 3, directly onto your laptop. So I can literally just turn off my Wi -Fi. You
can pull the ethernet cable. Be on a plane. It will still work. Total privacy. The data never leaves your hard drive. That sounds like magic. But there has to be a catch. You can't run a supercomputer on a MacBook without some consequence. And the catch is your hardware. These brains are heavy. They need a lot of RAM, a lot of GPU power. So if I'm on a five -year -old laptop... It's gonna struggle. The fans will spin up like a jet engine. Your battery will drain. And it's
going to be slow. How slow are we talking? Instead of an instant answer, you might wait 30 seconds. The interface itself might lag. Your computer is doing the work a data center normally does. You can feel the weight of it. So that's the trade -off, privacy for performance. And patience, which brings us to the rough edges. We promised we'd be honest about this. Yes. And the source material was pretty brutal about some of the failings. The biggest ones seemed to be audio.
Yeah. We talked about Notebook LM and its audio overview feature, which is amazing. It generates a podcast on the fly. SurfSense claims to do it, too. But in testing? It failed. Repeatedly. Just error messages. The author tried to generate a summary of that remote workspace, and it just wouldn't compile. And when it finally did work? The reports say it sounded robotic, like an old GPS voice. It sounded like a computer reading a text file. So if you want a smooth audio summary
for your jog, This is not the tool. No, stick to Notebook LM for that. This is for reading and analyzing, not listening. In the general performance, you mentioned that heaviness. The word used was clunky, like buttons taking five or six seconds to respond. That doesn't sound like much, but in 2024, six seconds is an eternity. It breaks your flow. You click save and you have to wait. It's a constant reminder that this is complex software running locally, not some slick
web app. Functional, not beautiful. A tool for a workbench. Not a toy. Exactly. And for some people, that's a deal breaker. If you want that seamless, Apple -like experience, you will be frustrated. So let's bring this all together. We have a tool that gives you total data ownership, but it's slow, a bit clunky, and some features are half -baked. Who is this actually for? I think it's for two very specific groups. First,
the privacy absolutists. Lawyers, researchers with sensitive IP, or just anyone who hates the idea of big tech scanning their data. For them, this is the only game in town. Right. And the second group is the organizers. The people who want to curate a library, who want to fix the typos, who want to build a second brain that is actually theirs, they'll put up with the friction for the control. It really does come back to that cooking analogy we started with. It does.
SurfSense is home cooking. It's messy. You have to chop the onions yourself. The fans will spin. But you made the meal. You know what's in it. And the cloud tools are fast food. Convenient, tasty, but you're eating what they serve you. And look, sometimes you just want a burger. There's nothing wrong with that. I use cloud AI every day for quick things, but for the important stuff. You might want a home -cooked meal. I think that's the perfect way to fame it. It's not about one
being better. It's about the right tool for the job. Exactly. So here's our challenge to you, the listener. Don't try to switch your whole life over to SurfSense today. Just try one small project. Just one. Maybe it's your tax prep or planning a vacation. Create a space. Upload three or four documents, try fixing a typo in the editor, and just see how it feels. See if the friction is worth that feeling of ownership. You might
find you actually like being the chef. And if your laptop fan starts sounding like a helicopter. That's the sound of privacy. Thanks for diving in with us. We'll see you in the next one. Take care.