Lorenzo Bernaschina is the developer of Gems Notes, a note-taking tool that uses artificial intelligence to find relationships between ideas. In this conversation, we talk about smart note-taking and how technology might be used to extend rather than replace human intelligence.
- Lorenzo’s personal website
- Lorenzo Bernaschina (LinkedIn)
- @ittaboba on Twitter
- Gems Notes
- Roam Research
- my mind
- Niklas Luhmann
- Alan Kay
- Douglas Engelbart
- Steve Jobs
Show notes include Amazon affiliate links. I get a small commission for purchases made through these links.
Lorenzo: Thank you. Thank you, Jorge. It’s great to be here.
Jorge: Well, it’s great to have you in the show today. For folks who might not know you, would you mind please introducing yourself?
Lorenzo: Sure. So, hi, everyone. I’m Lorenzo. I’m a software engineer from Italy and recently finished my studies. I have a master’s in machine learning and AI, but I also worked full-time for three and a half years as a full-stack developer. And right now, I’m building a personal knowledge management tool called Gems to help people make sense of complex topics and be more creative with the help of AI. And the reason why I’m doing it is that I see technology as a tool to augment human intelligence. I think that artificial intelligence and human intelligence ultimately solve very different classes of problems. And so, the magic happens when we combine the two, and one doesn’t exclude the other; they should actually reinforce each other.
Jorge: A lot of folks listening in might not be familiar with Gems. Can you describe how it works for folks?
Lorenzo: Sure. So it is a note-taking app that fundamentally is built around a visual interface. So this is probably one of the core differences compared to other tools. Most note-taking tools are built around folders, structured folders, and the note page, the blank page in front of you where you write your thoughts. This tool is built around visual concepts. So, at the center of the experience, there is a graph where each note is fundamentally a document, and you can connect these notes visually and group them. And also, behind the scenes, there is an AI that actually thinks along with you.
You can think about it as a sort of thinking assistant — personal assistant — a thinking companion that basically looks at your notes and says these are potentially connected or similar. And it does it for you. And so, here you see the collaboration between the human and the machine. So, this is basically the core idea of the tool: using AI to actually extend and expand human cognition and intelligence instead of trying to outperform or replace it.
Jorge: And it’s a web-based app, right?
Lorenzo: Yes, it’s a software as a service web-based app.
Jorge: I’ve been playing around a little bit with it; not too much. And I find the idea of having the primary interface be visual to be fascinating. On the Gems homepage, it leads off with a headline that says, “Make sense of complex topics ten times faster.” And I’m wondering about this idea of augmenting our ability to make sense of complex topics. How does Gems do that?
Lorenzo: So about this, I’m actually thinking about moving toward really more creators. So here, one of the challenges with these tools is this one. These tools are extremely flexible; you can actually use them for many different reasons. And so, I think in the very early days, it really comes to what kind of angle do you want to look at and pursue. And there are a couple of things. One is definitely research. So, give you a research assistant that helps you sift through the knowledge and find hidden patterns or connections of things you didn’t expect, to make new discoveries, to learn new things. To think about new concepts that you didn’t see by yourself, especially when there is a lot of knowledge. So, that’s one angle, and it’s the one on the current landing page. But there is another angle that, through interviews and talks with people, seems to be more promising at this stage, which is actually content creation.
So if you think about book writers, you know, scriptwriters, people who need to create and storytellers, write stories in general. There is really like this stage where they brainstorm their ideas visually, and they make their plot visually. And this stage is really like, a nonlinear stage where you think really non-linearly, and you really need a flexible structure like the one I give. And also, when it comes to complex stories and many facts, many events start to become also hard to be consistent and make sure that the story is coherent in all these parts. And again, here could be another very interesting application of AI that would look into the notes and say, this thing could be connected to this, and start finding things that you may miss and contradictions that you may introduce in your plot.
So, these are a couple of use cases. The one you see on the landing page is actually just one of them. From the communication perspective, the challenge right now is actually really understanding what’s the best message to show. So, I’m doing multiple versions and multiple tests of that.
Jorge: Yeah. Correct me if I’m wrong, but I think it’s fairly early days for Gems, right? You’ve only recently launched it, right? So it’s…
Lorenzo: Yeah, very early days. I did over a hundred interviews with many different people. This tool was born into the personal knowledge management field, basically. And so, people really interested in how to structure information, how to deal with knowledge, and how to make sense of it. So, the first conversations were really built around that, I think.
One of the important things about these types of tools is that they’re really flexible, as I was saying before. And so it’s really… when you design them, and you want to innovate in the space, you really need to think deeply about primitives. What are the fundamental interaction tools you give people to structure their knowledge and work with their knowledge? So, that’s what I did in the first place. I went to the experts in the field — people that are already actively using the other tools and are deeply thinking about how to use them. What are the problems? What are the good things about them? And I really try to understand the pros and cons of every one. And that’s what basically has driven the development in the very early days of the tool.
Now, the thing is actually taking these primitives that have defined it, like the visual aspect, the AI that helps you make connections automatically and everything, and how to translate them, how to transfer them to specific use cases. For example, writing a plot for a book, conducting research in a specific way, et cetera. So that’s the kind of the journey, and yes, it’s definitely early days in terms of this type of thing.
Jorge: You mentioned other tools. The tool that immediately comes to mind for me when thinking about web-based knowledge management tools where you’re building a graph, the one that comes to mind is something like Roam Research. And I know that there are others, but one thing that struck me as being different between Gems and Roam Research is that in something like Roam, when I use that, I start working on the notes themselves, which Roam treats kind of as an outline. Like, their basic building block is a block, which is like a paragraph or outline node. That’s where I’m doing the bulk of my work when I’m using Roam, and then I can switch to a graph mode where I can view relationships between these objects. But what struck me about Gems is that it kind of inverts that in that the primary interface I see when I log into it is the graph itself. And I’m wondering about the decision that led to that choice. Why start with the graph?
Gems’s visual interface
Lorenzo: Sure. So that was actually very deliberate. Specifically thinking about Roam or even Obsidian. So you’re right. There is this graph, but talking to people… turned out that this graph it’s just cool. It’s just something to look at and say, “here’s my knowledge base that is growing.” It’s just visual feedback of the connection and the thing you made at the note level.
So still, as I was saying before, the main interface of those tools is actually the page. And so, you’re still kind of forced to think in a very linear way, if you like. And then you have this thing of the visual aspect of it, but it’s not really like at the core. And it turned out that it’s actually very helpful, especially at scale. Once you start having a lot of information in the system, it turned out to be very helpful for people to actually make sense of and find hidden connections faster.
But again, it’s not like… since it’s not like the core of those tools, it’s still like a missed opportunity and hard to do in there. And so, that was basically the starting point for my tool. I said, “let’s turn this upside down. Let’s start from it!” Because if you think about it, the visual powers of our brain are extremely good. We’re extremely powerful when it comes to making sense and finding connections, doing it visually instead of with a page. So, that was really like the first aspect because, at the end of the day, the idea of this tool is actually to make you more creative. And so it does it in a couple of ways.
The first one is this one, the visual aspect. And the second one is the AI. So, how to make you creative is fundamentally how to help you think laterally, how to find associations and think non-linearly. And to do that… again, during my research with the user, the visual aspect and the possibility to actually have an intelligence assistant behind the scenes that work for you was very interesting.
Another thing I would say about the tool landscape, in general, is that I think there is one family of tools, which is actually very manual. And I think Roam, Obsidian, those types of tools fall into that kind of category. Even Evernote — the more traditional note-taking tools — basically, there you put your knowledge. They work as great storage, but all the work is on you on tagging and keeping things organized and structured, for the most part. Then there are some other tools like my mind that are actually the opposite. So, it’s almost everything is automatized. So you throw your stuff in it, and the system automatically tags and tries to keep things organized, et cetera.
I think that Gems stays a little bit in the middle of these two. I want to really give people the possibility to build their own structure because if it comes to making sense of things and helping you be more creative on your growing body of knowledge, that really matters. But on the other side, I also wanted to actually help you because when it grows and becomes larger, that’s where things start becoming very hard with the existing tools. And so, you kind of lose control. If there is someone behind the scenes that help you… keeping everything organized and consistent, that would be very helpful. So I’m kind of the middle. I’m trying to find a trade-off between these two worlds.
Jorge: Yes, that is fascinating to me. I am very drawn to this idea of the tool not doing the thinking for you but somehow assisting you in the thinking. I’m wondering about what organizational mechanisms Gems provides. Like, it feels to me like this would be a fine line to walk, right? Because I can imagine setting an AI onto your corpus of notes and letting it auto-tag it. Like I can envision how that would happen. But how do you draw the middle line between that and allowing the user to do their own organization?
Lorenzo: So right now, it’s actually not really auto-tagging. So what the AI currently does are a couple of things at the current development stage of the tool. So, for example, if you import your notes, your corpus of notes. Say… I don’t know, 200 notes? The system automatically clusters them by topic. And so it starts introducing for you some structure. And say, given these 200 notes, these ten are about this topic. These ten are about this other thing. These six are about another thing. And it basically creates different groups of them. But of course, you have the possibility to change them. So it’s just some help to kind of kickstart things and say, “Okay, instead of having you organizing all these 200 notes, I did some work for you.” But from there, of course, you can absolutely change everything if you want.
So that’s one thing. The other one that’s currently possible is similarities. So if you imagine you put your place on a note and you want to know the others that are similar to it, fundamentally, it is a way for you to ask, “give me potential connections and things related to this note.” So basically, you select a note, and you click this button, and you get those similar texts… semantically similar. And from there, you basically can create connections, if you want, to other notes. So you get a similarity score for each of them. So, you know how similar they are, and you can decide to connect them or not.
But again, it’s on you. You can choose to connect them or not. And all the features will definitely — AI features — will definitely be built around this philosophy of making you the last decision-maker. Just give you suggestions, just help you make more sense of what you have in the system. But you also have the last word. Probably automatic tagging is not one of those use cases because of this reason, because you cannot really control them. But for example, it could be a named entity or recognition. So, the AI looks into the notes and is able to detect if there is a name, if there is a place, if there is a date, anything like that, and help you, once again, find potential cross-references across the notes. But again, the last decision will always be on you.
Jorge: Yeah. And the way that this manifests for folks who haven’t seen it is like you were saying the system clusters these notes on the graph. Like, the notes are represented as dots with a label, and the system renders them in clusters, and it draws a circle around them, right? And you can label the circles, too, so you can give the clusters names. I’m wondering whether that is a two-way process. In other words, if I take a note out of a cluster and move it to another cluster because I think that it belongs better there, now am I somehow training the algorithm, or does that gesture have any meaning behind the scenes?
Lorenzo: So not yet, but that is definitely something interesting to explore: the possibility to actually train some AI, taking into account also your own representation. Right now, the models are actually trained [using] a completely different corpus of knowledge, but that’s definitely one thing to explore in the future.
Jorge: That question prompts another in my mind, which has to do with the privacy of my information — but not so much from the angle of privacy; I’m interested in the angle of training the AI in that I would expect that the larger the corpus, the more accurate the AI can become. Or that you are at least training it on a wider set of data. Is the AI only spotting patterns based on what I’ve uploaded in my notes? Or is there a way for it to somehow learn from all of Gems’s users?
Lorenzo: Actually, not. The models are basically large language models that are built on a completely unrelated corpus of knowledge that doesn’t come from the platform itself. So, at this stage, it’s just prediction. There is no training involved in the process, so the notes of the users are not really mined. Or not really used to train anything, nor at the personal level or at a more like collaborative level, like taking the notes of all the users and put them together in a big model.
Jorge: Which I’m sure is going to be a relief for folks looking to keep their things more private, right? Like, I’ve always thought that that’s a two edge sword. Because, on the one hand, you do want your stuff to be private. On the other hand, the more data you have, the better you can train these algorithms, right?
Lorenzo: Yeah, definitely.
Jorge: I like this direction that you are the last decision-maker. And I’m assuming that when you were talking about that, you meant you are the last decision-maker when it comes to structuring the information, this notion that somehow notes are grouped into clusters. I’m wondering if the system allows for viewing information in more than one organizational scheme. In other words, when we first started talking, you said that the way that you introduced Gems is in contrast to… let’s call them more traditional information management systems where you have folders or directories, and everything lives in a hierarchy of folders. But obviously, with digital, you can have more than one organizational scheme, right?
Jorge: And I’m wondering how, if any, Gems approaches that. Like, is the organization that it’s proposing and that I’m tweaking, is that the only way to look at things or are there ways to arrange things by multiple schemas?
The plan forward
Lorenzo: The plan is definitely to introduce many different visualizations. You can think of them as pair of glasses that you wear to actually look at your existing body of knowledge from many different angles. Of course, again, with the help of AI, a concrete use case I can give you is topic modeling. So if you want to know, given your thousand notes in the system, how they are correlated, you can actually do that and ask the system to see them plotted in a way that some of them are close together because they talk about the same thing, again. And that could be interesting to do in a separate window. And again, from there, maybe give the user the possibility to do some thinking and change some things that then are reflected on the main view.
Now, this is not currently possible, but it’s definitely on the roadmap, like many other different use cases. Again, the goal is definitely to give many different filters to look at your knowledge from many different angles. And I ultimately think that this is the way to take full advantage of computers and digital knowledge. Because one of the things with folders you mentioned, I think that the folder structure is fundamentally like a translation of what we did for centuries with shelves and filing cabinets. And the reason we do that in the early days of computer science was probably that we had really this challenge of making these new tools that were computers, familiar to people, and the easiest way to make it familiar was taking those kinds of ideas, and translate them into the computer.
Now, what happens is that it doesn’t really take full advantage of the inherent feature of the tool — of the computer. Because it’s not the best way to organize information when it comes to finding patterns, creating cross-references, surface ideas, or anything else, so, I will say about this, there is this zettelkasten thing that people in the note-taking field are familiar with by Niklas Luhmann, who’s probably the most audacious endeavor in this regard, trying to actually introduce this kind of cross-referencing and surfacing of ideas, despite the limits of filing cabinets and shelves. But again, it’s really hard. You really go against the fundamental limit of the nature of the tool, which actually works really great with computers. But again, we still pay for that decision in the early days of the computer. So we still are really hierarchical. Still really hierarchical, and so knowledge is still kind of stuck into this tract.
What I’m trying to do with this tool is really trying to make it possible to simplify this and take full advantage of the tool. And I think that this is really also connected to the ultimate vision of the founders of the field. So if you think about people like Alan Kay or Douglas Engelbart, and even Steve Jobs, they were really working on these tools to make them like tools to extend human cognition and augment human thinking. Steve Jobs used to say the personal computer was the bicycle for the mind. So, it was really like a tool to augment the efficiency and creativity of thinking. And this is really like the direction that I want to go to. And I think in the last 20 years, probably? Computer science, in general, didn’t really pursue this vision anymore. They changed their direction, which is fine. But still, I think that there is really something to do again in the field.
Smart note-taking jobs-to-be-done
Jorge: Sometimes, when you’re doing something like this, you have the idea, but you don’t know exactly what it’s going to be for. And when you read advice, people give you advice about this, they say, “you have to focus. You have to focus your… the audience,” you know? It’s like, “if you say that it’s for everybody, then it’s for nobody!” And all this stuff, right? But it’s like, “well, but this is a tool, and it’s hard to know who’s going to adopt it.” And you don’t want to optimize prematurely, right?
Lorenzo: Yeah. Again, I think it’s really a challenge of this personal knowledge management field. That’s what I have observed. If you ask, for example, ten different people why they use Notion — it’s a very popular, widespread tool. You will probably hear, if not ten, but like five to six different stories. Like, I use it for doing this, and there is this specific feature that really helped me. So the so-called “jobs to be done framework” really it’s hard to figure out in these types of tools. Because, fundamentally, there are tools that, again, define some sort of primitives. — like a way for you to work — but then, they’re very, extremely flexible. And in, in this flexibility, you can actually find many different ways to make it work for you, which is in total contradiction with what they tell you about the target and be very specific, et cetera. I don’t know; a CRM would be like… it does one thing. You know what it does. There is one use case.
In this case, there are many multiple different use cases, which make it really fascinating as a field. But on the other side, there is this challenge. And what I’m trying to do right now is actually auto-imposing to myself a sort of target or use case and testing many different use cases and trying to see what are the most promising, what are the most exciting — where people get most excited and find it useful — and trying to double down on that. I’m not sure; I’m not saying it’s the right approach. It’s just one approach, and it’s the one that I’m trying to follow.
Jorge: Well, you mentioned focusing on content creators, which makes a lot of sense, given the tool’s primitives, as you said. — particularly people who are creating content that is text-driven because this is a text-driven tool, right?
Jorge: I’m actually going to push back on the impulse to want to focus because I do think that these tools are general thinking tools. And we deal with information and knowledge throughout our lives. And the notes that we take — the knowledge that we aggregate over the course of our lifetime — can be useful for many different things, right? It’s part of our thinking apparatus. And we use thinking for all sorts of things. For creating, for learning, for becoming better parents, better citizens. Just for, you know… it’s part of our development. And I get the feeling that the impulse to want to focus these things is driven by the fact that they are competing in the marketplace with other tools that more naturally lend themselves to focusing. Whereas I think that these are general thinking tools.
Jorge: And we don’t have good analogies for that. Like you said, like, you know, we have all these metaphors that come from the physical world. You know, shelves and folders and all these things. We don’t have good metaphors for tools that augment our thinking in the way that you’re describing here.
Lorenzo: Yes. That’s definitely one of the biggest challenges that I see in the field in general. And I think that probably I wasn’t there to confirm, but I guess it was one of the challenges that also people who started with the, for example, the personal computer had because there were fundamentally some completely new tools that really introduced completely new ways of thinking and interaction.
And again, they really tried to make it easy. With scale morphic design and trying to make it as familiar as possible to people. Taking real-world objects: the desktop, the trash — all these types of things from the real world to lower the barrier to people. But still, I imagine that it was a huge challenge for them as well. Today, with a computer, you can do whatever you want. But I guess in the early days, they really had to think deeply about how do we make this thing useful for a very targeted and specific niche. And so I guess they had the same sort of issues. Let’s find a use case, a way to make this really useful for some people, and then we’ll figure out… and eventually they became the everything tools.
But I think that it didn’t happen the first day. And I see the same thing happening with this type of tool. So I really see some sort of parallel and connection with this. I think that personal computers were really from the industrial revolution. And now this is the same thing in the information revolution. But we’re really into kind of the same direction for this information revolution with this type of tools. We’re really headed to kind of the same potential revolutionary, you know, power and effect, in this field. That’s what really makes me excited as well.
Jorge: Yeah, I agree. I think that a lot of people think that it’s all spoken for when it comes to computers. I think that it’s still early days.
Lorenzo: Absolutely! Very early days.
Jorge: We see tools in the market such as Roam and Obsidian and Gems that are finally kind of breaking free of some of those metaphors. Like, it feels like we’ve reached a level of maturity where people don’t necessarily need the crutches of the metaphors that we bring over from the real world. It’s like, there’s a couple of generations now that could be called “digitally native.” And we don’t need the metaphors; we can transcend them. And if you transcend the metaphors, all of a sudden, the possibilities open up considerably, right? There’s so much more you can do.
Lorenzo: Absolutely. Absolutely. And I will add that probably if you look at this structure of the internet, this has always existed. Like what we’re fundamentally doing is going towards a more network kind of thinking. So instead of saying, “let’s categorize these things on many different shelves, many different categories, and many different boxes,” let’s free everything and make them communicate together in a networked way. And if you think about that, the internet, the structure of the internet, was actually dead from day zero.
So if you look at Google, or you look at Wikipedia, it’s actually that. You navigate into a graph, even if you don’t see it. The pages are connected, and you are on a graph. What happened, though, is that the personal level didn’t really happen. So, at the personal level, on a personal computer, or even in our shared folders and spaces where we work, we are still into that kind of early-day structure of folders for some reason. So, that’s probably another way to look at it. It’s really like, how do we bring the structure of the internet we are used to and works very well to navigate information and explore things, at the personal level where we still have folders and notebooks and these types of things?
So I think that ultimately, one of the things these tools are trying to do is to make this transition. And make it possible — even at the personal level — to take advantage of network thinking, which is really much more powerful to think. And it’s really where computers are much better than the historical information structure we have.
Jorge: Well, Lorenzo, that seems like a really great place to wrap up the conversation. It feels like a lofty aspiration. Where can folks follow up with you?
Lorenzo: So I am on Twitter. My Twitter handle is @ittaboba. You can find me if you type my name on LinkedIn. I have my personal website, which is called Ittaboba. And then, you can also check out my tool at gemsnotes.app.
Jorge: And folks can sign up for an account. Is that right?
Lorenzo: Yes! They can definitely join the waitlist. And I will be very pleased to meet them and actually let them try the tool.
Jorge: Well, fantastic. I wish you much success and luck with the development of Gems Notes. I can’t wait to see how it develops.
Lorenzo: Thanks, Jorge. Thank you.
Jorge: Alright, thank you.