Automated organization

Recent developments in the note-taking space introduce AI-powered automation when it comes to organization of notes. Examples are, and others. I think notions moves down that path as well.

Having a manual workflow that works quite well for me, I am hesitant. Wouldn’t the lack of a hand-crafted organization scheme give up on the merits of the generation effect? On the other hand, building an efficient workflow is a heavy task. Is it possible to just skip that step by means of AI?

What do you think about that? Do you have any first-hand experiences with that sorta tech?

1 Like

It’s an interesting question, and I’ll try to answer with a tangential anecdote.

In the early days of Supernotes, we were targeting students, and one thing that came up a lot at the time (esp when we were talking to VCs) was the idea of building a notecard marketplace. The basic idea being that students could buy and sell their notes to each other on Supernotes, which would be the ideal platform for this given how easily you can integrate the cards of others into your own personal “system” on Supernotes. And clearly there is actually a market for this, as students are buying cliff notes / study guides / etc all the time.

However we were unenthusiastic about the idea, because (having spent a lot of time researching the problem), we knew that this was behavior was probably counter-productive for most students. Struggling students see the student that is doing quite well and after a surreptitious look at the successful student’s notes think “wow, those notes look great. They’re so thorough and so well-organized. If only I could study from those notes I could probably do just as well. I’ll ask if I can photocopy their notes tomorrow after class”. Of course that’s not what happens in practice. The reason the other student is doing well is in part because they are creating those notes themselves, and forcing themselves to synthesize and internalize that information as part of the note-creation process.

Many of our current users don’t know this, but this is actually the origin of the card length limit on Supernotes as well. We designed it in the hope that it would encourage people to think more about how they are structuring content, and that by breaking it apart, users would be better able to retain the information that they are recording on Supernotes without actually having to rely on Supernotes itself.

Anyway, I think the parallels with using AI for the same purpose are strong. For various uses, the current crop of LLMs can be genuinely helpful, e.g. if you want to turn a series of bullet points into a polite email quickly. However when it comes to anything like a “second brain”, I think these features often give the appearance of being helpful while actually doing the opposite, just like copying the notes off the smart kid. In general, the goal of Supernotes will always be to try to build features that really truly make you more productive, instead of creating a productive facade. We actually built some GPT-3 powered features two years ago, but ultimately decided not to release them as we found the signal-to-noise ratio was much too low for our comfort.

That being said, we do think there are some places that LLM integration would be genuinely helpful (esp now that there are more and more models available and the best ones actually perform quite well), we just want to be very careful about where and why we add such features. Clearly many other products in the space feel differently, but bandwagoning has never been in our genes.

So please continue to suggest ML-powered features, as we are always ready to try out new ideas! Now that we have the “Feature Preview” system, we will be a bit more likely to just push things out and see if people like them rather than relying entirely on our own standards of quality to judge if something should be released.


Thanks for the insight! The idea of simply buying into the structure others have created is a nice illustration of what I think will not work. When learning, we use our hippocampus a lot - that part of the brain that is capable of remembering where something is, in a spacial sense of the word (as in behind that tree is my cave). If structure is not attained, but given (as with the notecard marketplace, or with AI), then I believe it is pretty much worthless.

That being said, the industry seems to be very appealed by the idea of using AI to organize things. I wonder why that is so, as I feel, for the above reasons, that there is not much potential in this. But I might be wrong. I’m not part of that industry, after all, and probably I am just missing out on a novel development - hence my interest in personal experiences in that playing field.

I’m an on-and-off user of mymind. Personally I never found its AI organization effective. The reason I use it is because of its (imo) attractive UI and the ease of capturing information. However, I cannot imagine it becoming my second brain. It is a private Pinterest board at best.

I think the mymind team knows the limitation of ML-powered organization as well. Recently it introduced the “Collections” feature, which is somewhat like folders defined by users themselves.


MyMind is actually pretty bad at search and organization. I use it as a glorified bookmarking tool rather than a second brain.

1 Like

My current thinking is that anything the user will almost certainly never do themselves is a suitable thing to do with AI, while anything the user is likely to do on their own without AI should probably not be automated.

Some examples:

  1. Using object recognition to categorize images. Basically no users will do this on their own, even if you give them a system for simple image tagging. AI is actually pretty good at this these days, so it is probably a suitable candidate for that type of automation. This is something MyMind does that is Probably Good™.
  2. Structuring your content with links / hierarchy. Users are definitely going to be doing this on their own if it is not automated for them, so realistically you want them to keep doing it in order to actually have the structure stick in their mind and help with recall when not using the platform.

Again, this is different from other software where I think “automate anything you can” is a reasonable goal in many cases.


Interesting take, I do agree. To add to this, I think it is a good idea to define a clear responsibility for any particular data set, that determines some field is filled either by manual or by programatical input.

Notably, as soon as there is an AI that fiddles with your data, you are no longer in control. I think that this amounts to a significantly different feel, as you don’t know if or to what extent you can trust your system.

To steer a middle course, I have always liked solutions that only suggest things (would you like to add the tag #tools?), but require manual approval. This is a way that adds value, but at the same time leaves the user in control.