Hacker News Reader

Comment by SirensOfTitan | Hacker News Reader

With all of the exciting directions knowledge management has been moving in lately, it's disheartening to see that Notion's product vision is already so uninspired.

Particular in the corporate space, the problem of finding relevant knowledge (and keeping that knowledge up to date) is a really hard problem. I think competitors like Mem are going to eat Notion's lunch here (categorization and perhaps tagging of stale content is a much more reasonable application of ML than this).

... further, Notion's performance is so absolutely awful that my very small company had to stop using it. Latency is incredibly relevant to note taking and writing apps.

The marketing page example is also so contrived: "Write a blog post introducing Notion's new AI feature." How would the AI even know what that feature even is? Where's the context? It seems like this just proposes static solutions to dynamic problems.

View on HN

Signal boosting in case anyone from notion is reading: please address performance, it is practically the only thing I care about. I love the product but it is almost unusably slow with my tiny amount of data. Cache more stuff on my device, avoid blocking network calls where possible, etc, please!

Yes, performance is absolutely awful. Not just individual (it’s a heavy react app), but it’s not really possible to use search when you have more than a couple hundred pages. I swear their test fixture is a site with no more than 5 pages and they all use Mac Pros so they don’t notice any performance issues.

I have a single markdown file that stores all my work notes for the year, and recently I imported it into Notion. My partner raves about how much she likes Notion so I thought I'd try it.

The files can take up to three seconds to load at times, and seemingly caches nothing on my phone so when I want to look at my notes on the tube, I can't load it.

Yes, I’ve been using Notion for years, but recently started migrating everything to Obsidian because it feels so much faster.

(Notion engineer here) I'd love to hear more! We're focusing a lot on performance and want to hear where it hurts the most. If you (or anyone else reading) are comfortable sharing, drop me a line at carlo AT makenotion DOT com

Agreed. I use a few filters which take seconds to update, sometimes disappearing and then redrawing on any change in the tables. And recently some of my filters and views have been changing, leading me to have to tweak and adjust some of my dashboard views, shaking my confidence.

The filters and relations are so powerful and I really love the flexibility: example, “show only tasks linked to projects with an active status.” But when they keep disappearing and redrawing, I just feel slowed down in the app lately.

seems like they wanted to launch an "AI" feature as fast as possible, so they took the easiest route which is basically just creating a UI wrapper around GPT-3 for generic text generation

that's what it looks like to me at least, I don't see anything here that 50 other generic GPT-3 tools don't already do. It's just embedded in Notion

I think they would consider it a success if the existence of that tool drives people to spend more time in Notion

Doesn't have to be revolutionary. It helps to be "in just 1 place"

true, plus I imagine it gives them potential for some usage based pricing to increase revenue from power users

I like to call this "AI (or ML) as a blunt instrument". You just plug in the model, map inputs to outputs in the most straightforward way.

In some cases that can work, but generally you want to massage the inputs and outputs a lot more than that.

Agreed about the performance. I even think this feature is cool (I don't use Copilot to write code, but happily pay the $120/year just for the time saved writing docs (the kind colocated with code, like rustdoc, JSDoc, etc) and unit tests, so I'd like a trained model fuzzy logic bot helping me with notes/docs too).

But there is no feature at all that could excite me about Notion now unless they fixed the performance first.

I stopped using it due to performance also.

> Particular in the corporate space, the problem of finding relevant knowledge

This really is incredibly hard. A large fortune 500 company I worked for this summer had a whopping 35000 confluence pages. There is no navigating a knowledge base that vast. You couldn't even clean it up, because there was no way to know if someone would need some piece of fringe information on a rarely visited page any time in the future.

Wonderful if everything is in Confluence though. Most companies have knowledge in various forms free floating through e-mail, sharepoint, sharefile, onedrive, teams, Onenote, wikis and so on.

I worked at a place that used confluence as a knowledge base. There was nothing that a git repo of txt files and grep couldn't do better. The search function was useless on top of being incredibly slow. Maybe they were not using it correctly I don't know.

Git, text files, and grep are a horrible user experience for anyone who isn't already experienced with those tools.

That's what wikis are for but IMHO wikis are annoying to setup and customize, and good luck convincing people to document their knowledge

I wish we had a librarian on staff on my team to just keep data organized

This is… a really good idea? Do any tech companies do this? I’ve had tech writers on staff but never a trained librarian.

It's also a matter of vocabulary and writing about your knowledge. It's almost impossible to find somebody that solved the same problem you currently have, because somebody that solved it didn't write about it. We're kinda OK searching code nowadays, but we're not able to discover architectures or abstractions properly.

I'd bet my had that most of it was meeting notes

I desperately wanted to tackle knowledge retrieval problems when I worked at Slack, to the point where I built prototypes during nights/weekends to try and get internal traction. Slack easily has one of the most valuable and important datasets in the world for training corporate agents.

IMO, the most useful rejects were:

- Subscribe to a search query. Get an RSS-like digest of Slack messages, external document updates, CRM interactions, etc.

- Auto-responder I called "escal.ai," named after Slack's internal naming convention for help/support channels. These channels were prefixed with #escal-<subject>. I built a variation of https://arxiv.org/pdf/1901.05415.pdf to handle asymmetric retrieval tasks, where the answer to a question was typically a Slack post in another channel.

I think I find the examples they give actually quite worrying in an age where we are dealing with hard problems around moderation.

"10 ways async communication is more productive" - alright, so you have an AI answering a leading question. Are you going to take the output as given and publish it as-is or spend time verifying and tweaking it? If you do the latter, how much productivity have you gained from rewriting a generated post compared to writing your own draft?

For automatic summarising and translating, okay cool, that really could be handy but this is almost commodity tech at this point.

I don't like my negative, pessimistic response here but I don't need tools that make it easier to write or create, I need them to keep things organised and up to date as you've mentioned. I need a better search that can help me find something that already exists, or actually point me to that article if I'm writing something that starts to look like a duplicate.

Can the AI say "this 'setting up the environment' post looks really similar to [some other setting up the environment post] which was written 6 months ago; is that what you're looking for/do you want to edit that?"

"If you do the latter, how much productivity have you gained from rewriting a generated post compared to writing your own draft?"

In my experience, AI providing the activation energy to get a first draft is highly valuable.

Particular in the corporate space

This is EXACTLY what people in the corporate space need! AI to help them generate nonsense corporate-speak so they can get back to work!

I'm still surprised that no one has just mined all of a company's e-mails to make a database:

- strip off all attachments --- these are archived on a central server w/ version control and links are passed around

- all e-mails are converted into threaded discussions in an hierarchical organization

Then, use an expert system to review and tag all messages and facilitate finding the correct forum thread/file.

I was trying to fix the problem of getting relevant knowledge and staying up-to-date with my startup https://www.gainknowhow.com/ . It never took off organically and I only sold it to a couple companies. My core idea in GainKnowHow was to structure data as a graph of skills. I think the graph data structure to store knowledge is better than how a lot of tools store knowledge today. I still think it is a great way to store knowledge and use it myself to train people, but the graph data structure was a leap too far for many people who implement knowledge systems at companies.

Some time ago, a Notion engineer was telling HN that their old principle of hiring slowly was blocking growth. This was around the same time Notion started launching half-baked feature like https://www.notion.so/help/guides/creating-links-and-backlin... which, last time I checked, still doesn't have the smoothness of Roam or Obsidian.

My gut says that every engineer they hired in the name of growth is trying to get their own little feature in to show impact. They prematurely scaled the company and it shows in the loss of focus.

FWIW my small company has zero issues using Notion, what you're describing sounds like a situation specific to your use case, not a generic Notion issue.

But more fundamentally, it sounds like you just don't believe the product works, which is fair, just not a very helpful comment. "How would the AI even know what that feature even is?" by learning about how its documented in the rest of Notion.

Fundamentally though, your comment has serious "Uber? No way that's going to work! Why would I get into a stranger's car?" vibes.

>Particular in the corporate space, the problem of finding relevant knowledge

I'm eagerly waiting for the day that I don't get frustrated with Notion's search. I'm not sure if I actually even bother with the search that much these days. I just try to drill down the page structure to find what looks relevant.

They could have used HuggingFace models to embed texts, put them in a vector search engine and serve results.

- models here: https://sbert.net/docs/pretrained_models.html

- fast vector search here: annoy, faiss, Milvus, Elastic, Pinecone

It's like the hello world of current day NLP.

This is exactly what Mem does using Pinecone and OpenAI: https://www.pinecone.io/learn/mem-semantic-search/

Almost any self-hosted ANN library will do if you're dealing with <1M vectors, but the indexes at these note-based applications tend to balloon very quickly into the hundreds of millions and billions. Given the popularity of Notion I expect their number of (would-be) vectors to be monstrously big. That's where rolling your own gets painful (or downright not feasible) and where Pinecone shines.[1]

[1] https://www.pinecone.io - Disclosure: I work there.

Thanks for the shout-out! For folks interested in playing around with vector and/or hybrid search: Milvus is open-source (https://github.com/milvus-io/milvus).

> How would the AI even know what that feature even is? Where's the context? It seems like this just proposes static solutions to dynamic problems.

The whole feature is kinda cynical in that it already kind of implies that because nobody will read it, you shouldn't put too much effort into writing it either.

They need to do offline first with online sync, rather than online first. But they simply won't do it.

What is the difference?

Writing goes first into your local storage, then gets send to your remote storage, rather than vice versa. Same with searching, search local storage rather than remote one. Everything would load faster, latency would be less, could use it offline and so on.

On the other hand, Notion wouldn't have full control anymore, usage stats might be hard to get when some users are offline, and so on, so unlikely Notion would do such a thing. Certainly their investors wouldn't like it either.

(I work at Notion)

> Writing goes first into your local storage, then gets send to your remote storage

We do this already; writes are queued in SQLite or IndexedDB first and then asynchronously batched and streamed to the server. Our apps (Mac/Windows/iOS/Android) cache pages locally, and all platforms search a subset of documents locally in addition to server-side search. We’re actively improving this area as well.

Can you shed some light on why something like Google Docs is able to operate this way but not experience the lag or latency that your app does?

Typing latency in Notion is generally worse than Google Docs because our renderer & input handling is less efficient than Docs. Docs has substantial lead over Coda, Notion, Dropbox Paper, etc because they use a custom layout engine for presentation that is entirely separate from the input surface. Notion, Paper, etc instead use the Web’s built-in but flawed API for rich text editing called “ContentEditable”. In ContentEditable, generally the rendered output DOM nodes are also the input surface. This is much less expensive to implement, but comes with a bunch of limitations and constraints.

Thank you for the response, it was very illuminating. Is the reason you choose the built-in API instead of building a custom layout engine just due to complexity?

The latency. Accessing browser storage is slow, live syncing is slow.

Maybe ask the Google Docs team? They seem to have low latency online-first doc management figured out.

The only thing that is making me not switch to Mem is lack of tables. I really heavily on tables for notes. And that has not yet been implemented in Mem.

What did your company end up using instead of Notion?

Reply on news.ycombinator.com