It's very fun to play around with, there seems to be value somewhere.

I wonder if you could make a video of how an individual could use it effectively in a pragmatic sense e.g. networking, research

(after more playing around I get the UX, very powerful indeed)

(feedback: add some controls for filtering (dates etc), relevancy, color islands slightly, just general UX improvements and I would probably pay a few dollars a month for this)

(feedback: when looking at a user, you have to zoom really far in before results start appearing. it should probably show labels/results if less than X amount of data points are in your viewport.)

(feedback: Is there "Explore more users" section the more related users to the current profile? It would seem so, but not immediately clear. And if it is, none of my related users have me as a related user in their profile.

===

This is really inspiring work, hats off!

Thanks so much and for the great feedback!

OK, I'm going to just be upfront here and admit that I'm an idiot. This is the third time I've seen this link come up on HN, and each time I've checked it out, but I can't really figure out how to interpret what I'm seeing there in a way that is useful.

How can that map be used to determine who is knowledgeable about what? Looking up myself, I can't connect what I see with my own areas of knowledge.

I think I need an ELI5 for this. Again, this isn't a criticism of the effort at all, it's a public admission of my own ignorance.

> How can that map be used to determine who is knowledgeable about what?

To be fair.. you can’t. Unless you think that mentioning specific words or phrases associated with a specific topic a high number of times means that you’re knowledgeable at it.

Take everything a user says and encode it into some vector. That's "embedding."

Then you encode a search query the same way and return the nearest N vectors/users. That's vector search.

It seems they did the above for HN comments.

Right, I understand all that. I just don't understand how to interpret the resulting map (or how to formulate a search) to provide useful information, specifically "who knows what".

Thank you for being so clear on where we are failing you. We're basically computing your semantic space as compared to the whole community. The map shows the relevant instances in that space. As you browse your concepts the space changes and we link to the relevant comments from the associated threads. Who knows what is prioritizing the user and the semantics / voice of that user amidst the noise of the whole community speaking at once.

We'd love to hear your feedback as you play with it more and we improve all of the above.

It obviously isnt intuitive, but I suppose one could learn to read it over time. For example, there is a lobe at (5, 10) that lights up like a Christmas tree if a user was engaged with Covid vaccines and mandate discussion. Of course, it seems like there is no discrimination between the different stances a user took, or the veracity of their commentary.

Is the knowledge graph weighted more to submissions than comments? I'm surprised to see how generalized the keywords associated with my user are "Company, Product, Market, Part, Article, Issue, Bit".

I'll admit, I am a generalist, and I comment more than I post, but I would have thought/hoped my knowledge base would trend towards neurotech, sleep, mental health, health, wellness, etc, as I feel that is the "less general" stuff I comment on.

Though I recognize this is how I like to see myself, I do often comment on business models, branding, etc etc, which are areas I have interest in.

I wonder if topics that have more activity from a user, but are in a less popular topic area would improve this?

Right now it does end up representing quantity + quality than just quantity alone. How can we do better?

I'm thinking you'd have a huge volume of people with similar common attributes, and if these most common attributes/keywords suggest a more baseline understanding of the community as a whole.

I'm not sure how you surface the more nuanced understanding of each users knowledge or interests.

The few things I'm thinking are 1) more weight to more recent content than earlier. We change and grow as people. I used to be in the music industry, then 3d mapping, and now neurotech. My music experience is now old, but neuro is fresh. That may be more interesting/valuable.

2) commonality with the viewing user. Where do we overlap, and again, where do we overlap where others don't? How often would a person be searching for someone whom they don't have overlap with? We are likely interested in the same things. Though perhaps somebody who knows nothing about marketing, is looking for marketing help, so I'm not sure how you surface that.

I guess I'm wondering what you imagine the ultimate use case being? Or the early adopter use case. You may be prioritizing the algorithm for uses I'm not expecting.

How are you measuring quality?

Looking at my own profile the words (topics i know about?) seem sort of generic. "issue", "company" - not sure what to make of that other than I assume those are words I use a lot in comments?

We'll take a look at how to do better, thanks!

I feel targeted when you call me a "system server user network number service code issue" ... not that it's inaccurate ;)

Ha, exactly!

When you click on some keyword it's actually to go back from that keyword.

It would be nice if map showed titles of all articles on extreme close up.

Overall it's interesting tool but I think for people with small amount of comments this is not much informative.

The visualization of the space is excellent.

Thank you!

Thanks for making and sharing!

Could you share some info about how you generated the 2D space embedding/visualization?

Thanks! I'll ask Wilson to say more in our next update.