If you want to get emails when we post new product updates, you can sign up here!
Tweetscape Product Update, 07/29/2022
Over the last year, I’ve often found myself thinking “The hard part is figuring out what people want. Building it would be easy.” And that sounds nice, until you figure out what people want, and it turns out that it’s actually very hard to build…
Well, here we are. It turns out that taking the sprawling mess that makes twitter, twitter, and grouping it in ways that are useful, interesting and sensible is actually… pretty hard. Update: Julian corrected me. He said that it is fun. We're both right. It is the fun kind of hard.
Get a Feel for Streams
I’m a big fan of the image Julian created that I shared in a tweet this week:
It gives a great feel of the type of experience we want to build. Current twitter keeps you stuck in A, but many of us are dying for a better way to segment that into focused streams that fit our interests.
i want my twitter to feel like this pic.twitter.com/DwEV2YmjAs— Nick Torba (@nicktorba) July 28, 2022
Of course, there are lists. But the more people I talk to, the more people I meet who have tried and failed at creating lists that work for them.
I think that’s what lists are for, but I don’t use them.— danielle (@danielleboccell) July 28, 2022
Another friend explained all the work it would take to make this a reality now:
This requires "completeness without overlap", aka ability to see that the lists I've put ppl in 1. Put everyone in at least one list, 2. Put nobody in 2 lists. UX now makes this a nightmare. You have to check one at a time. It's like 5 clicks. Hundreds of accounts.— Airship⚛️Evangelist (@repost_offender) July 29, 2022
And here is another reply with my take on why lists and communities don’t meet the mark.
for lists, the UX is bad and it shows by how few people use them, even when they "feel like they should"— Nick Torba (@nicktorba) July 29, 2022
communities are nice in theory, but applying top-down organization is anti-twitter. the solution needs to be bottoms up https://t.co/rmGCJVFfg3
I hadn't been tweeting much recently. It felt good to tap back into the hive mind and get useful thoughts and feedbacks on our thoughts. I should probably keep doing that...
Last week, I said I was going to deploy an app to show some simple account recommendation methods. Here it is!
This “experiment” is nice, but it’s not really an experiment. It’s a fun little tool with some bugs.
However, it did generate some excitement:
Oh the account recommendation method is suuuuppeeer cool and stands out a lot— pranab (@altsanabo) July 28, 2022
Both "min seed acct followed by" and "multiple interactions". Seemed relevant for the longevity example!
I tried to clearly lay out the logic and functionality in the app, but it still needs a lot of work. If you have any thoughts, as always, DM ME!
I still think there are a lot of simple methods (I'd just call them "low hanging fruits", but that term has been ruined for me) for useful like this that people will be happy to use, especially when the mechanisms that make them work are transparent.
Andre’s Grind: Advanced NLP
Andre spent this week using a doc similarity matrix between the tweets of different groups of users to make account recommendations - a much different strategy than the following and interaction based methods I used above.
This process generated many more questions than answers, but in the good way. Here are a couple of images Julian created to help visualize our current thinking.
In addition, here are some of the questions we’re thinking about, in no particular order:
- how large of a dataset do we need?
- what can we use as labels?
- current lists?
- current communities?
- What should we use as the base unit for analysis
- obvious answer: tweets
- what’s the minimum length a tweet should be to be included?
- less obvious:
- obvious answer: tweets
- Should we use keyword filters to make datasets better for specific tasks?
- How much should we consider external context of a tweet?
- media and external urls
- referenced tweets
An important part of our work next week is creating a list of detailed questions we want to answer and experiments we want to run to test them.
These will range from more simple methods, like pulling the overlapping keywords between users, to more advanced methods using large language models.
On top of creating a better defined list of questions, we are also going to outline clear, intentional experiments. The point of the small experiments is to put our different hypotheses to the test. To do that reliably, we need to design our experiments in a way that makes it clear if the users were able to accomplish a specific goal.
Next weeks product update will have more details on how those experiments will be structured!
My friend Kenny sent me the Instagram equivalent of Tweetscape this week - Un1feed. It’s looks like a very cool project.
What surprised me was how similar the copy on their website is to the copy we used on our first landing page mockup.
Custom Feeds, create feeds to match your interests
streams allow you to efficiently organize the content you’re interested in
Easily explore topics you want to learn more about
I’m gonna reach out to their team soon to get to know them and learn more about how they are thinking about the space.
bye, bye, bye
That’s it for this week. See you next time!
Roote has dope merch. Here’s a pic of me being a cool kid with our bentoism quadrant shirt.
@roote_ definitely has my favorite swag of any company i've worked for pic.twitter.com/l0Fl6irUql— Nick Torba (@nicktorba) July 29, 2022
Hit me up if you want one! We have more!