Apr 20, 2022

Personal feeds

Intriguing idea from Matt Webb, what if people got an MRI scan regularly and could then run some algorithms on the results?

Imagine you get a full-body MRI every 6 months. Nothing wrong necessarily, it’s just like going to the dental hygienist. Then 100s of different machine learning models run, one looking for a particular liver condition, one looking at another organ, another looking for such-and-such anomaly elsewhere, etc. It’s purely precautionary; a way to pick up issues before they get serious; a Check Engine light for your body. You’d get a notification on your phone the next morning.

He goes a step further, let’s say it’s not just something one company or hospital does, but an ecosystem looking for different things?

So I see something that is more like a software ecosystem. As a consumer, you pay (or your insurance pays) for the twice-yearly scan. A portion of that fee gets divided amongst the hundreds of separate companies that provide computer vision modules that run across your full-body image, like paying for Spotify streams.

Which reminded me of a post by Jon Evans that I’ve mentioned repeatedly and includes the idea of third party developers of feeds.

I have a not-especially-modest proposal for how to solve almost all of Twitter’s problems. It’s very simple: let third-party developers build feeds. Extend their API and allow external developers to design, and users to install, custom tabs with custom feeds. So a user’s Twitter interface could include the Twitter-built Moments tab, if for some demented reason they actually wanted that … or, instead, an NBA fan who lives in Toronto could have a custom-built NBA feed, and a custom-built Toronto feed. […]

Or the StockTwits feed. Or the Nuzzel feed. Etc etc etc. All built by third parties– who share the income from “Promoted Tweets” within their feeds. Sure, give new users a default, Twitter-built curated feed. But also let them choose from a “Featured Feeds” list … or, better yet, from the Feed Store. […]

In short: make feeds Twitter’s apps.

To me, those two things are (very) roughly the same thing. “Here’s a certain corpus of data I’ve defined, run an algorithm on it, give me some results.”

Corpus + Intent/task + Maths = Stream/feed

We can then further define each of these.


  • All the data. The whole internet for a search engine, all the tweets for Twitter, everyone’s scan data in Matt’s example.
  • My data. My browsing history or my bookmarks, or all my subscriptions or lists for Twitter, all of my scan data in Matt’s example.
  • Their pick of data. Waverly (see below) is a great example of this, their recommendations are chosen from a corpus of quality sources they picked themselves, not the internet at large.


  • My intent. What I specifically asked for. A specific thing I asked for.
  • The platform’s intent. Facepalm whose aim is to keep me there, not to give me what I’d really ask for, were I asked.
  • Advertisers’ intent. Roughly the same as above but different.


Same warning and opportunities as usual with algorithms; are they fair, transparent, understandable, can you ask how they work, etc. In here, I also dump widely different things, like a feed that just shows tweets with links (just an if/then statement), a more complex calculation that weighs various arguments, or an ML model that spots tumours.


  • Exactly what I asked for or would expect.
  • Something else that feels like what I want but just wants to nudge me into something.
  • Matt proposed a couple of guard rails, but one can easily imagine a company emphasizing something in the results to nudge me into a more expensive service.


Feedly is kind of close to Evans’ proposal above, I give their “AI” Leo access to my Twitter lists, a bunch of RSS feeds (the corpus), some priorities (my intent), and it returns me a subset of that corpus as a much smaller/densely interesting feed.


Since it’s only on mobile, I haven’t tested has thoroughly as I’d like, but the really promising and intriguing idea is that you write your own “Waves,” which are basically just plain language little stories about what I want to read (intent).


For those who’ve been around long enough, we might see the above examples as early hints of the famed “agents” from the earlier web, during the last AI winter. Some kind of service to whom you could say something like: “Find me a great hotel in Florence, you know my tastes, then book me a ticket for the best time and price around the middle of summer.”

If you squint just right, all the examples and posts above can be seen as more limited and more targeted “agents.” I picked Feedly and Waverly as examples because I use them today for slightly different reasons. Both are smartish services using maths to give me some good stuff, centaur-like.

I’m always talking about keeping a smaller haystack to search through, that’s kind of what Feedly does, by using only my sources, and the plain language intent in Waverly could be very powerful. Imagine an app or combination of services that could work something like this:

  • Corpus. Use everything I’ve bookmarked or written, everything written by the 150 people on this Twitter list, and every first-degree contact on LinkedIn. Or even better, every member of the Near Future Lab and Sentiers Discord communities.
  • Intent. Here’s a 400-word document describing what I’m looking for.
  • Maths. That company’s algorithm, audited for fairness and transparency by some third party foundation.
  • Stream. Put everything about 90% confidence in an RSS feed and email me the top three items every morning.

That’s actually 80% doable right now for this kind of purpose, and will likely be possible for entirely different domains over the next decade. The auditing foundation (or something similar), and companies that actually fulfill those requirements is probably the hardest part, not the technology itself.