Confession: Artifact's Future Is Brighter
Maybe a machine-learning news feed is actually quite good after all.
The Teardown
Thursday :: March 2nd, 2023
In my last newsletter, I wrote about Instagram cofounders Kevin Systrom and Mike Krueger’s new news-feed startup Artifact:
Artifact’s innovation is, I think, to make friends better at suggesting interesting content to each other.
Artifact’s launch suggests two things should happen:
The app will definitely suggest things I’ll read that I don’t yet know about
I’ll engage my friends about these things
It is the second point that doesn’t quite land for me. Today and looking at the foreseeable future, I’ll share the stuff that I find interesting over group message, text, or other targeted and personal medium. My friends are already there, over text, over group message, and etc. Suggesting that they join another app or community is a steep uphill effort. So, if it’s hard for me to push my friends into another network to have better conversations, is the app just another algorithmic newsfeed? SV companies often share a utopia in which everyone is perfectly and efficiently connected, but lots of us are doing just fine, thanks.
The two cofounders offered lots of interesting insights into their strategy, decisions, and deep thinking in two interviews - one with Casey Newton at Platformer, and another with Ben Thompson at Stratechery. I’ll dive into some quotes from Ben Thompson’s interview, but first, let’s back up.
Conceptually, you might think of a social network as something simple: people interact with each other using software. Facebook comprehensively iterated that idea, making it possible for you and me to interact with each other, but also with content from organizations. It also mastered the art of stuffing ads in our faces based on all of our network interactions and masterfully engineering ad clicks. However, for quite a long time anyways, Facebook generated lots of signal from people interacting with each other within its software.
TikTok introduced a paradigm shift in the social network concept. It wasn’t about your interactions and connections, but instead, your contributions to the entire corpus of users. And also, information about your raw consumption habits. Did you watch videos with people dancing in church? Dancing on the double-yellow line in the middle of a street? Working on a bicep? Haircare? If so, TikTok showed you more of what you directly indicated you wanted because you watched those videos. So did TikTok, with its technology.
TikTok turned this data into a massively viral business, one that caused lots of copying and iteration by other companies to attempt the same success.
Artifact is about text, something that’s very different from video, but the fundamental technical approach is similar:
Artifact doesn’t care about your friends, or my friends, or really any of our connections, at least in the traditional model I discussed earlier
Artifact place articles in front of you as soon as you open the app
And, Artifact watches your behavior and puts more stuff (i.e. news) in front of you that it’s increasingly confident you like. The more you click and read, the more it knows what else to show you.
Again, it doesn’t matter if Jeff and Zoey (my imaginary friends) are best friends that share news with each other through WhatsApp. It matters if Jeff and Zoey click and read news using Artifact. Artifact will figure out what they like and suggest articles that match their preferences. Kevin Systrom (KS) highlighted the process in the Ben Thompson (BT) interview:
BT: Over the last couple weeks I’m like, “Wow, I’m actually finding stuff that I wouldn’t have found otherwise”, which obviously is the value proposition. So I have a couple questions on that. Number one, how much of that is the function of learning me versus as you just mentioned it actually is the entire corpus of people using it? So now when there’s thousands or tens of thousands or however many there are using it, what’s the function of improvement there? Is it the mass audience or is it me personally?
KS: I’ll touch on this a little bit. From a machine learning standpoint, the idea of cold start is as old as the domain of machine learning. The idea is, okay, you have someone who shows up, how good can you be and how quickly can you be good for them? The problem is, Ben, when you show up, you select a handful of interests. I’m not sure what you selected, but you selected some of them. We probably know you’re in a specific location, meaning general locale, we don’t know exactly where you are, but general locale. So the question really is how quickly can you go from basically blank slate default, which you said wasn’t useful, and I’d agree in general, to more and more useful? Mike has been using this product for, I don’t know, a year and a half, maybe longer, and we literally know every nuance of what Mike’s into. Mike really loves F1 car racing and he’s super into it.
Why does this matter? Well, a publisher such as the New York Times has to decide what will go on its paper. It also has to decide what will go on its website. It’s impossible to change one of those formats once it’s crystalized in print. The other lives in a bit of an ether, informed but not driven entirely by numbers, and also informed by editorial decisions. Those decisions come from humans, like you and me.
Artifact doesn’t editorialize at all, at least not in the human way. There are few more bits from the article that talked about that initial technical cold start vs. data-informed suggestions:
The interesting part here is that with lots of data, we can group you into an archetype of a person pretty quickly because it turns out, and I was telling Mike this the other day, the most fascinating thing about running this company so far and really working on the data personally, is that people like to think they’re very unique people with very unique consumption experience or tastes, and it turns out there are probably literally thousands of Ben’s out there and I speak in general.
The app quickly associates you with other people (i.e. the archetype) as a way of generating initial suggestions. And if you read just enough, it will be quite good at putting what you want front and center in the app from that point forward. Ben Thompson wrapped the detailed commentary into something succint:
It feels like a sculpture approach. Instead of forming a persona, you quickly sort them into a persona and then chip away the parts that actually aren’t right and that’s sort of a shortcut to get there more quickly.
In summary, Artifact does two key mechanical things:
In the early days, Artifacts creates a rough understanding of what you like
It then watches what you do, feeds that data into its models, and uses the output to refine its understanding and learn about you, improving future suggestions.
I’m excited to see how it works for me. Here’s the simple view of what Artifact knows about me:
The top three categories make sense. I was already reading quite a bit about electric cars, cars broadly, and apps and software, before Artifact. Now, I use Artifact to suggest more of what I want rather than finding it on my own.
I titled this post with the word confession because I wanted to admit my misunderstanding. Artifact may add some traditional-looking social features, but for now, it only asks me to invite my friends to the app if I provide permission to my contacts. Otherwise, I can share articles with Artifact links and the app will tell me when my friends click those links. My original view included some fear that the app would try to capture all of my interactions with my friends, but so far, that’s not the goal.
One more interesting conclusion from my revised understanding: you might want to interact more with your friends if you know that they’ve seen (or at least clicked) the link you sent them.
Artifact just might be the ultimate text-oriented growth hack for any network that exists outside of Artifact. Fascinating.