Twitter, Clubhouse and engineered serendipity
My response to the question ‘why do you use twitter?’ is always the same — you can find things there that you would never have found any other way. Naval and Nick Cammarata have both substantially influenced my perspectives on the world, and the reason Twitter, despite all its flaws, is a such an important product to me is I would never have found these people without it.
Clubhouse has something of that same magic: stumbling across something you would never have come across otherwise, but yet something you really value — something that quickly becomes part of the furniture of your mind. This is wonderful, powerful stuff; the internet at its very best.
But it comes at a cost, and that cost is to your time and attention. Having had one of these moment of ‘magic serendipity’ stumbling into a conversation on creativity in clubhouse, I found myself wondering how many minutes I had wasted listening to second-rate conversations to get here. What is the hit rate? The ‘magic moment per minute rate’? I have found myself opening Clubhouse less and less often recently, and I think it is because this hit rate is simply too low. The quality of the median moment of listening is not good enough, compared to more considered audio content: radio, podcasts.
The basic ‘material’ of Clubhouse is unstructured and transient: its audio — unstructured by default — and it disappears as soon as it is spoken. This poses particular challenges for Clubhouse in driving discovery for users: how do they know what rooms to suggest to people? How could they possibly try and engineer for those ‘magic moments’? Twitter has a somewhat related problem, although tweets aren’t transient (though they do decay fairly quickly).
What you need is signal — some way that Clubhouse knows that the user is experiencing that moment of ‘wow this is such a great conversation!’. But clubhouse has no mechanism for this (only the opposite: which is leaving a room). What you need is signal on what users like, then a way to process new content for these same features so you can recommend the next thing. This is how Tik Tok works: skip/stay on a video is the signal, combined with algorithmic understanding of the video content (e.g. cat videos). This creates an incredibly tight feedback loop to learn what people like, quickly, and is the driving force ultimately for propelling it to being the fastest growing social app ever:
Clubhouse’s ‘leave room’ bears similarities, but is a much more sparse signal. More problematic though is that there is no real way to tie this back to the next bit of ‘content’ to recommend. Tik Tok has a bank of videos ready to queue up, twitter has tweets, but with Clubhouse the content is all live — there is no queue. I fully expect Clubhouse to introduce some signal-generating features into the app — like the reactions you get on an instagram live — to start getting more signal on what users are enjoying. But the problem of the content being completely live is more challenging. I rather suspect that the reason the app exploded so quickly in Silicon Valley was that the discovery and curation side — the magic moment per minute — happened by default due to the closed-community nature of how the app emerged. In other words — they did not need to worry about algorithmic discovery because the network did it for them. Well now its a diverse, global platform with sky-high ambitions, and it will have to start figuring out how to get users to content they are going to really enjoy.
Twitter and Clubhouse really are marketplaces for ideas, and the effect of that is that the core ‘unit’ is deeply, profoundly unstructured. Spotify and YouTube deal in unstructured formats (Audio and Video) but they can be somewhat parsed out — think of how this content gets classified and presented to you in the home screen: funny videos, Deep House. With pure ideas these underlying categories just don’t really exist. Ideas are free-form and atomic, they can’t be folded into some schema, however advanced the AI (well, not yet…). Can you ever really say to algorithm — ‘this tweet, more like this please’ and expect it to be that effective? The essence of why you found it interesting is too subtle and too idiosyncratic to be reducible to a common schema. It in some way reflects your ‘world view’, which is unique, and the interaction between this and grasping the gist of the ‘idea’ is beyond tractability for algorithms, probably by a long way. Instagram incidentally is similar — yes the format itself is images, which can be parsed with very high accuracy these days, almost trivially, but why did you like that particular shot? What was it about it that made you hit like? This complex interaction is not digestible by an algorithm — most of the real information would be lost by trying.
This is why discovery in these ‘world view’ or ‘idea’ marketplaces is driven not by recommendations on the base unit — the song, the tweet — but indirectly, via human curation. On Twitter you follow people, as on Instagram and Clubhouse. You rely on these people to curate, surface and create content for you. Ultimately these people husband a world-view into being, through the native content of the platform, for all to participate in — to try on, to give themselves to. There term ‘influencer’ is somewhat maligned in popular culture today, implying a kind of grubby, vain, status-seeking, but actually it is hard to overstate their influence collectively: they are the routing nodes for ideas in the 21st century. That is a big deal, its a big deal in the way the printing press was a big deal — by affecting information flows through society.
The takeaway for these apps is that the trajectory of their success is driven in large part by helping users find the right people to follow. Some users are prepared to do the heavy-lifting themselves, actively curating the set of people they follow, but at around 100m users Twitter began to hit a ceiling where the newer users just weren’t motivated enough to do this. They approached it more passivley, wanting more of the heavy lift to be done on Twitter’s side. And Twitter has always felt like its a little half-hearted about this, yes there’s a ‘people to follow’ module but it feels like an aside to the main product.
This is why Twitter’s new ‘super-follow’ feature holds so much potential: it is very, very powerful signal. How does Twitter know that Nick Cammarata is one of my very favourite people on Twitter? Only indirectly, by observing how often I like and re-tweet his stuff. But this is polluted signal: how does twitter delineate between my interest in the person and the ideas (tweets) themselves? What if I don’t like or retweet that much? What if I like and re-tweet all the time? With a charge of $5 per month, a super-follow is an unequivocal, unambiguous signal of real love for that person/account. It really is sharp, crisp, irrefutable. And now with this Twitter can get to work: an affinity matching algorithm would look at who else also super-likes this person, then looks at who else they are following that I’m not following. Do this writ large, across all of the followers of that account, and you can get some pretty powerful recommendation machinery.
The lesson in the end is to take seriously the power of accounts if driving the value through the ecosystem, from content to user. Monetising makes complete sense seen through this lens (Facebook is also experimenting with helping creators generate revenue) — it is following the value-creation of curation in the ecosystem. This ‘value-finding’ has perhaps been forced by the explosion of competition in Social over the past couple of years, and the necessity of these older platforms who thought they had it all sewn up with the network effects, needing to up their game. A lot of focus is on the ‘creator’ economy these days, which conjures images of Tik Tok stars churning out vids in a shared house in LA; less has been said about the role that individuals play as distribution nodes for information in a world that has never been more saturated in the stuff. Was it the printing press itself where power lay, or the publishers?
