We really want human curation because that's better than algorithms. We still have to do matchmaking, right? And an algorithm is really good for that. So all we haven't replaced the algorithm; what we've done is add really significant qualitative signals into the algorithm that wouldn't otherwise exist.
Good morning, good afternoon, or good evening depending on where you're listening. Welcome back to AI and the Future of Work, episode 315. I'm your host Dan Turchin, CEO of PeopleReign, the AI platform for IT and HR Employee Service. Our community is growing, thanks to all of you, our great listeners. As you probably know, we recently launched a newsletter. Go check it out. We'll share a link to The Beehive newsletter in the show notes. If you like what we do, please tell a friend and give us a like and a rating on Apple Podcasts, Spotify, or wherever you listen. If you leave a comment, I may share it in an upcoming episode, like this one from my new buddy Deo in Seattle, Washington, who is in marketing at a little company called Amazon. Deo listens while making dinner. His favorite episode is that great one with Allison Baum Gates, legendary venture capitalist and author from Seer Verin, about where she's investing in AI and the secrets to breaking into a career in venture capital. We will link to that great episode in the show notes. We learn from AI thought leaders weekly on the show. Of course, the added bonus: you get one AI fun fact each week. Today's fun fact: Christine Patsi writes in the Harvard Gazette that ethical concerns mount as AI takes a bigger decision-making role in more industries. She writes that many worry whether the coming age of AI will bring new, faster, frictionless ways to discriminate and divide us at scale. Part of the appeal of algorithmic decision-making is that it seems to offer an objective way of overcoming human subjectivity, bias, and prejudice. But we're discovering that many of the algorithms that decide who should get parole, for example, or who should be presented with employment opportunities or housing, replicate and embed the biases that already exist in our society. At present, three major areas of ethical concern for society exist: privacy and surveillance, bias and discrimination, and perhaps the deepest, most difficult philosophical question is the role of human judgment given the power and expected ubiquity of AI. Some argue that its use should be tightly regulated, but there's little consensus about how we regulate it. As frequent listeners know, it's a topic we unpack frequently on this show. I agree with many points made in this article. We'll continue to unpack the topic of ethical AI and responsible use of AI in upcoming episodes. In fact, maybe we'll delve into that in today's conversation. And of course, we'll link to that article in today's show notes. Now shifting to today's conversation: Tony Stubblebine is the CEO of wildly popular and very important publishing platform Medium, which recently surpassed the million subscriber mark and was launched in 2012 by Ev Williams, who also co-founded Twitter. Tony's a serial entrepreneur who also publishes Better Humans on Medium and previously founded habit coaching companies Coach.me and Lift Worldwide. Tony's a vocal advocate for writers everywhere, and he speaks frequently about human creativity and why the bots won't take over. Tony received his BA in Computer Science from Grinnell College. Go Pioneers! And without further ado, Coach Tony, it's my pleasure to welcome you to AI and the Future of Work. Let's get started by having you share a bit more about that illustrious background.
Thank you, I'm glad to be here, Dan. And it's certainly Grinnell College will be happy to get that shout out. Very small liberal arts college. When I took over Medium about two years ago, a little bit more than two years ago, my whole career got condensed just into the phrase 'technology veteran.' And when I saw it, I was like, 'Ah, that's actually pretty dead on.' I've been almost every role in a tech company, but mostly smaller companies. I think the largest I ever worked for is maybe 150 employees, and Medium is only around 75. I've been a programmer, I've been a CEO, an entrepreneur, a manager, a writer, a publisher, an editor. And as an entrepreneur, I've done literally every role in the company. Like, I know what QuickBooks sounds like when you're doing bookkeeping. There's a beep that becomes almost addictive; you want to like each entry, get that pace of beeps going. So I at this point actually really like that as the example. But it's kind of funny for you to bring up Grinnell College because I think that the thing that made it work for me is I knew I wanted to go into computer science, but I knew I wanted a broader education beyond that. And so I got that there, where the humanities is really what taught me how to write. And then fast forward 25 years, here I am running a publishing platform, and I don't think that would have been possible if I had gone a more focused route.
Medium is a household name for so many of us. I guess there are around a million of us who are readers and publishers on the platform. But let's say there are a few out there listening who aren't familiar. Describe what Medium is and also the vision, the mission of Medium.
The voice that I think is most important on the internet, and that I think kind of got lost in the modern hustle economy, is the amateur writer who has professional or deep personal experience with something. This is what got me excited about the internet in the 2005 era, where there's this trend of user-generated content. And at the time, I was working for a professional media company, O'Reilly Media, which is writing programming books. A lot of the internet at that time was built by people who are reading these O'Reilly books. So I knew what professional information looked like, and we saw all this amateur information popping up that was covering all the surface area that we would never cover and doing a better job than we ever would. So at that moment, I had a real respect for amateur writing, because it's not just the quality of the words that matter, it's the quality of the experience underneath it. Fast forward to where we are right now, where the incentives of the internet really require, in order to be heard, you have to be good at optimizing for all the algorithms around us. You have to write for Google search algorithm, you have to write for virality. And that was already the status quo before we got to AI-generated content. And now there's a bunch of AI-written content flooding the internet on top of all of that. I think that combination of things made it really hard to find the people who are too busy, you know, by definition the people I want to hear from are too busy living to bother learning all of those hustles. So when I took over Medium two years ago, the major and primary change we made was to redo our recommendation systems to put human curators back in the loop so that they could spot people based on the qualities of what they were saying rather than the popularity signals that most recommendation algorithms tend to be based on.
There's the cliche: if you're not paying, you're the product. And that's true, as we all know, in social media. The kind of—I'm going to coin this term—the 'short-formication' of content on social media is to me kind of the anti-Medium. It's where you go to find a fraction of the truth or a fraction of the full thought. And Medium is a place where you go to hear fully formed thoughts and have more kind of thoughtful dialogue. This wasn't intended to be like an ad for Medium, but maybe talk about how you think about social media platforms versus what you're building at Medium.
I think you hit it there that there's an incentive problem, and it actually came down to business models. I'm not—and not to—I think almost on a social media platform, any tweak you make to the platform ends up changing the whole dynamics and system of it. And the major tweak we made was to think: what would this look like if we had an incentive system where you were the customer rather than the product to sell? So in an ad-driven business model, you are the product; your attention is being sold to the advertisers. And we knew as writers that changes the incentives around the quality of the writing. It's almost counterintuitively: it's bad if the writing is too engaging because then you won't notice the ads. You want it to be just engaging enough. And that's why you get so much content that is so emotional, because it gets you to click through. We changed, built the whole thing around a paid subscription instead. So we have a million paying subscribers out of somewhere on the order of 100 million people who visit in a month. And that allowed us to focus on this different goal: how can I put a piece of writing in front of you that you will be happy to have paid to read? A very different bar to shoot for. And essentially everything else we did falls from that one decision.
The human curation aspect of Medium—anyone who's experienced Medium knows it's such a core part of the experience. We all know that the algorithmically generated feeds in the more traditional social media platforms suck because they're designed to manipulate. It's a steady drip, an IV narcotic that gets you to just stay focused and scroll to the next thing. So that sucks. But I got to ask the corollary, which is: man, Tony, you and your team take on a lot of responsibility when you're the arbiter of what I am interested in. You're making decisions about what I should and shouldn't see. How do you think about how do you balance that responsibility?
You know, it's funny. This is almost like an idea that only exists on social media platforms, that there should be no arbiter, even though in practice then the algorithm is the arbiter and the algorithm is terrible at it. You come from media; your starting position is the opposite. It's like I'm being paid to be the arbiter. And so the way I was able to balance this is when I was coming on board, literally my job proposal started while I was on vacation in the Caribbean, where the point of the vacation was just to read by the beach. And the book I was reading was Neal Stephenson's book 'Fall,' which I don't think that many people read, but it had this concept of 'edit streams' where, depending on how much money you had, you would spend a certain amount of money getting someone to edit the internet for you. And I really like that idea. And what that led to for Medium is like, wait a second, we can provide a sane, thoughtful, very informed default curation experience, and then we can give people ways to edit it by opting in and out. So that's my view of the design: we already come from the background of being very comfortable as editors and curators, but we also come from the internet where freedom is so important. And you can have both, but let's at least start somewhere, which I think we've done a good job of.
When I read the Washington Post or New York Times, I know there's a bunch of stuff coming from the Associated Press and it's labeled, and I know that there's some writers they hire. So I choose to subscribe to that publication because I kind of know what I'm going to get. The vast amount of content that the Medium editors have to sort through arguably increases the importance of that editorial role. I don't have to go too far afield to help the audience with examples. If I'm interested in global warming, or if I'm interested in job automation from AI, or if I'm interested in the hurricane, somebody's making decisions that subtly impact my perception of those topics. There's no way around some amount of human subjectivity seeping in. How do you coach the curators to navigate that super tricky process?
I think when people bring that concern up, I always want to get them into more specifics of what is actually being decided. Between two views, the number one kind of decision point between two sets of views is almost always one of the views is much more thoughtfully done and well researched. So you got two views on a normal social media platform that are competing over how much attention they generate. But then on Medium, you have two views where one of them is your gut reaction based on something your kindergarten teacher told you, and the other is backed by a huge amount of science and research. So you're going to get the other view. And it's not us subtly changing something on a whim; it's like no, we just really value thoughtful deep takes. And my hope is if more of the internet worked that way, that this opposing view would have to compete on the same terms. So I think in practice, why wouldn't you choose this one view that has so much more depth behind it? And then the other one, which is kind of hidden to people, is generally the start of our curation process starts with the subject matter expert. I mean someone who is an expert on a lot of different ways you could say that, but the main way is that they've seen most of what's been written, at least on Medium, about this topic. So they have some sense of 'this is the best version of this' or 'maybe you're not aware, you might be looking for A, but actually you should be seeing B.' That's something that being well-read can really bring into the mix. So this isn't like someone coming into your kitchen who likes cilantro more than you. These are actually really specific forms of taste that I think if you got into it, you wouldn't worry at all. You'd understand immediately what a value it is.
As an example, if I subscribe to Better Humans, your publication, presumably there are adjacent publications that people who subscribe to Better Humans also subscribe to. So I kind of seeing like my curated feed might be everything from Better Humans as well as some subset of things that I might like because I like Better Humans.
It can be that way, yes. That's an interesting part of the idea of curation paired with algorithms. So when I first got to Medium, I think I explained this in a way that was kind of confrontational. I'd say we really want human curation because that's better than algorithms. And I've come to understand the necessity of the subtlety of it, which is we still have to do matchmaking, right? And an algorithm is really good for that. So all we haven't replaced the algorithm; what we've done is add really significant qualitative signals into the algorithm that wouldn't otherwise exist. So if you're interested in one topic and another topic is related and is getting strong qualitative signals, and everyone who reads the topic you follow tends to like that, that's a good opportunity for us to suggest it to you. I mean, that's actually where the algorithm is shining, because we're just asking it to play matchmaker. We're not asking it to play quality control.
That's why I ask the question. Age-old debate in algorithmically managed feeds: if I say I like something, what am I saying? Do I like the author? Do I like the article? Do I like the perspective? Do I like all of it? Maybe I like parts of it. There's sparse information that we're collecting. So the algorithm, in the case of an advertising-supported platform, it's very clear there's only one objective: what it means for me to like it is feeding the algorithm to show me more stuff that's going to keep me on the feed longer. But with Medium, you have a different motivation. So I was wondering what a like means in Medium when you're trying to recommend other content I might like.
We're much less personality focused. I think a lot of other media is you're following a person because you're building a parasocial relationship with them. And on the other side, as a creator, they need that because they need the consistency of your follow. But a lot of the voices we care about are rare, because as I said, I like someone that's almost too busy living to get this information out, because that's where all the really deep wisdom is. So an example of how this might play out is you follow a paleontology publication. The example I've been using recently is 'Fossils at All,' which is run by a paleontologist, but they're publishing other paleontologists who might just publish twice in their entire life. And that kind of connection—you like paleontology, and there's a subject matter expert curating the world of paleontologists who want to write—so we can reasonably assume that we might be doing you a service to bring the stranger, this paleontologist who's a stranger to you, and put them in front of you in your feed. I think that's one half of the standard way we're just doing an interest-based graph like every other recommendation system. And one half of it is that we've put a curation signal in, and this paleontologist who runs the Fossils publication is the curator who's helping to say, 'If you like me, you're also going to like this person.'
Okay, so a related topic that's really important for us to discuss. Let's say that paleontologist, famous paleontologist, chooses to summarize the latest paleontological work in the field using LLMs because it's a great way to curate a bunch of recent information. And the policy on Medium is non-humans cannot publish. But my question is just to get a little more nuanced: how human do I have to be? If a portion of what I published was written by me and a portion was not, does that violate terms? And if it's just binary—all non-human or all human—does that scale? I mean, where do we go from here?
I think a lot of times policy ends up being so practical that people miss the point. The point of Medium is to deepen people's understanding of the world. And so then the practicalities of that are that we want to draw a person's human wisdom out in a way where that wisdom can be passed on to someone else. I use 'wisdom' more than 'knowledge' because it's not just the facts; it's how and why and when are those facts important. That sort of life experience that packages the knowledge is the point. So in this example that you're giving, you could plausibly say, 'Oh, we're still getting a lot of this person's wisdom. I mean, they chose the research, right? That alone is huge.' So that's the key thing: is this worth reading? That's the question. And a lot of times it can be boiled down to: 99% of the time, if it was AI generated, it's not worth reading. But yeah, if you can find a 1% edge case, then that's the 1% you should read. So to have this as a policy, it runs into a lot of practicalities. We can't prevent it from being posted because it's impossible to spot, and there's such a wide range of usage—how much AI assist is in a piece. But we can put a lot of human curation between what gets posted and what gets read. And that's really what's happening. The human curation is like a jury in America: theoretically can overrule the law, but almost never does. So in practice, AI is just not allowed for distribution on Medium because in practice, the way that people use it generates content that's not worth reading. But if that were to change, I think we could allow quite a bit more flexibility. The policy today is that the writer self-reports, self-polices. Within the editors and curators, they're really doing the most to prevent it from spreading to the rest of the community. Because what's important to us is not what exists on Medium, but what gets read on Medium. That's the thing that we think we can control. There's actually a lot of AI-generated content on Medium, but it's just not part of what gets recommended and sent to users because there's a human filter.
