Do Things That Don’t Scale… Then Scale Them!
How we scaled the unscalable to create a Trust & Safety leader
Ralph Waldo Emerson would probably not have been a very good entrepreneur. He writes:
“If a man has good corn or wood, or boards, or pigs, to sell, or can make better chairs or knives, crucibles or church organs, than anybody else, you will find a broad hard-beaten road to his house, though it be in the woods.”
Now, Emerson lived 150 years before the iPhone. But many believe the idea still resonates today. If you build a better mousetrap, people will find it naturally, and your startup will take off.
Simple, right?
In reality, building a better mousetrap is usually not enough. It’s not enough to be very good. To get something from 0 to 1 typically requires something remarkable. It takes solving a problem others haven’t solved.
Startups take off because the founders make them take off by doing well what was never done before. Usually it takes some sort of push to get them going.
Paul Graham, the founder of Y Combinator, famously wrote about this concept in an essay called “Do Things that Don’t Scale.” It captures the hustle mentality that founders need to build a successful business.
What does it mean to do things that don’t scale?
A startup may have to recruit its first users manually. In the case of Airbnb, the founders had to get their friends to list their homes or spare bedrooms on the website. Facebook had to get its first users on the Harvard campus before they launched at other universities. They had to get a critical mass and validate the business model of social media before they opened up to the world.
This scrappy approach to building a business can be counterintuitive for tech entrepreneurs. It’s not always neat or elegant. It’s not like writing software, which can scale infinitely.
But the lesson is that when you’re trying to get a new product or feature off the ground, it requires a different approach. You have to roll up your sleeves, roll that boulder up the hill, and reach critical mass, before the flywheel effect kicks in.
The Rise of Livestreaming
I intuitively followed this philosophy throughout my career, and most recently at The Meet Group to roll out our successful livestreaming product.
In 2015, I was on a trip to Asia, and I saw the success China’s dating app companies, like MOMO, had integrating livestreaming into their apps. I thought this could be recreated in Western countries. We know that people often come to dating apps to meet new people, but the downtime between chatting one-on-one with other users can leave them feeling even more alone. Livestreaming video helps to combat that aloneness by giving people a way to be entertained, often in the downtime between chats.
We wanted to build a livestreaming platform where creators could thrive — one where users could showcase more of their personality, and potentially generate substantial revenue for themselves and the company, all while entertaining a potential user base of millions.
The opportunity for livestreaming was clear. But there were significant challenges to delivering a live video solution that was safe, moderated and acceptable to the app stores. We drew up a list of the things we’d need for a safety process: moderation resources and manpower, technological and machine learning capabilities, safety and moderation standards expertise, streamer management tools and processes, and relationship building and maintenance resources to work with law enforcement and government safety agencies. We knew that every live stream had to be moderated for the stream’s entire duration. And even if we could successfully build the safest solution, the underlying processes, teams and capabilities all needed to be able to scale.
With millions of active users, we planned for at least one million minutes of live streaming per day. We knew that failing at content moderation came with severe consequences. App store providers have careful content standards and real enforcement that has seen the removal of even large apps for inappropriate content. Advertisers don’t want their brands to be anywhere near bad content. Payment processors won’t facilitate payments associated with adult content. At that time, we were a NASDAQ-listed company. Our shareholders were clear that they did not want to be invested in the distribution of adult content.
In short, we saw a huge opportunity in livestreaming, but we couldn’t mess it up. For it to have a chance of working, our livestreaming product would need to have careful, proactive moderation in place from day zero.
At the time we began developing our livestreaming capability, The Meet Group already had an experienced Trust & Safety team, and a growing revenue stream of many tens of millions in annual revenues from our existing apps and from advertising. This financial profile made investing in safety possible.
We first launched live video on our MeetMe app. It was successful right from the start. It was the fastest-growing revenue product in the company’s history.
Now, here’s the part where we had to do things that don’t scale.
We had to figure out a way to proactively moderate all this live video content. We began by programmatically taking screenshots of every video stream, multiple times per minute, and we reviewed those screenshots for appropriateness. At the beginning, we had some legacy AI moderation systems that were somewhat helpful, but human moderators at our locations in India did the bulk of the review. If a screenshot of a stream was flagged as inappropriate, the associated live stream would be ended, and the larger safety team would be alerted to review the matter.
For the first couple of months, a team of hundreds of human moderators reviewed millions of images per day, at a substantial financial cost to the company.
Using human moderators was the first step. Relying on humans for one hundred percent of your content moderation doesn’t scale — but it was one hundred percent necessary for us to get our livestreaming product off the ground. The next step was to take this human-centric approach to content moderation and find a way to scale it to our millions of users.
We worked with our in-house Machine Learning engineers to create proprietary algorithms for content moderation. Our engineers told us that if we could provide them with hundreds of thousands of images that were carefully labeled by humans as to their exact content, they could use those images to “train neural nets”, and ultimately give us an AI solution that can interpret the content of an image — in milliseconds, for fractions of a penny.
At the time, there were off-the-shelf solutions available, and we tested those for detecting bad content and nudity. But we found that our homegrown solution was far better, having lower false positive and false negative rates, because it was entirely trained on our actual content.
Scaling the Unscalable
After 6 months of training and retraining and then continuously retraining, we had achieved a high level of precision. Pretty soon over 95% of streamer moderation samples were being handled by AI. This enabled us to exponentially increase the frequency at which we sample streamers. That was a good thing too because the amount of content would grow dramatically. We know mange the livestreaming video for 16 apps, including large properties like Tumblr. We’re now capable of generating and analyzing as many as 100 million samples per day in a cost-effective manner.
The vast majority of live streams are abuse-free and can be handled precisely by the algorithms trained to uphold our specific content standards. The small percentage of live streams that aren’t solved by AI require careful analysis by our Trust & Safety team.
We started using Google voice-to-text transcription technology to discover abuse that may be taking place in the audio portion of a live stream. We used age estimation algorithms for detecting minors who manage to sneak onto the app.
These tools were critical to building a livestreaming capability that transformed the company. We were able to scale the unscalable by creating a leading safety organization in social apps.
Today, our careful, scaled moderation contributes to attracting advertisers, building strong relationships with payment processors, and providing a safe space for a vibrant live streaming creator economy.
We did many things to underscore our commitment to safety, including cross-industry collaborations with the Internet Dating Excellence Association and Spectrum Labs, a company equally committed to preventing online toxicity with content moderation. Additionally, we joined forces with Oasis Consortium, an organization of thought leaders across gaming, social media, and dating created to accelerate the development of an ethical Internet, who has pledged a commitment to online safety through the User Safety Standards pledge. We were among the first crop of dating companies to receive the Social Safety badge from the Internet Dating Excellence Association (IDEA).
All these efforts enabled our success today. The Meet Group currently enables & moderates livestreaming video for seven of the top grossing 25 social networking apps in the US.
Lessons for other Founders
As I’ve reflected on what worked for us, and what made our livestreaming product so successful, I’ve been able to distill 3 big lessons from our journey.
1. Think small before you think big.
Founders often want their companies to seem big, and they end up copying the flaws of big companies, like being indifferent to individual users. This seems to them more “professional.” But it’s better to embrace the fact that you’re small and use whatever advantages that brings.
One advantage of being small: You can provide a level of service no big company can. Delivering a great customer experience becomes harder as you scale to millions of users.
Running a startup requires a creative approach to problem solving. A startup can be fragile. You need to solve problems every day to survive. Whether it’s acquiring users, or keeping them happy and engaged. You should always be racking your brains to think of new ways to delight them.
Sometimes it’s better to launch small. Run experiments on a small subset of users, see what works, before investing in an automated solution and rolling it out to everyone.
2. Do things that don’t scale.
Doing things manually may be frightening to some tech founders, but it’s less frightening than eroding your customer experience and losing users.
This can mean coming up with creative ways to find your early users. It could mean recruiting people at conferences, trade shows, campus events. Participating in niche online communities, like subreddits with people who share a set of common interests.
For a hardware startup, it could mean assembling your first products by hand, before you can invest in automated manufacturing.
In an ideal world, you’d create a feature, roll it out to everyone and it would scale perfectly. In reality, when a problem comes up, you need to solve it by any means necessary. You can’t let the perfect be the enemy of the good. You have to do things that don’t scale.
3. Scale them!
In any business, there may be things you think can’t be scaled, but time has a way of solving your problem for you. For The Meet Group, it may have seemed crazy to think sampling a video stream every 3 seconds and reviewing millions of screenshots a day could be scalable, but it is.
The arrow of technological progress points toward automation and scale. Compute costs fall, algorithms improve. Global work forces allow cost efficiencies. An ecosystem of third-party tech pops up to support your business.
You may think you can’t scale something like a safety program. But as long as the core part of your business works — as long as users adopt it and spend money — then a cottage industry of tools will emerge and you can rely on them for almost-free R&D. You just need to survive the early years when your industry is still developing.
You need patience, discipline and imagination. Take the hand grenade pitched at your head and make it just one of a number of balls in the air. That’s how you go from having a great idea to delivering something that impacts the lives of millions.
