Growth = Step-Function Changes + Incremental Improvements

October 2024 marks my fifth year at Snap, making this my second-longest tenure with an organization, aside from my graduate program. While there are still many challenges to tackle at work, I want to take a moment to reflect on the journey that brought me here and to share some of the insights and lessons I’ve gathered along the way—many of which were learned the hard way. In this post, I will share my observations on growth.

Why Snap?

Rewind to 2019: I was looking for a role in a domain or product that was close to end customers, business-critical, and technically challenging enough to keep me busy and dedicated for at least 3-4 years. When the opportunity arose to build the Friend Recommendation system on Snapchat, I knew it checked every box I’d hoped for. So, I joined Snap to help build Friend Recommendation, a product now used by hundreds of millions of users daily.

Quick Add

The Friend Recommendation product on Snapchat is commonly referred to as Quick Add. In the industry, friend recommendation systems are better known as ‘PYMK’ (People You May Know). This term was initially introduced by LinkedIn (I had thought Facebook coined it first) and was later popularized by Facebook as one of its most crucial growth products.

Friend recommendation is an intriguing and complex problem with unique challenges. The first challenge is scalability: we’re recommending friends to billions of users from within the same billions of users, which is essentially an N² problem. Secondly, because we recommend users to other users in the same social graph, strong network effects make A/B testing and evaluation difficult. Finally, like many other networks, there’s a Matthew effect, where ‘the rich get richer,’ which could hinder growth if not managed carefully. In addressing these challenges and many others, I have repeatedly observed the following pattern.

Growth = Step-Function Changes + Incremental Improvements

Looking closely at the success metrics of Friend Recommendation over the years (adjusted for seasonality), a clear pattern emerges in the chart: a few major surges followed by steady, gradual increases. In other words, growth is driven by two forces — numerous small, incremental changes and a few significant, step-function changes. Both are essential and irreplaceable.

Step-Function Changes

First, let’s discuss the significant step-function changes and why they are crucial for growth. Just as the Development of Language, the Agricultural Revolution, and the Industrial Revolution were pivotal in human history, similar events have shaped the evolution of Friend Recommendation products. I can trace each major surge in the success metrics to specific feature launches. For instance, the first online feature store enabled feature engineering at Snap’s scale, embedding-based retrieval introduced a new method for candidate retrieval, and the continuous improvement of user embeddings, along with significant UX and UI updates, reduced friction in the friending experience. These moments are truly exciting, as they laid the foundations and unlocked possibilities once considered out of reach.

While these step-function events can shift a product’s growth trajectory, they are relatively rare. As a product matures, these foundational shifts become less frequent, often only occurring with the emergence of new technology or innovative product ideas. However, it’s more often the case that these shifts don’t happen because the team hasn’t made bold bets—or enough of them—to drive transformational change. Step-function changes don’t necessarily come from moonshot ideas; they often arise from ideas in which we have high conviction, supported by success in similar products in the industry. The key is to make deliberate decisions to invest in these initiatives, beyond just focusing on tactics.

While big changes do occur, they are inherently rare and insufficient to drive growth by themselves alone. Now, let’s turn our attention to the small, incremental changes.

Incremental Improvements

The human brain tends to latch onto a few significant events. In Morgan Housel’s book ‘Same as Ever’, he wrote that many people remember that Dwight Eisenhower survived a heart attack in 1955, while far fewer notice the remarkable 70% reduction in the death rate from heart disease since then thanks to advances in detection and treatment because that is 1.5% improvement each year on average, too small of a number to be noticed by most.

Similarly, in the domain of growth, people tend to celebrate and remember the big wins, while the incremental changes often go unnoticed and are frequently overlooked. Here’s an interesting story I experienced firsthand. When we launched the first version of user embedding using Graph Neural Networks, we celebrated the achievement with an email to executives and a party. Since that initial launch, we’ve kept iterating on the embedding quality and shipped more than ten versions of embeddings subsequently (not counting the unsuccessful attempts). To my surprise, when I later combine the compound metric gains from these later iterations, they far surpass the gains from the initial launch. The compound effect works its magic. My takeaway is that we shouldn’t rely solely on big, step-function changes to drive growth while overlooking the seemingly small, incremental ones; when enough incremental improvements accumulate, their combined impact can be astonishing.

An Analogy from Machine Learning

I see a parallel between this pattern of incremental changes and step-function changes in machine learning—specifically, gradient descent. In gradient descent, the goal is to find the global optimum of a function. Incremental improvements can be viewed as performing local optimizations; they are essential because they help us identify the best solution within the constraints of existing technology. However, if the team exhausts all opportunities offered by the current technology without introducing new innovations, productivity can stall. This underscores the importance of investing in technologies that can help us break free from local optima and facilitate step-function changes. These significant changes elevate the business to new heights and create numerous low-hanging fruits, which can then be realized through further incremental improvements. And then the cycle repeat itself!