At Pinterest, AI implementations start small and add complexity over time
Silicon Valley is known for its cutting-edge technology, but true scale and application will come from simplicity. "There is no technology that's really useful if it cannot be explained with plain words," said Vanja Josifovski, CTO and VP of engineering at Pinterest, speaking Thursday at The AI Summit in San Francisco.
As AI has advanced, algorithms have steadily become more complex. In turn, rather than maintaining that complexity, the challenge becomes lowering the bar for use. Much can be achieved without a technical background, according to Josifovski. People just need to understand the relationship between the installed features and the outcomes.
At a company like Pinterest, one of its key operating principles is making do with a lack of resources. The platform has 250 million monthly users with 175 billion Pins on the platform, but only so many engineers, Josifovski said. So when engineers inject new functionalities into the platform, they have to understand the impacts on the business and what will change for the user.
While AI has evolved over time, people are largely using it to work with data. Previous AI hype cycles have come and gone, but hardware improvements have revitalized its popularity.
"What we call AI today is more a result of an evolution of us being able to collect and store data and the evolution of the hardware," Josifovski said.
With growing calls for AI democratization, ease of use takes center stage. Pinterest conducts internal classes and employees from project managers to people outside product development participates. Using classes, the company can work to grow an in demand workforce and create more opportunities to find new AI applications on the platform.
Pinterest has a unique data set to interact with, with picture and text components. Technology is available to purchase and implement, so the company focuses instead on building everything around the core parts of the stack.
The most recent advances in AI are in increasing the intelligence of the visual components. Data at Pinterest takes on three shapes: The image; the info from the shared web page, which can contain user comments; and the graph data, which represents a specific interest for the user.
With graph data, a user could Pin a kitchen with a blue fireplace, Josifovski said. But for another user, that same image might appear on a board for a modern kitchen.
The goal of AI is "making use of your data," Josifovski said. While Pinterest has many AI applications, the company has to ensure it is collecting and storing the right data while channeling its analysis for platform improvement or back office operations.
It starts small. To grow its AI platform, Pinterest had to make sure the simple projects worked. From there, it was all about layers.
Follow Naomi Eide on Twitter