Open-source traction vs enterprise profits: how do we find balance for product success?

Founder Vision is a one-on-one podcast that digs into deep conversations with business leaders from emerging markets as they get vulnerable about their experience in the early- to median-stage moments of their founding journey.

Activeloop is an open-source framework that was built to solve a big problem right now in the ML world: dataset management. Machine learning algorithms have advanced a lot, but managing a dataset remains extremely hard. When your data set becomes image data sets, or audio, or video, it becomes much more difficult to use them to train models — to reuse them, or do anything else useful with them. That’s where Activeloop comes in, allowing users to build and scale data pipelines for machine learning and computer vision.

Brett’s conversation is with Shashank Agarwal, Activeloop’s CTO.

Catch the full episode in the player below, or on Spotify, Audible, Apple Podcasts, Google Podcasts… pretty much wherever you like to listen.

Key Takeaways

Embrace Open Source

Open source vs. enterprise is a line many tech companies have walked, and it’s a hard probem to strike that balance when you need revenue and you have investments. Activeloop is going strategically hard in the direction of open source.

“We have a big community of individual researchers,” Shashank says. “Some of them are employees at big companies, such as Google. Some of them are college students. They all want to reuse data sets. They contribute and help make the product better. They help us try out new features. It has been really helpful for the community so far, and to us in building the product. At the moment, our entire focus is on open source.”

This was a decision based on much research. Shashank says that the Activeloop team talked with some other leaders in the industry (who must remain nameless for now), who made “a very famous open-source project” which raised $100 to $200 million dollars before it became enterprise.

“We asked them this question specifically,” Shashank remembers, “How much should be held back from feature releases? We asked because at that point we didn’t have a very compelling enterprise product to sell. He said ‘no, double down on the open-source features and never trade traction with anything else. You cannot buy traction.’”

So far, so good: Activeloop is trending in the top 10 projects around the world on GitHub, and they’re aiming directly at becoming an invaluable resource on the world ML stage.

“I can’t tell you the name of enterprises who are using us already,” he adds, “but we already have some initial deals where [Activeloop’s framework] is already part of some environments where people are using it to manage their datasets, update their datasets continuously and use it to train on a very large cluster, on the terabyte scale, already.”

Think Huge

“If we are successful in our goal and what we want to achieve, in two years, we will allow or enable a lot of these amazing datasets to become available,” Shashank muses. “For example, if somebody wants to train or make an autonomous driving car, they will have a dataset that´s already available already with one click. The code is also there. They can start immediately working on it and tinkering with it and running with it. We’ll soon have over 5,000 or 10,000 datasets for every single application, which are readily available for training — and, in most cases, we’ll have the training code already available and being contributed to by the community.”

“Imagine a community working together with these datasets and trying to solve problems, trying to make a better model in the open-source world,” he smiles. “That is the solution we have in two years.”