AWS Startups Blog

Before pitching yourself as a machine learning startup, you better be one, says Matt Hartman of Betaworks

Matt Hartman of Betaworks speaks at AWS Startup Day in New York.

Today, in practically every venture capital pitch room around the globe, the technological darling of the moment is artificial intelligence—and more specifically, machine learning using deep learning techniques.

It’s for good reason. The potential of deep learning techniques today—and other approaches to machine learning in the future—to remake entire industries is profound. No wonder that every startup seems to make some claim to a machine learning pedigree, as startups did with big data, mobile, and the web before. But, as Matt Hartman, a New York City-based partner with accelerator/investment firm/tech factory Betaworks, told the crowd at the AWS Loft Startup Day in NYC, every startup needs to be clear, especially when talking to potential investors, about what flavor of machine learning company they are.

“It comes down to whether you truly are a machine learning company,” Hartman says. “Or a company that uses machine learning.

The distinction is important, especially when you are pitching folks like Hartman and other investors. It’s whether your differentiator as a company—your competitive advantage—rests primarily on machine learning or something else like your brand or the network effect you’ve achieved through savvy partnerships. Either approach can lead to success, but they look different structurally and your startup better too.

“If you truly are a machine learning company, where are your PhDs?” Hartman says. “What incredibly difficult computer science problem have you solved that no one else has?”

He offers the example of Uru, a New York-based startup that has developed a technique to seamlessly inject new advertising or brand content into an existing video (or any content, really). Yes, Uru has developed partnerships with brands and integrations with video players, but it has set itself apart through its computer vision chops. “Inserting new content into a video is actually a very difficult computer science problem,” Hartman says. Not only have Uru’s PhDs developed a solution, it’s proprietary, and that is the value they bring to the market and potential investors.

Maybe your machine learning-based solution isn’t proprietary, well, what about your data? Can anyone get the same data and train it, or do you have unique access? Does the proprietary data you have trained get more valuable, the more of it you have? That can be another key point of leverage and value from an investor’s perspective.

From Hartman’s vantage point, if you can’t check those boxes, if you don’t have either proprietary solutions to a tough problem or proprietary data (ideally both), you aren’t truly a machine learning company at your core. You are using machine learning.

And that can be just fine. Gif search engine Giphy initially built its business by being there first and then solidified its position with a strong brand and a business development team that locked in key partnerships with external brands. Machine learning might help Giphy tag gifs better and therefore make the product better. “But a competitor won’t win even if it copies the machine learning tagging,” Hartman says. “They key to Giphy’s success is its brand and business development, and therefore machine learning doesn’t need to be a core competency.”

If machine learning isn’t at the center of what you do as a company and if it doesn’t offer a fundamental competitive advantage, that also suggests a different structure. You don’t need those PhDs. You don’t need to spend the time and money solving tough computer science problems yourself, Hartman says. Instead, buy the answers to your image recognition or natural language processing from someone else and focus on what does make you different.

“Think about it this way,” Hartman says. “Machine learning will eventually be where analytics is today. Does every company need to leverage analytics? Probably. But does every company need that to be their very core competency? Maybe not. The same will be true when it comes to machine learning. It will become table stakes for a lot of companies, but it won’t be their core competency.”