Dmytro Shandyba is the CTO of Newzmate, a new Ukrainian artificial intelligence (AI) company using machine learning tools to drive traffic to web news sites. Shandyba brought his startup to the annual innovation conference TechCrunch Disrupt San Francisco in the hope of find funding and potential business partnerships.
While tech giants like Google and Amazon are setting the pace in AI R&D these days, startups like Newzmates can also find success by targeting niche markets
This year, TechCrunch invited leading figures in AI to advise attendees how to build a successful startup. Google’s Senior Vice President John Giannandrea, Udacity’s Founder and CEO Sebastian Thrun, and Founder of Deeplearning.ai Andrew Ng all delivered optimistic speeches concerning the prospects for AI startups.
Giannandrea, who leads research on Google’s search engine and AI, says today’s startups are taking advantage of growing public access to key resources such as ImageNet (as an open-sourced dataset), Tensorflow (an open-source software library for Machine Intelligence), and research papers. Over the first five months of 2017, US$9.5 billion in venture capital was invested in AI startups, almost three times the 2014 total.
Giannandrea says Google has played a key role in encouraging both academic and industrial AI development. “We have published image and video datasets, literally giving away millions of dollars worth of labeled data.”
Udacity’s Thrun pointed out how his company helps entrepreneurs and engineers access a low-cost machine learning education; and announced new Udacity Self-Driving Car and Flying Car nanodegree programs. These will enable anyone with minimal programming experience to acquire basic understanding of these emerging vehicular technologies.
Another instructional AI initiative is Andrew Ng’s startup Deeplearning.ai, which earlier this year launched five deep learning courses on Coursera — a top online educational platform which Ng co-founded in 2012. The curriculum will cover state-of-the-art machine learning topics such as CNNs, RNNs, LSTM, Adam, Dropout, BatchNorm, and Xavier/He initialization.
Before Ng founded Deeplearning.ai, he had established centralized AI teams for both Google and Baidu. He sees similarities between Internet startups in the 2000s and today’s AI startups: “The rise of AI technologies offers so many opportunities that there is no way for any one large company to exploit them all.”
Ng believes that small startups should generally avoid challenging tech giants in verticals such as web search engines, which rely on unique data assets that are difficult for startups to acquire. However, there are other verticals — such as healthcare or agriculture — where tech giants do not necessarily have an unassailable advantage. ”As long as you have an application that is good enough to launch, it will get some users, and that allows you to create more data. This creates a virtuous cycle,” says Ng.
Meanwhile, Mohsen Hejrati, Founder and CEO of San Jose-based startup Good AI Lab told Synced he believes that in spite of open-sourced datasets and tools, AI is still largely a club for big, well-staffed and wealthy companies.
“AI is exclusive because algorithms and science are still limited, knowledge and expertise to use it is rare, and infrastructure is hard to build and costly to operate,” Hejrati says.
Hejrati says a typical research team will try one hundred ideas before one moves forward. Then the team will run 50 experiments for one week each on 50 GPUs, which is equal to 42 million hours of computes. This will cost the team around US$10 million on a public cloud platform such as Amazon Web Services.
As such, Hejrati is determined to drive down computing costs. At TechCrunch Disrupt he announced Cluster One, a system that can aggregate computing resources from computers anywhere in the world. By downloading an app, users can donate their surplus computing power. Hejrati says he’s aiming at amassing 100 billion AI cores by 2024, which would make Cluster One the largest AI supercomputer in the world.
“Everybody can take part, from cloud giants to small cloud providers, from corporate computer fleets to your own desktop,” says Hejrati.
When asked whether this really is the best time yet for AI startups, both Hejrati and Newzmate’s Shandyba responded positively. Says Shandyba, “If you find a specific unsolved problem in a defined market and you have technology to get it resolved through a reasonable business model: go for it. That’s what I would call an ideal prerequisite for engendering successful tech startups.”
Journalist: Tony Peng | Editor: Michael Sarazen