Source Code: Your daily look at what matters in tech.

enterpriseprotocol | enterpriseauthorJoe WilliamsNoneAre you keeping up with the latest cloud developments? Get Tom Krazit and Joe Williams' newsletter every Monday and Thursday.d3d5b92349
×

Get access to Protocol

Will be used in accordance with our Privacy Policy

I’m already a subscriber
Protocol | Enterprise

Microsoft, Facebook, IBM and Google have dominated AI research. Asapp is trying to change that.

Cheaper computing, more sophisticated tools and abundant funding are making it possible for new competitors to emerge.

Smaller AI players are challenging the giants.

Smaller AI players are challenging the giants.

Image: Uriel SC

In the modern era of artificial intelligence, a handful of companies dominate industry-backed research — Facebook, Google, Microsoft and IBM. No surprise there: Those firms have war chests the size of nation-states and have made AI a core business focus.

But now, empowered by more sophisticated tools, ever-cheaper computing power and plenty of funding, startups are trying to challenge that dominance.

One is Asapp, which sells software to digitize the call center. It's a hot upstart in customer experience technology, a market that could see sales of $24 billion a year by 2027. Investors including industry heavyweights like John Chambers and John Doerr have poured over $255 million into Asapp at a valuation of $835 million, per PitchBook. It counts household names like American Airlines and Vodafone as customers.

Since its founding in 2014, Asapp has assembled a research team with some of the top minds in conversational AI. Leading the group is Chief Technology Officer Will Robinson, who spent over a decade at Google. The company recently hired Ryan McDonald, another longtime Google employee who focused on natural language processing research there, as chief scientist. Other key personnel include Tao Ma, the former head of speech at JD.com. And its mission is much more singularly focused than that of the industry giants: It aims to pinpoint research that can improve its product suite.

At places like Facebook and Google, "scaling very, very big and very, very broad is almost assumed with anything that a researcher there is doing," Robinson told Protocol. "We are laser-focused on this customer experience performance platform. And the further we dig into the problem, the more surfaces and the more places we are going to find to apply this advanced artificial intelligence."

While firms like Microsoft and Google produce research at the same pace of top educational institutions like Stanford and MIT, the direct commercial tie-in is not always apparent. Facebook, for example, is building robots that can beat humans at complex games. It aligns with the company's goal of getting AI to "match human intelligence," but the outcome is unlikely to dramatically change its product landscape right now. Similarly, in 2018, IBM unveiled a machine that can debate humans. While teaching an algorithm to make persuasive arguments is impressive, there aren't many businesses clamoring for that technology. IBM is now trying to integrate that technology into its broader suite of AI products, but it's unclear when that will happen.

Those firms' contributions to the field of AI are still impressive. Google's BERT, for example, is used widely to train NLP models, as well as by the company itself to help improve user searches. It's those kinds of achievements that serve to advance the field overall, but don't necessarily mean a big revenue boost for Google. By focusing narrowly, Asapp and others are tacitly admitting that the longer-term, blue-sky research is much harder to accomplish in a small company.

But Asapp is confident in its thinking that there's room to go deeper, not broader. It's a mindset other burgeoning businesses have as well. PolyAI, for example, also regularly publishes research focused on conversational AI, as well as studies that benchmark its own software against models from the tech giants.

"[In] the field of natural language processing there's rapid innovation occurring that is, in part, being driven by startups," said PitchBook senior analyst Brendan Burke. "Startups are learning how to train models in unique ways to solve particular problems … and the tuning of the model once it's in production can exceed what some large companies are able to achieve."

For that reason, Asapp is actually forgoing research outside the core speciality. And that focus is helping bring on talent like McDonald and Ma, according to Robinson.

"There are plenty of people with world-class research chops or machine learning, engineering chops who very much want to see their work immediately having an impact in the world and on a business," he said.

Now, Asapp is working to even more closely integrate the research arm with the business itself to amplify that commercial value. For one, researchers often spend time directly with customers, regularly sitting with service agents to learn how they do their jobs. Understanding the pain points then helps inform where Asapp should focus its research efforts.

And despite its smaller stature, Asapp has scored some big research wins over the past few years. Notably, it built an architecture called SRU++ that Asapp says can compete with Transformer, which is used as the foundation for many pre-trained systems like OpenAI's GPT and BERT. And its technology is rated as some of the best; tested against LibriSpeech, a dataset derived from audiobooks that has 1,000 hours of spoken English, Asapp's had the lowest error rate.

Ultimately the rise of Asapp's research capabilities shows how far the field has advanced. A decade ago, access to the compute power and hardware required to test algorithms was much harder to come by. But it's "started to spread out," said Robinson. "One of the things that you see resulting from that is companies like Asapp propping up and … applying this expertise more like a scalpel."

Latest Stories