Uber isn’t about to be left behind in the race for artificial intelligence.

The ride-hailing service has been staffing up in AI over the past year, thanks in large part to its December 2016 acquisition of Geometric Intelligence, a start-up incubated at New York University. The idea is to help the company do everything smarter — rides, food delivery, self-driving cars — as it faces competition from the likes of Alphabet, Amazon and Lyft.

Rather than letting the start-up continue to function in New York, Uber brought the team to its San Francisco headquarters in March 2017. And while CEO Gary Marcus did leave just a few months after joining, Uber was quick to promote another cofounder, Zoubin Ghahramani, to be the new chief of the burgeoning lab.

The lab has roughly doubled since then, with nearly 30 people out of Uber’s more than 16,000 employees, Ghahramani told CNBC in an interview. And he’s structured the team to avoid some of the historical issues that other big tech companies have faced in operating research organizations.

“Having interacted with many companies over the years as an academic, I’ve seen companies where research was sort of like this side thing — there was always this frustration that it wasn’t interacting with the core business,” said Ghahramani, who has advised Microsoft Research and has received funding from Alphabet.

To try and avoid this sidelining, the Uber lab has two main programs: core and connections.

Core is about advancing basic AI research, which is a function of many corporate AI groups. Connections, meanwhile, is about working with all different parts of Uber.

For example, the lab needs to help figure out the best way to run its main ride service in different cities — but it must also provide insight into the time required to prepare and deliver food in the Uber Eats business, automate some of the work of Uber’s customer support department, and even contribute to the work of Uber’s self-driving car group, Ghahramani said.

“Ironically, I’ve spent my whole academic career engaging with academics, publishing papers and spending my time in universities, and so on,” he said. “What really excites me is the possibility of having an amazing material impact on our future world in a positive way. And so that’s going to happen through the connections, the business of Uber.”

At the same time, Uber’s AI researchers are still centralized. They are not embedding themselves inside other groups at the company, making them slightly different from the research teams at, say, Facebook or Pinterest. (The lab falls under Uber’s advanced programs department, whose mandate is to work on future products.)

To help with recruiting, Uber publishing research for anyone to read and build on. And on Friday the group took another step forward by releasing its software for free under an open-source license for the first time.

Uber is using the software, which goes by the name Pyro, to try to predict supply and demand for rides a few hours ahead of time. Pyro works on top of PyTorch, an AI framework that Facebook open-sourced earlier this year. The researchers are also bringing Pyro to Uber’s finance department to build forecasts of the next few weeks and months of Uber’s business.

The fact that Uber’s business is fundamentally based in the physical world, rather than being pursely about software or the internet, also helps the company recruit.

“The fact that we are so in the physical world, so in real time, so engaged with things like moving about and self driving and things like that that the actual artificial intelligence, cognitive science and machine learning problems are really pretty challenging and pretty different than a company that only interacts with people through a browser or something like that,” Ghahramani said.

“The physical world is pretty complicated and it makes you really stretch the stuff that you do.”