Interview with Andrew Ng, founder and CEO of Landing AI and deeplearning.ai
Few executives have stood closer to the leading edge of artificial intelligence (AI) than Andrew Ng.
Ng is the founder and CEO of Landing AI, an enterprise artificial-intelligence company. He has directed Stanford University’s AI lab and led the Google Brain AI research team. As chief scientist at Chinese search giant Baidu, he developed innovative image- and speech-recognition applications. When it comes to grasping both the technology and AI’s business potential, few can match Ng’s resume.
The rapid rise of AI means executives have lots of questions, such as how to manage computing systems that are beyond human comprehension, and how to ensure they’re used ethically. Ng considers another set of questions even more important: What type of organization is needed to succeed with AI? What skills do you need to recruit for? What organizational structures work best? Do companies need an AI boss in the C-suite?
“A lot of companies are realizing the importance of how AI will impact their business," Ng says. “And they’re starting to build in-house AI teams.”
Ng spoke with Forbes AI about how companies can build well-rounded teams to get the most out of their AI.
Andrew Ng
Executives don’t have much experience building AI teams, primarily because AI wasn’t as big of a necessity as it is today. How should companies go about selecting AI team members?
You need talent that has a deep knowledge of modeling. They need to understand the capabilities of computer learning. Then this centralized team can work cross-functionally with all business leaders to develop specific AI applications. Ultimately it will be the teamwork that drives the application.
In most companies, AI will be used for multiple projects. For example, AI has been reducing work in agriculture. There are dozens of applications—everything from weed control to interpreting satellite imagery to optimizing the harvesters. It’s very difficult to hire AI talent that knows all of these vertical applications.
That’s why it’s critical to test a few applications first to really understand what AI can do in one particular area of business.
Learn more about how companies are leveraging AI today.
What team structure enables this?
A small team empowered to act quickly and even fail should be able to test an application within six to 12 months. Giving the team its own budget, rather than forcing it to be funded by the business units, can give it a faster start.
What roles and skills do you need to recruit for when building an AI team?
The leader of the AI team—whether a CAIO (chief AI officer), VP of AI or other role—needs to understand enough about the technology to have a sense of what can and cannot be done. Further, they need to work cross-functionally with business leaders to identify and drive projects forward in AI+X, where X is whatever sector AI is meant to add value to.
The AI team also needs the engineering talent to execute on the AI project. Depending on the specific project, this would include machine-learning engineers, data scientists, applied scientists, data engineers or other roles. Some teams will also have product managers.
How have you seen the role of chief AI officer structured?
They’re often very high up in the organization. The chief AI officer’s responsibility should be to build out AI capabilities for the whole company.
Remember that a hundred years ago companies were hiring vice presidents of electricity. They were brought in to support the whole company and dictate how to handle this newfangled thing called electricity.
AI today is in that early stage of development. So you need that chief AI officer working closely with the CEO or CIO to give them the ability to drive change across the company. AI is a general purpose technology; it’s useful for a lot of different applications. If one single business unit deploys AI it will only transform that one business unit. At Google we [Ng’s team] would have only transformed the speech functionality if we hadn’t been reporting to Google’s CEO. Instead we could change the whole company.
What are the core functions of a centralized team model? Who does this well today?
Google and Baidu do this well. Both are great technology teams, and have proven their ability to work cross-functionally to create concrete AI value through improved web search, advertising, speech recognition, product recommendations and many, many other projects. Facebook’s FAIR and Microsoft’s AI work have also made huge contributions.
One challenge of the central team is that it takes years to build. At Landing AI, where we have invested extensively in technology capabilities for accelerating AI adoption by enterprises, we’ve found that making these capabilities available to our partners can significantly speed up their ability to identify and realize practical AI business value.
How should this core team connect to other business units?
The central AI team can build company-wide platforms. For example, at one of my previous companies, my team owned the user data warehouse which served as the central repository for all user-related data. This connected with multiple business units and aggregated data from all of them to then also drive value back to all of them (subject to privacy). No single business unit could have done this, and it made sense for the AI team to have this be a centralized platform for the whole company.
The job of the chief AI officer is to look across all lines of business and say how can we use AI to improve. When I was leading AI at Baidu, we had a product management team whose role was to go to every business unit and figure out whether they could help improve areas like the quality of web search, the contents of a map, recommend better movies or how online shopping worked. Their job was to find opportunities.
And of course, we always had far more opportunities than bandwidth. That’s just part of that world. So the other piece was building business cases and really understanding from the business leaders how badly they needed AI. And then—fortunately or unfortunately—we had to prioritize. There were always more ideas than we could execute on.
What was the structure that made that possible? Do you consider the Baidu structure a good model in general?
The Baidu structure delivered great results. Many elements, such as the idea of having technology teams and a systematic process for marrying breakthrough technology and concrete business applications, are useful to every company. But organizational structures are not one-size-fits-all, and depending on the business context, a company may invest more or less in platforms, or data engineering, or hold different portfolios in basic research versus advanced development, and in different sectors such as computer vision, natural language processing, general machine learning, etc.
Was there a dedicated group at Baidu that was responsible for building business-use cases for AI?
I had a team of [product managers] responsible for building bridges to the business units and identifying, prioritizing and then tracking execution of AI projects in different vertical areas. The product managers, some of which were former engineers, had to have sufficient understanding of AI to have a sense of what it could and could not do, though they would also often check with the more senior engineers for a more definitive conclusion before committing to a project.
We welcomed ideas from anywhere though. For example, some of the best ideas for new AI projects came from managers in business units, others came from the technology units. The more important thing was to have a process for screening ideas and letting good ones get resourced.
A small company of 100 people can’t just deploy a new AI team. How should they think about AI for their business?
The advice I give to everyone is figure out how to jump in. Even if it’s just a junior programmer on small projects, get the wheels going and feel what an AI application can do. Companies tend to want strategies around everything. But if a company’s never done an AI application, they can’t strategize properly, so C-suites develop strategies that look cut and pasted from a newspaper headline. Someone else’s strategy is rarely right for them. So I say: Just hire a couple engineers to see what they can do, and keep growing from there.
Do you recommend in-house AI even for companies that can only hire one or two developers?
It depends on whether you can come up with a concrete thesis for AI value creation, and whether it is more efficient for your company to build or buy. If an application is likely to become an industry standard, it’d be better for a small company to buy. But if an application is valuable for a business, and is sufficiently unique to the company that it cannot be built efficiently by a third party or if the company needs to hold the IP asset closely, then they should build.
CREDIT: Hero Images/GettyImages;
