Andrew Ng knows a thing or two about artificial intelligence.
The former head of Google Brain and prior chief scientist at Baidu, Ng also co-founded Coursera and regularly teaches popular courses on the technology online and at Stanford. And he runs Landing AI, which provides manufacturers (and soon, other industries) with an AI platform to help developers more easily build and deploy computer vision models.
That experience has given Ng a deep understanding of the benefits that AI can produce — and the limitations of the tech. As Ng expands his work outside of consumer internet companies, he's seeing a pattern: Organizations are setting their AI ambitions too high.
"I still see companies jump in and make investments in projects that I would consider technically impossible or technically not feasible with today's technology or the near-term generations of technology," he told Protocol. "Learn to walk first. It's fine that the first project you do is not a $10 million AI project."
But even landing on a smaller project can be difficult due to the nature of AI initiatives, which often span departments and require cross-functional leadership that many organizations are still working to develop. It's why Ng says a central AI group is so paramount.
"That'll take some top-down leadership to put in place," he said.
Protocol talked to Ng to learn what he thinks enterprises are getting wrong about their AI strategy and why investments in MLOps should be the wave of the future.
This interview has been edited for brevity and clarity.
As we see the excitement around AI continue to grow, where do you think the broader market is? Is the tech more hype than reality at this point?
AI isn't one monolithic thing. So there are some segments where the hype is definitely disproportionate, but also some segments where there's a lot that's not as visible to the wider public. [Artificial general intelligence] still has a little bit too much hype, though it's come down a little bit. On the flip side, there's a lot of industrial, B2B applications of AI that are valuable but not as well understood compared to a B2C application, which is much more relatable.
Where would you say enterprises are at in this journey? I know there'll be differences, but it seems companies are now starting to take what were smaller pilot projects and expand those.
Very early. For the large companies, even the Fortune 500, some are further ahead but many have one AI project that was put into production through sheer heroics that is reaping substantial amounts of value and potentially dozens of pilot projects that could be promising but, at least on the current path, may take heroics again to put into production. The more traditional industries, where the digitization wave came a little bit later, are still very early.
You've been vocal about the need for quality data over investments in models. Are there sources of data that companies are ignoring?
I feel like the answer has got to be yes. The more common pattern is there are a lot of companies with data sitting around that [is] already good enough to create tremendous value. All data they can very easily create. Take ecommerce. Tons of companies have tons of user data already sitting in their data warehouse and an AI team would be able to go in and drive insights.
What is the best way for organizations to begin this AI journey? How do they have to be set up to be successful?
One of the most important steps is to deliver a quick win. Small pilot project and then take it to a successful outcome. And that initial quick win often teaches an organization lessons that would then be useful for the second, third and fourth projects. Too many companies start off wanting a grand plan. But until it's learned to walk, it's very difficult to plan out what to do when you cross the finish line of the marathon. Learn to walk first. It's fine that the first project you do is not a $10 million AI project. It's fine that the first project you do is a $200,000 project — or even a $50,000 project. The purpose of that is not necessarily to create massive ROI. The greater value is the learning.
Are there any other ways companies are getting their AI strategy wrong?
One of the challenges of AI is it takes a decent amount of technical knowledge to figure out what is and isn't technically feasible. I still see companies jump in and make investments in projects that I would consider technically impossible or technically not feasible with today's technology or the near-term generations of technology.
For example, building a chatbot that can handle all customer service requests in a fairly conversant way. It's clear that's not possible. Fairly recently, someone asked if I could help them build the equivalent of a self-driving car with six engineers in six months. I don't think I could do that.
Project selection is still really difficult, because it takes cross-functional business and technical judgement to prioritize projects. Only a centralized AI group can build horizontal platforms that span the entire company, so that'll take some top-down leadership to put in place.
What did you see in the manufacturing industry specifically that made you start Landing AI?
Speaking with a lot of C-suites about AI adoption, [I] saw many of the same problems over and over in terms of practical deployment. There are lots of $1 [million] to $3 million projects. And it's challenging to get the AI talent and the staffing to make the economics workout. Tons of projects were stuck in proof of concept, because even if a company developed an AI model, it's difficult to write all surrounding software — MLOps is sometimes what we call that — to take that system into production.
We ended up building LandingLens, which is a data-centric, MLOps platform for computer vision. We help companies — starting in manufacturing but we have interest in other computer vision vertical applications — be 10x more efficient and often much more successful as well in building and deploying computer vision systems.
It seems the prevailing notion for why companies adopt AI is to cut costs. Do you find that, based on where we are at in the life cycle of AI, that immediate outcomes should be around quality improvement? Does it make a difference in terms of success which metrics prioritize first?
Cutting cost is a worthy thing to do and improving revenues or improving margins is a worthy thing to do, but I find that the latter category of projects often has more momentum than just cost-cutting. It's easy to get momentum on projects that create value beyond cutting costs.
What would be your one piece of advice for enterprises struggling with their AI strategy?
Find the right philosophies and MLOps tooling, because that will give organizations a big boost in AI adoption and performance. We've moved past the era where it's about the engineer using their own tools. Until now, a lot of AI was developed using very broad tools. We did that for the past decade, we're now moving on. I don't write any code in assembly myself. In the future, we'll find that the tooling will make machine learning engineers much more efficient.