How Companies Can Get Machine Learning Right: Q&A with Sean Gourley

Sean Gourley, the CEO of machine intelligence company Primer, talks to In The Future about what algorithms can do that humans cannot—and what risks companies need to consider when replacing human talent with machines.


Conversation has been edited and condensed for clarity and brevity.

Data has value—and for many companies, capturing the right data can be transformational. To better analyze and understand increased amounts of data, companies will need to use machine learning and automation. The science is there, but figuring out how to use it—and finding the right strategy—is daunting. According to Sean Gourley, an expert in machine learning and automation, companies need to start engaging themselves in Artificial Intelligence (AI) technology while asking some important questions along the way.

Sean GourleyGourley is the founder and CEO of Primer, a machine learning company building machines that can read and write—and automate the analysis of large datasets. Before he started Primer, Gourley was the CTO of Quid, an augmented-intelligence company he co-founded in 2009. The TED Fellow talked to In the Future about what it is exactly that algorithms can do that humans cannot—and what risks all companies need to consider when replacing their human talent with machines.

In The Future (ITF): How does Primer use machine learning to help companies grow and improve their businesses?

Sean: We focus on natural language understanding (reading) and natural language generation (writing)—and if you think of these two fundamental skills, you can start to replicate human activity. This is the kind of the work often analysts inside of large organizations, like banks, do. They process information—they read it, then try to understand it and write a summary of it. For us, we’re coming in and automating some of these processes so that analysts can spend less time on this type of work and spend more time engaging their critical thinking. We want to take the stuff that machines are good at, and the stuff that humans are good at, and hopefully create a better performing intelligence that’s a combination of human and machines.

ITF: In what areas do algorithms have an advantage compared to human employees?

Sean: There are a few advantages. First, machines tend to be relatively cheap compared to humans—they are a tiny fraction of an employee’s cost. The second thing is speed. For a human to process any piece of complex information, it will take around 650 milliseconds, and that’s just one unit– any decision is made after thousands of these units. Reading through a 10K filing, for example, can take huge amounts of time. And related to that speed is scale—humans tend to be experts in a small subset of things, but they start to lose the expertise that comes from connecting across different spaces and disciplines because they simply don’t have the throughput to allow for more complex interconnectivity and analysis. The final aspect is dimensionality: Machines are much better at dealing with high-dimensional structures.

ITF: How does this change the everyday fabric of work when it’s applied?

Sean: It manifests itself very often in analyzing information to improve recall. For example, when humans look at things, we tend to be quite precise and we think that we’ve seen everything. But the reality is that when machines do the same thing, they surface a whole range of things that a human missed. Where speed and scale come to bear for machines is really on this recall side of the equation. Because humans tend to look very narrowly at things, we’ll often miss connections that are important because they are outside of our purview. But machines can make those connections, and this ends up making them better at predicting things—connectivity is a very important feature.

ITF: Can you give an example of where machine learning would benefit a company?

Sean: Let’s start with the world of finance. If you’re an analyst covering a particular security, you’d start looking at keyword searches for specific mentions, which would pick up events associated with those key terms. But it relies on someone making a connection. Let’s say there’s a home improvement company and there’s a timber shortage based on a strike occurring in East Asia, at one of the ports. You will never find it unless someone has connected it to the specific company. But a computer can look through that and know there’s a high risk of supply issues out of East Asia for timber for that company, and it can monitor all of the events unfolding around the world and it can connect that, even though no human has made that connection. Humans tend to filter information down using proxies, in this case the name of the company, but we don’t tend look at all the connectivity that occurs and we tend to miss the issue of recall. The machines can make these connections before we are able to.

ITF: Do you think companies today are making significant progress using machine learning?

Sean: Not yet, since so much of the technology has yet to be commercialized. You can go back to the early 2000s and a lot of companies said they were using the Internet well. It raised this question—when does a company evolve from company, to online company, to tech company? What had they done? Taking a company’s experience and putting it on the web doesn’t make it a tech company— in reality they were still brick and mortar companies, just with websites. The real transition to a technology company happens when you start doing things differently and remove the need for physical infrastructure. When we look at this question of how well companies today are using AI, it’s not a question of whether they’re deploying machine learning and some models, because of course they are. The question instead is whether businesses are transforming their practices to take advantage of new intelligence made possible by AI.

ITF: With that in mind, how can a company measure the success of its AI strategy?

Sean: If you think about what AI really is, it’s the ability to automate cognitive processes: human thinking. How are companies restructuring their businesses to take advantage of a new kind of intelligence, machines? That’s the metric I would posit is the right one to gauge whether companies are effectively leveraging AI. If you take it a step further, businesses can also evaluate whether they are now doing things that have already been enabled by that new kind of intelligence. These are the things that they could have never done because humans couldn’t think fast enough, for example, enabling a processes to unfold that require the reading and comprehension of millions of documents in under a minute.

ITF: And how do algorithms fit in alongside human employees—should we be concerned about our own fate?

Sean: There are some tradeoffs to be made here. If a human costs $200,000 each year in salary and benefits and a machine $2,000, you end up getting different things. The mechanics of this is not whether you’re replacing a human and getting the exact same output. Another piece is that humans and machines don’t think the same way or react in the same manner, because we’re different kinds of intelligences. Humans and machines aren’t going to make the same set of mistakes—they’ll be different. Now if you look at automation, for example, it can smooth out the day-to-day variance and a lot of things become a lot smoother, more stable and safer.

ITF: Algorithms can go wrong. What kind of risk do companies then need to consider?

Sean: Automation can bring serious tail-end risk with it, so when bad things do happen they tend to be quite catastrophic. You can see this in the automation of financial markets with flash crashes. As a society, it’s one of the questions for us—are we prepared for these big tail-end risks? That is the cost of this much smoother, much more stable and efficient process. I think that is a question to weigh, not just from companies but also from regulators. I think that if you automate something fully and remove the human from the loop, you remove checks and balances that can expose you to these tail-end risks. Keeping a human inside the loop allows you to have two different intelligences that will actually provide two different catch points that should hopefully yield a better performing system. What is comes down to is designing interfaces that allow these two parties to work together effectively.

ITF: What questions do companies need to bring up in order to mitigate those risks?

Sean: First and foremost, for any algorithm that you’re dealing with, you should ask: what’s its precision and what’s its recall. When it makes errors, what kind of errors does it make? Are they little or big? What data was it trained on, and was that data representative of the world we’re going to be in? How vulnerable is the algorithm to attack? At the board-level, having that conversation probably gets people a lot further along the understanding of the risk.

ITF: What advice would you leave with an executive tasked with an AI strategy?

Sean: We’re just seeing the start of this transformation. The science that’s been done in the last five years has been huge, the advancements have been massive, and very little of that has been commercialized. I would get a team to look at that science and build prototypes of things that could change the way you do business. If you’re not actively engaging that science and building things, you’re not going to build up the muscle as an organization to deal with it—which means you’ll be beholden to someone who is, or worse off, you’ll be left behind.