NLP is Quickly Improving. This Will Create New Opportunities

In our class this week, Yuri Balasanov covered the history and recent breakthroughs in NLP (Natural Language Processing). He was excited to share how much NLP has improved just since 2019.

He used this slide to show the explosion of pre-trained models and transfer learning (to make it easier to tune the models to different applications).

What does this mean for organizations?

This rapid improvement made me think about Azeem Azhar’s book, The Exponential Age. He mentions that our intuition fails us when things are growing exponentially. That is, exponential growth quickly overwhelms us and if we are not careful, others quickly pass us.

He uses digital cameras as a simple example. They had been improving and decreasing in cost in a exponential way. Slowly, over like a 10 year (or more) period, we moved from analog to digital cameras. Then, almost overnight, they showed up in our phones, on both sides of the phones, on our computers, and now new cars have multiple cameras. Not many of us saw this coming. And, this then created whole new types of companies like Instagram and Tik Tok.

An important point is that exponential changes didn’t just have us swap out our old film cameras for a digital one, it created all kinds of new opportunities.

I had the chance to see something similar in the supply chain optimization space. I joined LogicTools in 1998. We had a product to help companies determine where to best locate their warehouses and factories. To do this, we had to solve a complicated mathematical optimization problem— a large mixed integer and linear program. At the time the best general-purpose solver on the market was CPLEX.

However, in 1998, David Simchi-Levi, Edith Simchi-Levi, and Pete Cacioppi had built a custom solver for this problem that was a lot better than what CPLEX could do. We could do this, because we were solving a very specific problem. If the problem changed even a little from what we had specified, we would have to re-write the algorithm (which I think we did a few times).

With our faster solver, we won a lot of business.

But, a threat was looming: CPLEX was improving exponentially.

Around 2001, CPLEX was faster. I was slow to grasp this. I thought that if we made some tweaks to our approach, we would re-gain the lead. In retrospect, since they were improving exponentially (and could focus on just this), we would never come close again.

Luckily, we reacted fast enough. We mothballed our solver and used CPLEX instead.

It turned out that the faster solver was just one benefit of the switch. We were able to refocus our efforts on things like ease-of-use, new features, and new products. Improvements in just one area— the solver— made us re-think the value we brought to customers.

This brings me back to NLP. It feels like camera and solver stories apply here.

If people on your team are telling you that you need to swap out your home-grown NLP capability for these latest advances, they may be right.

More importantly, it feels like a good time to ask strategic questions about how you might use NLP in new ways, use it in different parts of the business, or how it might disrupt what you are doing.