LLMs, Learning, and the Value of Toil

Large Language Models (LLMs, like Chat GPT and others) are changing how we create; perhaps more profoundly, they’re going to change how we learn. As we learn new ways to learn, we’re all going to have to recalibrate our attitudes on what constitutes value.

I started my career as an analyst at an investment bank, where a large part of my job was analyzing financial statements and building models to see what future cash flows might looks like. The job could be broken into two parts: 1. Obtaining the data and 2. Analyzing it. Finding the data involved heading down to our library which housed trays of microfiche (look it up). I’d print out the relevant pages from a 10-Q or 10-K on particularly odiferous paper, then take it to my IBM PC to manually input it into Lotus 123 running off 5 ¼” floppy disks (again, look it up!). This first part, while necessary, was an intermediate step that provided no insight, even though it did provide raw information that was somewhat difficult and costly to obtain.

The next part, though, involved poring through those financial statements and reading the footnotes to construct a forward-looking model that drew on the past for insights. It was time consuming, but as I examined financial minutiae at midnight I was also learning the ins and outs of cash flows and accounting. That base knowledge made me a better banker later on and was, I thought, necessary to have a deep understanding of how companies could use their financial statements to shape perceptions of reality.

Indeed, one concern I had with offshoring the work of junior analysts a decade ago was, “Where will the bankers of tomorrow get their training? How will they learn what’s anomalous? When will they gain the exposure that makes them proficient at pattern recognition?” The advent of LLMs, though, has made me reconsider what was, in retrospect, a naïve view.

When you say out loud, “because I suffered to acquire knowledge, my knowledge is more valuable than your equivalent but more easily acquired expertise,” you realize how absurd it sounds. Yet that attitude is the default for many who slogged early in their careers, whether assembling models and pitch books; creating presentations cobbled together from in-person interviews and extracts from physical government publications; or debugging thousands of lines of code.

Photo by jesse orrico on Unsplash

Today we don’t necessarily need to understand the inner workings of a financial statement to be a financier, just as we don’t need to understand how an engine works to drive a car, or how to write code to use computers. Over time, in fact, we move farther away from the underlying base principles (which, to be sure, will still need to be understood by a small number of hyperspecialists) to more and more abstract user layers. I’m not sure that my yet-to-be-born grandkids will ever know how to drive a car; my kids are certainly the last U.S. generation to know how to drive a stick shift!

Now, don’t get me wrong, toil has its merits; it fosters grit and resilience, for example. I’d submit, though, that the difficulty in acquiring knowledge often leads us to overvalue the worth of the lessons learned. This is particularly true when deep knowledge doesn’t necessarily translate into practical understanding or even common sense. Looking back in the other direction, my parents’ generation knew how to balance a checkbook, repair household appliances, and sew clothes. Online banking, replaceable gadgets, and fast fashion offer simpler alternatives.

It’s natural to ascribe value to toil, and to believe that if we had to struggle, so should others. But we should focus on the end – the knowledge, insight, and even wisdom – not the means used to acquire it. And to be sure, there’s a danger that people will have a superficial understanding of a subject without recognizing the subtleties that elude them, and we’ll certainly need to guard against that.

Nevertheless, practical understanding is often – but not always – more important than deep comprehension. We can bemoan this new soft generation, just as the ancient Greeks did, or we can recognize that they have more time to focus on creating genuine insights and drawing new connections (even if AI-aided) now that the grunt work of collecting data has been reduced. We need to focus on the outcomes, not the inputs, and judiciously use whatever tools we have at our disposal to make the process both more efficient and more effective.

I’m now convinced that toil for its own sake isn’t worth it. Darn kids.