Pylimitics

Minas Karamanis has done some thinking about using LLMs. How they can help and, on the other hand, how they can undermine your own abilities and career. This is probably obvious, but it all depends on how you use them. Karamanis is an astrophysicist at Berkeley, but I think what he says applies to any profession.

Karamanis points out that “…you cannot learn physics by watching someone else do it.” That’s why every physics textbook includes problems; the process of trying to solve them, and especially the process of failing and trying again is how humans learn. If you just read the solutions and think you understand, you don’t. You won’t be able to solve a real problem, because all you’ve done is scan the solutions, not trained your mind to understand the problem in a way that you can find your own way to the solution.

There’s a problem though. In academia, just like in computer programming or countless other endeavors, people are paid to “be productive.” I don’t know anything about academia or astrophysics, so I’ll use the computer programming field as an example. I might have the task of writing a program to do a particular task. The tl;dr of what I do is figure out what the task involves, how to proceed step by step to accomplish it, write instructions in a programming language to do those steps, then try the program to see if my solution works. Let’s say it takes me a week. Some people could do the same thing in three days, some might need two weeks, and most people, who are not programmers, would never manage it (and would look at you incredulously if you asked them to do it). An LLM, though, could write the code and test it in a couple hours (probably less).

The company I work for is urging employees very strongly to use LLMs. Ostensibly to make us more productive. This urging is as always communicated in corporate jargon that emphasizes excitement and success and opportunity and all being in this together. Many commercial organizations compete to produce more and faster, hoping to outperform their competitors. We have to get our tasks done this month, this week, today because the end of the quarter is looming, as it always does. And the end of the quarter counts as “long term.” Executive jobs are supposed to involve lots of planning and thinking ahead, but they give little attention to the future that’s five, ten, twenty years away. You can’t think that far ahead.

Except that we can, and we do all the time in regard to ourselves and our personal lives. It’s our organizational selves that seem to be constrained to immediate, very short term thinking. Get that product to market now; if there are problems we can just address them later. Nobody seems to notice that when later comes along, we don’t have time to address the problems we skipped because by then we’re busy with a new immediacy.

Programmers almost have to use LLMs these days just to keep up. And remember those people who would stare at the boss in amazement if asked to write a program? Well now they can. All they have to do is prompt an LLM to write the program, and even test it. Anybody who can write a sentence can do that. What’s missing is learning.

Whether it’s programming or plumbing or managing a business, learning a skill is not a short term project. You make a lot of mistakes along the way, and the result of a few of those mistakes is an aha moment when you suddenly understand something. It’s not really sudden, but it feels that way. You have to make all the mistakes to get to the understanding. Karamanis calls it “building the infrastructure inside your own head” that enables you to understand whatever you’re engaged with.

It’s too easy to use LLMs in a way that doesn’t build that mental infrastructure. It might not just be easy; it might have some short term benefits. You seem to be more productive. You might be able to finish more tasks (especially mundane repetitive ones) in the next couple weeks. Your organizational self might even be rewarded. But your personal self? Five years from now? That self will be less capable, have fewer skills, and be diminished in comparison with could have been. Using AI can be like deciding about eating dessert. Short term, yes, it’s pleasant. Long term though, might be different.