I have been thinking a lot about how much AI models have improved in what feels like a very short time. From the release of GPT-3 to GPT-4, the growth of intelligence was staggering. However, with recent incremental steps, it's becoming easier and easier to rely on the outputs generated by AI. This has led to something I have come to view as knowing without understanding.
By that, I mean that we're seeing a growth in what the average person is able to produce, without necessarily understanding how or why it works. I have previously referred to this as normalization as a service, and a lot of what I reflected on back then still holds true, but it also just feels different now. When you view creative output - PowerPoints, text or images - it's quite easy to spot what has been created by a person and what has not, but now - I'd bet - most code is being written by AI. At the same time AI labs are releasing one mathematical breakthrough after the other.
Despite these breakthroughs and the objective capabilities of newer models, I am also noticing more and more of a gap in the usage of the models. An intern is perfectly capable of writing a technically sound and grammatically competent board meeting summary, but while it might objectively be of high quality, it might not be subjectively good. Another, more experienced colleague might use the same tool and get it right in the first shot.
The example above is not just a made-up example, but an anecdote I actually heard the other day. I find it fascinating how having access to a tool capable of both transcribing and doing the following analysis of the transcript can lead to such different results depending on the input. It also makes me wonder if that is how it will be forever, or if this is simply because we're still in the early stages of development of what we now commonly refer to as AI. A rather positive point of view on this phenomenon could be that no matter how much technical competence the models grow into, we'll still need to hire for junior roles, simply to train people to have the taste that LLMs don't seem to possess.
At the same time as experiencing taste as more important than ever, I experience something that I almost struggle to identify as anything but a mixture of FOMO and imposter syndrome on a daily basis. I have convinced myself that I'm good at using AI models to learn new skills, and if you look at the way I use Excel compared to a year ago, that holds true. If you look at the SQL expressions I'm running, Power Automate solutions I'm architecting or websites I'm building, I'm more competent than ever. But at the same time, the amount of JavaScript, Python or even HTML and CSS code I've written in the past six months is significantly lower than it was the six months before that. I feel my competencies growing and gain better insights, but still feel I might not be understanding the nitty-gritty details the same as I might have before.
It's something that I'm very aware of and take seriously - using LLMs as tools, not as a replacement for my own thinking. It helps that I am reminded on a daily basis of what I'm not trying to become - frustrated that I can't do my work because I'm out of tokens, complaining that the AI is not good enough at the task I've given it, or simply replacing my own understanding or desire to learn new things with the most recent output from Claude.
Friends and colleagues know me as a notorious notetaker, so the example with the board meeting summary hit somewhat close to home. Being able to take notes during class is what carried me through university, and was later a crucial ability during interviews when I worked as a recruiter. Yet despite seeing myself as a decent prompter, I can't get an AI to write a better meeting summary than the notes I took during a meeting. On the other hand, to get around that, I found that especially Claude in PowerPoint is quite good at turning these same notes into highlights in a PowerPoint that I can later share with my colleagues alongside a .md file, if they want to dive deeper into the key points that are made in the slides. I find it endlessly fascinating how it can struggle with one side of the coin - the summary - and effortlessly put key points and graphics in a PowerPoint based on the same input.
I also find that somewhat analogous to what is going on in mathematics. Given that I'm able to have a conversation or participate in a meeting while taking notes, I don't find the idea of an AI notetaker awfully appealing. It's a skill that I have, something that I enjoy being good at, and something that I don't feel a need to replace. From what I understand, something similar is happening for mathematicians. What they have been working on for years is now being done by AI labs. The papers might not be interesting beying the result, without taste, but the end results are real (except, of course, when they get taken back later!). But where I might get a bit better at structuring my summary while writing it - real, but not significant - a part of mathematical discovery is the way you get there, the understanding you build along the way, and the connections and understandings that are being built on the way to the result. As such - and depending on the field - the final output is not always as important as the process to get there.
I find the example of math analogous to other places where AI is being used more and more. You might be able to build or write something faster, but without understanding the whats and whys, the result is often lacking.
When I started in my most recent job, much of my onboarding was of a similar nature. It was all do this, do that - only focused on how to get to the end of a task. When I asked about why we did certain things or suggested changes to the procedures being followed that seemed more intuitive to me, I couldn’t get a clear answer on why, or on whether my proposed solution would work. It was then, after identifying why we do the things we do, that I realised the importance of understanding instead of just knowing. Back then, if I had just accepted the way we worked, we would still be doing things the same way today. Of course, that's not to say they were bad ways, but by building an understanding of the processes and systems we rely on, we have since been able to build improvements and more efficient processes in almost all areas that our team touches today.
This is also why I feel a genuine sense of worry when leaning into using LLMs for cognitive tasks. It's a real balance that one has to identify for themselves, and one that I'm constantly reflecting on. How do I become more efficient, and do the amount of work that I know I'm capable of with these new tools, without losing the skills I've built and while improving every single day? It's something that I'm not quite sure I've figured out yet, but looking at where I am now compared to a year ago, I'm thankful I'm aware of it, and reminded to stay on top of my game.
As with my experience when I started in my job, and with the importance of the process in math breakthroughs, I think it's important that we understand LLMs or any other new technology that enters our lives. Of course, that's close to impossible, given that LLMs are closer to being grown than built, and no one truly understands them. At the same time, I believe most people will be able to understand the basics. Invest the time in understanding context windows, experimenting with what context gives the best results, and how you prompt your way to a result you like. I'm also somewhat fortunate to have taken a Python class during my bachelor's where we did some light text and semantic analysis, using libraries to identify word stems and sentiments of a text. That course in particular is something I'm only growing more and more thankful that a friend convinced me to take back then.
Every day I'm talking to colleagues who are facing issues or inefficiencies in their use of AI that could be fixed or significantly improved by a better understanding of How the models actually work. As I reflected on in my last post, a way to do this is to identify what in a process is testable, and what can be turned into something testable - how can we turn our processes into code?
Knowing gets you the result. Understanding gets you the next one.