I recently had the opportunity to take a course on AI from Columbia University and Wall Street Prep. The course promised both a crash course on how AI can be leveraged by corporations, as well as exclusive conversations with professionals leveraging AI. The tl;dr is that, while the course was thorough, it ultimately taught the wrong thing and was basically out of date on launch.
In the past, I’ve taken several online classes in Analytics & Machine Learning, which were the buzz words at the time. The main work of these classes was threefold: review the basics of statistical analysis, teach the tools that now make the hard part easy (i.e. all those calculations and graphs), and develop some intuition about the types of calculations and graphs that will lead to insight. The overall effort made me better at engaging with data to solve problems. It was also something I could genuinely apply immediately (or almost immediately, if real world data were actually tidy and clean, which it never, ever, ever is). I was excited, if a little bit dangerous, about what I could accomplish.
Enter “AI,” and by that I mean LLMs that are also Machine Learning, the learning they’ve done is just on a massive corpus of human language and not proprietary data sets. As of December 2025 when I took the course, what was being promised was literally replacing the entire workforce, meanwhile most individual users struggled to find genuine applications for it. (After all, what’s the ROI on writing an email in the style of Shakespeare?) And of course, we’re all assuming it’s user error, and not the immaturity of the tech.
To my surprise, the course was 80% regular machine learning stuff (or as they called it “classical machine learning,” I’m so old) : a bootcamp on python, data management, statistics overview; then a few modules on how to query ChatGPT through API tokens to get it to tell you things.
Most of the examples involved getting ChatGPT to do some of that data tidying that is oh so troublesome in real life, such as matching a string in a text input that might have typos and other variations, which seemed reasonable, if limited, to me. But there was one example I found problematic. In order to “test” an idea, they queried Chat and asked it to rate it as if it were a Grad Student, a Teacher or another role, essentially using its vast knowledge of the human experience to rate a business idea. I was incredulous and kept waiting for them to define what they thought each of those personas was, but no, they were just vibing with Chat and letting it decide what, if anything, those people would want. That seemed shady to me, and if someone had given me that as an application of something they’d done, I would make them go to their room without supper.
Overall, my main impression is they hadn’t cracked what exactly these LLMs are really for.
I watch a lot of old TV shows from the 80s and early 90s. Most of it is nostalgia for the slowness of transportation, the slowness of communication, and the complete lack of existential dread. It’s the little things. But I usually bail on the episodes that involve the plucky protagonist having to solve the murder by using these new-fangled computers. What exactly they think they’re for or how they actually work feels so hokey here in 2026 (though I’m never bothered by hailing a cab, using an answering machine, or the unassailable conviction that education is the root to advancement in society.) I feel like LLMs are still in this moment. They’re here, they’re new and we know they’ll do stuff, but until we collectively agree what that stuff is, some of the uses seem a bit try hard rather than innovative.
The one thing I will say is, most of the code I wrote for the class I wrote with a lot of help from Chat. So if there is one take away, machines do a damn fine job talking to other machines.