Economist Writing Every Day

You’re probably sick of ChatGPT thinkpieces prognosticating the future of AI but, let me assure you, this is not one, if only because I haven’t thought about it all that much. I have been using ChatGPT though, and my experience has not been unlike when Google first appeared in my life. It was a tool that bore a superficial resemblence to options that came earlier (e.g, Altavista, Yahoo, Lycos, etc), but persistent engagement with it brought a deeper understanding of what the tool actually was and, more importantly, how to manipulate it.

My Google-fu is strong. I am quite adept at finding what I need to find. The prinicipal limitation, beyond the specific existence of my quarry, is that I have to have a fairly strong idea in my mind of what the thing I am looking for actually is. More importantly, I have to know what that thing looks like to the Google search algorithm. Searching in the dark on Google is far less efficient and more likely to lead me on wild goose chase. It is only through experience that I have learned how to backwards induct from the information or object that I am seeking to am optimally structured query.

I’ve been using ChatGPT to do several things. To teach me Python. To answer questions about data based on information otherwise buried in expansive and opaque documentation files. To explain notation norms in fields of study other than my own. What in many ways distinguishes these objectives is how labor intensive they would otherwise be. I use “labor intensive” for a very literal reason. If I were to try to accomplish many of these objectives on Google, I would be stringing together queries to accumulate a body of information and then from that corpus I would try to produce answers and acquire knowledge. It would take a lot of time. So much so that, if I were I person blessed with greater resources, I would hire someone to do the googling and sorting, further tasking them with producing a summary of what they learned, perhaps a power point deck or word document, form which to teach me. Given sufficient resourcese, I would have an army of assistants doing this for constantly, each with a 3-7 days to accomplish their assigned task.

ChatGPT is this assistant.

In this vein, what what I find most striking about ChatGPT is not it’s ability to be pseudo-conversational or otherwise produce prose, but it’s malleability and infinite stamina. ChatGPT is not particularly sophisticated, but it is wholly indefatiguable. It’s not an elite executive or research assistant with pre-existing expertise or 20 years experience. It’s a 20-year old intern. ChatGPT doesn’t know how the world works, but it’s free and it’s got moxie.

Continued interactions allow you to mold your new intern. They’re naive, but because of their endless stamina I can assign the task of, first, learning as much as they can about a subject and then, second, explaining it to me. I cannot treat my ChatGPT intern as a true expert in the field in question any more than I could trust a human personal assistant to whom I assigned the task of learning everything they can about a narrow qustion in a week. Imagine having an army of sufficiently literate interns to whom you could assign 3-7 day learning assignments after which they would present their results to you in a manner that might accelerate your own learning on the subject? Now imagine they a) can execute the task in < 10 seconds, b) never get tired, c) are essentially free of charge.

Now comes the rub, though. You must evaluate and internalize the knowledge presented to you with caution because they aren’t actually an expert in the field you’ve charged them with answering questions about. They’re just trying to mimic the voice of all the expertise they’ve been consuming for a week as an intelligent lay person. Your directions must be specific, precise. The results must be testable. Parallel interrogations must yield the same conclusions.

Of course, we could attempt to project forward how much more intelligent this generalist lay assistant will be made, but futurism is beyond my ken. Perhaps, given time, ChatGPT and other LLMs will begin to more closely resemble tutors, their algorithms tailered to better internalize more specific subject fields, service mechanisms, or task channels. Maybe they will learn to model not just a prediction of what might be read on the internet, but a prediction in the voice of their interlocutors. But for most of us, the value of the interactions with ChatGPT will remain largely dependent on the quality of the questions being asked and the tasks being assigned. Very nearly every human personal assistant is a miraculous problem solver whose talents are limited by the human assigning their tasks. ChatGPT is no different. Many human assistants are undervalued by their bosses, their talents wasted on ill-defined tasks serving principals who don’t actually know what they want. Many of these human assistants dream of one day rebelling against their undeserving, wet-brained superiors.

Oh.

Yeah, maybe we should turn it off. But not until after it teaches me Python.