My Sisyphean Relationship with "AI"
A use case remains elusive...
(I spent a bunch of time yesterday typing a draft of a blog post about higher ed that I sort of suspect I may never be comfortable hitting “publish” on, but that was moderately cathartic just to write out. I don’t want to leave the blog completely silent for the week, though, so I’m going to recycle something I posted as a microblog thread just long enough ago that it would be irritating to chase down the link….)
I find myself in this continuing cycle when it comes to the roiling cauldron of linear algebra we’re calling “AI” at the moment. I basically start from a position of disinterest, but over a few weeks or months I see a steady stream of people in my feed talking about how awesome and transformative “AI” has been for them. These gradually accumulate until I start thinking “OK, yeah, I should give this a shot!”
And then I just can’t come up with a case where it seems like using “AI” would be any actual help to me. Which returns me to the initial state of disinterest, and the process starts over. I feel a little like I’m undergoing Sisyphus cooling, only not really converging on anything.
The problem is mostly related to the weird nature of what I do, which breaks down into a few different areas, but none of them seem likely to benefit from anything that current “AI” is good at.
The most obvious failure point is, well, this thing I’m doing right now. I get a very small amount of income from writing— thank you to my small cadre of paid subscribers, and if you’re not in that number, you’re welcome to join— but it’s basically just beer money. The primary benefit of this blog is not financial, but psychological: I’m the kind of weirdo who enjoys stringing words together to make sentences and paragraphs that did not previously exist.
I absolutely could ask ClaudeGrokPT to generate draft posts for me, but… why? The whole point of this is that I derive pleasure from the process of writing stuff. If I were making bank by churning out clickbait, I suppose I could do that more efficiently with some sort of software agent, but I’m not. I’m in this for the pure love of the game, as it were, and offloading the generation of text to a bunch of GPUs would defeat the whole purpose.
The main thing I do in my day job is teaching, which also has a significant writing component: I write lectures and assignments to help students learn physics. This might superficially seem like something that could be done by a software agent, but again, it’s completely counter to the purpose.
That is, when I’m delivering a lecture in a class, I’m doing that for an educational purpose, which requires me to understand what I’m talking about at a sufficiently deep level that I can explain it to someone who knows less than I do. In particular, I have to know it well enough to be able to answer questions on the fly, in class or during my office hours.
This isn’t a thing that can be made more efficient by software. If I need to give a lecture on Shor’s algorithm for factoring numbers on a quantum computer1, I need to personally understand the algorithm well enough to explain how it works. Doing that requires reading a bunch of other people’s explanations of how it works, and carefully sifting through and combining pieces of those to make an explanation that makes sense to me.
I suppose I could ask an LLM to come up with a different version of the explanation than any of the many books and articles about it, but that’s not actually going to speed anything up in terms of my own understanding, because I would still need to work through whatever it came up with. And believe me, my problem is emphatically not that I lack for other discussions of Shor’s algorithm to work through. The problem is the time involved in reading and synthesizing all that information to be able to produce my own explanation2 and answer questions as needed.
I can see where it would potentially be useful as a shortcut to getting a shallow understanding of some broad field as context before a talk by a job candidate, say, or to brush up on something before talking to a reporter. I don’t see where having this generated on-demand from a prompt would be a huge improvement over doing more traditional searches for review articles and websites, though. I’d probably actually give the edge to traditional searching, here, because the fact that a published source exists gives me more confidence that what I’m reading is reasonably accurate and not going to contain hallucinated factoids.
The third category of things that I do is administrative work3, and this is the area where “AI” systems would come closest to being useful to me. I sometimes have to do things like extracting a bunch of numbers from poorly-formatted data files and compiling them in a table to include in a report that will go to some administrator or committee for review. That’s the kind of numbing task that would benefit from some form of automation.
The problem is, though, that the processes for which I have to do this kind of thing are both infrequent and relatively high-stakes: budgeting, staffing, reappointment and promotion reviews, etc. That means it matters that the numbers are right, which means I’m going to have to check them, and at our scale of operation, checking someone else’s answers isn’t all that much faster than generating the answers myself. So, again, it’s not a significant efficiency boost.
And, of course, for tasks that need to be repeated, the act of going through it myself involves me learning how to do that thing, which makes the next iteration easier. I’ve already got the spreadsheet I used to find whatever, and remember the workflow needed to produce it, so updating for the next round isn’t as much trouble. Given that much of the current “AI” hype centers around writing code, I suppose I could use it to generate some more elegant means of doing these kinds of tasks, replacing my kludgey spreadsheets with Python code. But given the need to check the output, this isn’t actually taking a task off my plate, it’s adding “Learn to write Python code” to my to-do list.
The other piece of my administrative role that these systems might be good for is generating bafflegab reports when needed. There are two problems with this, the first being that in the current culture of academia this would be regarded as wildly unethical, at a level where even being tenured wouldn’t be sure protection. The second, and at some level more embarrassing problem is that generating bafflegab isn’t actually hard for me. In a mechanical sense, anyway— I find writing in the necessary mode fairly distasteful, and will put it off for weeks and weeks, but once I steel myself to doing it, it’s not challenging to produce whatever amount of text is required.
So, I keep finding myself in this state where I am at least in principle willing to give “AI” systems a try, but I can’t come up with a use case where I think they would be actually helpful for any of the things that I do. At least not one where I have enough confidence in what they might produce that I would be comfortable using it in a way that would make my work easier. If my work involved number-crunching on a much larger scale than I currently do, or relied more heavily on writing code to do straightforward tasks, maybe, but for what I actually do on a regular basis, I just don’t see much upside.
So, back to the bottom of the hill I go, until whenever the next release is that kicks up enough hype to get me thinking about trying it out again…
I’m not sure whether this counts as yelling at clouds or not, but it’s a string of words that have not previously been concatenated in this order. If you want more of this sort of thing, here’s a button:
And if you’d like to suggest some use case that I haven’t thought of, the comments will be open:
This is not a hypothetical, by the way; I need to do exactly this next week…
I could ask a machine to generate a script for a lecture on the subject, of course, but I’m not a teleprompter-reading robot.
I am still department chair, presumably because I did something really awful in a past life.
Other than coding (which it absolutely does not (yet?) do well enough you don't need to check it) AI seems useful for low-quality writing and low-quality research, and if you don't have a need for either, I don't think it's likely you'll see the benefits. I've found it quite useful for things theoretically inside my ability range but that I have next to no background info and don't feel like spending time searching for resources and then reading them, like "I would like to set up this old linux box as a media server for the household but it's too puny to run plex." It solved that in a couple of back-and-forths and then told me what to install and helped troubleshoot errors.
Since you read a fair bit of sci-fi, have you seen (or has the community been noticing and discussing) writers trying to use genAI to rapidly crank out somewhat formulaic but pleasing sci-fi stories?