This is actually a post I wrote in May 2025. I’ve been sitting on it because…well, honestly, mostly laziness, but also a bit that it felt unfinished for how to-the-point it is. However, commentary on the role that software engineers’ pride seems to be playing in AI company marketing (see, e.g., Cal Newport), among other buckets of cold water dumped on its use in software engineering since last summer, has pushed me to just ship it as-is. So fuck it, here it is.
————————————
Tools are expected to have a defined and articulable function, and they are assessed according to that specification. This is how, for example, we know what to write about when we review things.
Mere widgets, on the other hand, are not expected to have a defined function–they are simply a placeholder for the economic output of a company. By referring to “mere” widgets, I mean to focus attention on those economic outputs lacking a plausible, accompanying explanation of economic viability. For example, who is the target audience? What problem(s) will it solve? How is it better than the alternative? What are the alternatives? The production of mere widgets has a long and storied history in technology, especially post-WWII in the U.S. And we see, historically, that reviews of them–google glasses, the meta verse, etc.–are often difficult to write and fail to predict the way they will be assessed. Rather, these reviews focus on the salient (read, flashy) features and the “promise” of their kind.
AI–in the form of LLMs–is very clearly a mere widget. While coding has now been latched onto as the function, and they are spending hundreds of millions on extremely exploitative labor practices to make it appear their capacity for coding is “innate” or “emergent”, this is post hoc, a poor motivation for the DNN structure of LLMs, and does not actually guide our assessment.
People defending AI as a tool, however, do have a defined and articulable function. They are, at bottom, the launderers of hype: they transmogrify what is clearly bullshit PR into plausible-sounding claims. For example, “AI will replace coders”–clearly bullshit PR covering for overhiring during the pandemic and the tax law changes that meant they could longer immediately write off R&D expenses–has now become “AI is an essential tool for coders,” which sounds much more plausible. Indeed, most ask themselves, why shouldn’t copilot be compared to SDKs, high-level languages, or macros?
There is, of course, a good answer to this: while each of these grew out of widgets and basic research, their practical implementation came after identifying and tailoring each to a function specified in advance.
Fun fact about programming languages: they have known vocabularies, their parsing grammars are (comparatively) simple, and for most (i.e., mundane) specifications, we could identify, just by brute force and knowing a little about the problem structure, a program that implements it.
Fun fact about programming: it’s mostly the same basic specifications, the same few languages, the same sorts of divide-and-conquer moves, and requires the same sorts of comments. (Think: if it weren’t, why would we be able to get away with students just memorizing a few basic algorithms in our algorithms courses and never learning how to document their code?)
Fun facts about DNNs: they can’t learn even regular languages, and they can’t do composition reliably (i.e., don’t generalize) and, when they can, it’s because they have high data density with that very specification.
Given this, even if we do view an LLM as a tool, it sure looks like a shitty one. We were doing just fine with our pencil, hammer, and saws and were given a Swiss army knife to replace them. Compared to the tools we could easily have built instead–your table saw, bench planer, router, etc.–and even compared to what we had already, it is manifestly worse. So tell me again: who benefits from viewing AI as a tool for coding? It sure isn’t our students, and it sure isn’t the public!
And, this function of “an LLM is a tool” progenitors was not glommed onto post hoc. These companies knew, with near certainty, that this would happen. This is, after all, the entire unicorn strategy: prop up tons of garbage startups just long enough for one to make it to IPO so investors can cash out. We have also seen it in Ed Tech, not to mention education, for generations: as long as you measure just the wrong things (e.g., student performance on rote problems) at just the wrong time (e.g., during the time students are actively engaging with the material, and very importantly, not after they’ve supposedly “finished” learning it), it’s quite easy to make these widgets look like useful tools (e.g., the calculator). And these non-experts on learning from CS and other engineering/management disciplines have been more than happy to ensure the shiny widgets get the breathing room the companies are looking for.
But this is also simply baked into the social structures at play here. After all, benefitting themselves is why tech companies have gone out of their way to capture the curriculum and research interests of first terminal degree programs and now undergraduate curricula. Here tech companies have ripped a page directly out of the playbooks of tobacco and the fossil fuel industries. Between most CS PhDs going into industry, an incredibly low emphasis on the humanities and social sciences in undergraduate curricula, and selection against students who have cares beyond their own income, CS education is a network primed and ready to transmogrify industry bullshit into superficially plausible claims.
They also know that these people’s libertarian sensibilities do not extend beyond the usual talking points, so that it will occur to very few of them that the “get onboard or gtfoh” attitude is a constraint on free choice. Indeed, as effective tools for these companies’ PR, these individuals serve to concentrate the power to constrain ways of living, with little to no public recourse, in just the way that libertarians have always asserted was the whole point of enshrining individual liberty.
Even worse, many will disclaim their pronouncements that “AI is necessary” by denouncing its attendant moral harms. In effect, however, this is worse: what it tells listeners is that these people are not ignorant of the harms but, rather, they have explicitly prioritized the short-term promise of economic gain over their putative moral values and the long-term experiences of their students (and, indeed, the world). That is, while the harms matter, they matter *less* than the rewards we may see in the near future. Of course, adding in a zero-sum conception of their choice–I need to do it for my students, else someone else’s students will reap the rewards–serves to overdetermine their judgment.
Thus, while an LLM seems to measure up poorly against the rubric of toolness, the progenitors of this view measure up exceptionally well.
- LLMs aren’t a tool, but you might be - July 21, 2026
- Quick note on OpenAI and Erdős’s unit distance problem - May 26, 2026
- Who cares about Haag’s theorem? - April 19, 2024