Quality in the Age of Slop

My first programming job was as an assistant for a computer artist - it was a good opportunity to learn the craft of both art making and computer programming. And the analogy of coding agent as artists assistant is an interesting one.

It should be noted that many artists use artist assistants in many ways, some significantly beyond "painting the background". Jeff Koons and Damien Hirst are notable examples of high profile artists whose work is almost entirely made by assistants. It's easy to say these are bad artists who churn out big flashy conceptual works for millions on the art market with little artistic value. The analogy to the software world is easy to make.

But there are other artists whose work I think has more artistic merit despite being actually fabricated by their assistants (Jenny Holzer and Olafur Eliasson are two that come to mind). Sol LeWitt is an even more interesting case: his pieces are just their design, there is nothing physical about them. When they are put in a museum or gallery, it is up to the install team at the venue to actually make the work itself.

While it's only been in the last year the coding agents have come onto the scene, issues around authorship, forgery, labor, craft, mass production, etc have beset the art world for far longer - it might be worthwhile to look at it for guidance. I guess I'll have to add "Art in the Age of Mechanical Reproduction" to my reading list.

would you be interested in keeping an actual junior dev as an artist's assistant? asking for a friend

I honestly think SkyNet is a red herring. A superintelligent AI taking over humanity is likely to be better than the superimmoral ruling class that we have today. I think the real threat is them concentrating even more power through AI and ever steepening the social pyramids, which is far more likely and far more scary than superintelligent AI.

And good luck getting them to legislate against that...

This is what scares me most about the Maw. It seems to want to swallow forever the distinction between good and bad, leaving a world in which there is only code that works and code that doesn't, and no code that is beautiful, or excellent, or virtuous, or funny.

That nails it for me, yes.

Beautiful article. If we've gained nothing else from the AI apocalypse, it has led to a much deeper introspection over the relationship between software engineers and the code the write.

I am biased because this piece talks about the book that single-handedly changed my life. I like it a lot, overall, though I think prefacing it with opinions of pseuds from goodreads is not a good idea.

Gumption traps come in many varieties, but all will be familiar to working software engineers.

Gumption traps are very relevant to programming, and I think you are bound to encounter each one that Pirsig lists at some point in your career. (I've written about them myself, too, though this was before widespread adoption of LLMs.)

The advice in ZAMM is so relevant to programming that I wondered whether Pirsig, ZAMM's author, had ever done any programming himself.

Of course he did, and in the sequel to Z&AMM, Lila, he even namedrops COBOL!

This is hard to wrap your head around. I'm not sure I understand it. I refer you to ZAMM if you want a better explanation, since I probably haven't done it justice.

Quality, I think, is best described as a layer sitting above subjective and objective. For the most concise explanation I think we have to go to Lila:

Any person of any philosophic persuasion who sits on a hot stove will verify without any intellectual argument whatsoever that he is in an undeniably low-quality situation: that the value of his predicament is negative. This low quality is not just a vague, woolly-headed, crypto-religious, metaphysical abstraction. It is an experience. It is not a judgment about an experience. It is not a description of experience. The value itself is an experience. As such it is completely predictable. It is verifiable by anyone who cares to do so. It is reproducible. Of all experience it is the least ambiguous, least mistakable there is. Later the person may generate some oaths to describe this low value, but the value will always come first, the oaths second. Without the primary low valuation, the secondary oaths will not follow.

The reason for hammering on this so hard is that we have a culturally inherited blind spot here. Our culture teaches us to think it is the hot stove that directly causes the oaths. It teaches that the low values are a property of the person uttering the oaths.

Not so. The value is between the stove and the oaths. Between the subject and the object lies the value. This value is more immediate, more directly sensed than any “self” or any “object” to which it might be later assigned. It is more real than the stove. Whether the stove is the cause of the low quality or whether possibly something else is the cause is not yet absolutely certain. But that the quality is low is absolutely certain. It is the primary empirical reality from which such things as stoves and heat and oaths and self are later intellectually constructed.

(If you want, I also have some more notes about it.)

In Lila, Pirsig endeavours to construct a complete metaphysical system and therefore qualifies his definition into splitting static patterns of quailty (tiered into inorganic, organic, social and intellectual) above which sits this indefinable dynamic quality that is the focus of Z&AMM.

Pirsig might say that my opinion on AI is neither subjective nor objective because it derives from a perception of Quality. I'm not sure I can make heads or tails of that. I would instead say that my opinion is subjective—is hopelessly biased—but what ZAMM has helped me appreciate is that everybody else's opinion is too, even when it might not appear that way.

The question I would ask is whether adoption of AI is in itself a low-quality event or whether it is possible to integrate language models into the work of programmers in a way that is high quality. I think people are feeling that the way they engage with it is low quality, but because they don't have the words or the frame of mind to engage with it on a level that is not just subject-object dualism it's difficult to articulate, and they instead choose to reject it outright.

I think on some level AI enables romantic approach to programming in the sense that as long as you're able to engage with an AI-generated artifact on a surface-level and don't care to go deeper into it, it might be fine in a given moment. But when you actually try to look under the surface, into the code, you notice that there is no classical structure to it that you can reveal because it's been made by a model that pretended to work this way, but actually didn't. I suspect this is why people who have no inclination to engage with technology other than as a means to accomplish something else (executives, product designers, investors, solopreneurs, &c.) don't exactly understand the developers' frustration with AI-generated code.

He talks about how, in manuals for consumer products, every line conveys the idea that the grill/lawn mower/dish washer/computer has no relationship to you, and you have no relationship to it, other than to operate it. What counts as good grilling, or lawn mowing, or computing etc. is always taken for granted, even though that's the most important part.

This is still true of documentation and manuals for software libraries and tools. I read the documentation for Pi agent the other day and walked away frustrated because it just assumes that part - that you already know what is good way to use it and you're only looking for ways to tailor it to your preferences. And my colleagues, whom I asked "well, how do you use this thing well" just looked at me puzzled.

Reminds me a bit of Vim. If you just read the Vim manual, you come away puzzled. We needed some decades of people writing pieces on how to use Vim well. And then it turned out that Vim itself might not even be the best platform for good Vim usage, so we got Kakoune or Helix.

My wife and I are expecting our first child later this year. A daughter! I'm beyond excited.

As a father to a two-year-old daughter, congratulations! You're in for a ride of your life.

If you're looking for more material in the same vein as Z&AMM, I would recommend Hofstadter and Sander's Surfaces and Essences (making the same point as Pirsig's analytic knife but in a less philosophical or metaphorical way), Lila as a sequel that tries to structure the contents of Z&AMM more, and the work of Sevilla King and John Vervaeke.

the kind of book your freshman-year roommate (the one who wrote haikus at 2am by moonlight) would have gushed about

It's rude to put me on blast like this.

I read through the entire post. I don’t know if that’s a sign of your fine prose or just my ability to stick with a long read, but I think it’s the former.

Regarding quality, one point Robert makes is that people differ about Quality not because Quality is different, but because people are different in terms of experience.

So before I comment on quality to my team, I ask myself whether our experiences are the same. If not, telling them to “improve quality” doesn’t work. I need to be specific about what to improve.

Extending this to AI-generated code, I wonder if “quality” differs according to experience, too?

It's an immense relief to confirm that other folks paid attention to Pirsig. Previously, on Lobsters, I found that we were having almost exactly the argument that Phaedrus had picked with the classicists over whether essays have quality; the only difference is that the essays are written by chatbots rather than human students.

I can't see how I could offload programming to a coding agent without losing this sense for "the quality of the work." I have found LLMs extremely useful as a search tool and as a kind of super-charged rubber ducking partner. But to use an LLM to write code—when LLMs include randomness as an essential part of how they work—when the whole selling point of these tools is that they can produce more code than I can keep up with—would only be putting a layer of friction between myself and what I'm building.

I want to pick at this a little bit. I'm going to focus on Pirsig's framing: when a human subject looks at a physical object, the Quality of their interaction — that is, the source of the subject's value judgments about the object — emerges from the interaction itself rather than being subjectively determined by the human or objectively determined by the physical attributes of the object; we say that judgments are contextual or participatory. Not all objects participate in the same way; when a human looks at a photon, there is an inherent degree of freedom in how the photon responds, due to the Kochen-Conway theorem, whereas when a human looks at a tree, the tree is too busy maintaining its homeostasis to afford to change much in reply. In-between, M. pudica and D. muscipula do respond to touch and noise but not to gaze; it's not a one-dimensional spectrum.

So, how does an LLM harness or chatbot respond to observation? Well, it doesn't. Rather, it's a finite (and relatively small) mathematical object! All of its properties are objective and there's no choice or variation in its outputs. We can set up a pseudorandom harness in order to take random walks through its (un)likely tokens, and we can steer those random walks by force-feeding tokens of our choice, but that's about it. LLMs merely feel deep because they have hyperbolic topology and navigating hyperbolic spaces feels like zooming in with ever-increasing detail.

With Pirsig's reasoning, I can get to one of two perspectives on LLMs, but no further. Either: the LLM is a contextual system, where human inputs are the statistical context for statistically-likely replies. Or: the LLM is an objective system, where human inputs are the initial segment for statistically-likely utterances. Either way, the LLM is a sort of mirror which reflects its user at themselves, with the user only choosing the angle of approach; the user's chosen inputs are the primary means of cybernetic control by which the user attempts to reach their desired information/state and the model merely provides an array of preset options which is so large that it overwhelms humans. Perhaps the reason that chatbots provoke the ELIZA effect, leading readily to psychoses, is simply that it is a high-fidelity mirror, with flexibility and nuance, designed to distort the user's appearance with sycophancy and love-bombing in order to get them addicted to chatbot usage.

Discordians get five paragraphs! I want to connect this to your "layer of friction". You perceive using the LLM to write code as a barrier which sits between you and the computer. Vibecoders acknowledge this but insist that, like any other API, the barrier allows for abstraction and isolation, highly desirable software-engineering concepts. I perceive using the LLM as like a mirror, so I suppose that the mirror is not between me and the computer, but off to the side of me and the computer. Instead of using the computer directly, I aim at the mirror and carefully zoom in on a precise region; once I'm precise enough, my mere ability to see the computer from a specific angle is sufficient to instruct it. But that's not an abstraction at all, and it's clearly less isolated; I'm just putting in extra work in order to find a vantage point which might not exist. Why might a vibecoder do that? Perhaps they don't know how to observe the computer; HCI is participatory and a human has to have a programming context (a Naur theory, discussed previously, on Lobsters) before they can write code. Perhaps they prefer to look in the mirror because it reflects themselves back at them; it's mere vanity. But I think that those are really the only two ways in which it could ever make sense; there's enough problems between the keyboard and chair already and it doesn't make sense to add one that doesn't improve one's expressive/abstractive power.