Reflections on Quality ZMM #2
I’ve been struggling quite a lot lately with this topic of Quality. It’s been a broader concern even before I was reading ZMM, but the reflections in my prior post about it really had me confront the reality. For frankly my current education choices are at odds with the core of this philosophy.
As I said in the prior post, I’ve started learning how to code after the advent of genAI, causing me to feel quite dubious whenever I’m doing actual analyses. On the one hand, learning it gives me the ability to understand what I’m doing better and become one with the activity of analyzing or building something. Which is what Pirsig prescribes as a necessary condition for quality to appear. On the other hand, though, the most experienced developers, such as Peter Steinberger but also those who are giving me guest lectures in my study program, admit themselves that the idea of sitting inside the code is becoming a thing of the past. You say what you want to do to an agent, such as Codex or Claude Code, and then review at every step.
When I was setting up a project a couple of weeks ago, and had an idea I wanted to implement, I would explain to codex what I wanted and tell it to look at the data in my folder in order to give me a plan. Then I would give the same explanation (broad level overview of goals) to ChatGPT together with a couple of snippets of what my data looks like. After this, I had codex generate a plan, which I then paste into GPT, which automatically suggests improvements to the plan. I’m basically letting two agents hash it out in order to build the perfect pipeline.
The result is a bunch of scripts of which the outcome is exactly what I wanted, and more. I mean this code is so clean I wouldn’t have even known where to start on this. Moreover, a full preprocessing pipeline of the data took me a couple of days. Probably less than a day if I didn’t have other stuff to do.
So where do you go from here? In order to actually build stuff, it’s most efficient to have agents help each other in order to build the thing I want. But there is no skin in the game, and I don’t particularly feel any connection with what I’m creating. I know the idea of each file in my folder, but couldn’t tell you exactly what’s in each. I suppose this is also the reason I don’t get additional ideas about the project while doing this. Normally, when you’re truly engaged with something and you become part of the activity, you care, and as a result of tangling with the data and topic, you naturally build on top of it because your brain has a lot of faculties coming together. That’s not the case here.
Same thing goes for this blog. I recently transferred the Jekyll GitHub pages blog to Astro + Vercel, and pretty much struggled my way through it and had GPT and Codex do the layout and email setup fixing. It took 5 hours or so to have it be the exact blog I had before, and it looks and feels great. But what now? How do I know if the path I chose with this is the one that led to the best outcome?
Both for the project and the blog, I lack counterfactuals. If I had taken two weeks to do the blog transfer and really combed through each topic to become familiar with web development, I would’ve possibly had tons of ideas of what to do. Moreover, I would’ve understood it, and in the future be able to adapt and add things without again relying on an outside system. The biggest thing of all, though, is that I would get to partake in the joys of compounding. Same with the project, I’m not finished and will go on until I have an MVP, but I have no idea whether the alternative route would’ve given me better results.
You could argue that I now must focus on figuring out the laws of demand and marketing. That is true, and I likewise referred to this in the other reflections. If I don’t have to focus on learning how to operate a computer, but instead just talk to it in human language and it conforms it’s manufacturing system automatically to what I want, I get to focus on what really matters. What has always mattered. Laws of supply and demand.
So we really have two opposing viewpoints, each with merit:
1: For quality to emerge in what I’m producing, I must be one with the activity, and build myself along with it.
2: As my about section’s description says, but Nassim Taleb also refers to with skin in the game: “Primum vivere, deinde philosophari” - first live, then philosophize. I cannot get 10 years of experience in programming and web development while the ability to act on my thoughts and ideas has increased exponentially.
So the problem is quite clear. I can and must, to survive, act in the most efficient and effective manner available to me. But the very mode of doing so takes away from the probability of the outcome being Qualitative. There are no counterfactuals in observational data, so how to move forward on this? An experiment?