The past few years have been an interesting study in the field of artificial intelligence. It’s a space I am heavily invested in, given the sheer volume of speculation we’ve seen since the early days.
I started writing software in 2013, though I only began producing meaningful work around 2015. Back then, much of the process was tedious, in a generative sense - most engineers considered the entire development cycle to be prescient. It was a craft you were privileged to inhabit.
When we encountered friction, we had platforms like StackOverflow. It contained almost every permutation of a solution you could imagine. From the most pedestrian - how to center a div - to debugging something as esoteric as an assembly-level program. But the interesting, although retrospective, part about StackOverflow is that it always possessed an answer. It meant that whatever bottleneck you faced, there was a corpus of solutions left behind by someone who had already bled for the answer. It was a conscious good; a community of altruistic fellas rooted in the idea of mutual auspice. If you had a problem, you went to the well.
Circle to 2023
Circle to 2023, the decline of Stack is humongous. Devs barely have to go on the site anymore; traffic has declined. In December 2025, the site saw questions drop by nearly 80% from its peaks. The library of human struggle has been ingested. We spent a decade helping each other for free, unaware that we were just labeling a dataset that would eventually make the community itself redundant.
Opensource
Opensource is not a new concept. For those not adept in the field, it’s the idea of making software and its source code transparent and accessible for others to build upon. Again, the software community has always taken pride in being rooted in benevolence.
We’ve seen a great deal of beautiful software emerge from this ethos. Linux, the kernel that serves as the invisible engine of the modern world; Android, which democratized mobile computing; React, which redefined our digital interfaces; and even VLC, a tool built by volunteers that remains the only thing capable of playing any file format ever conceived. We released the source code because we believed in a rising tide. We didn’t realize we were providing the architectural blueprint for a machine that could replicate the work without requiring the worker.
Determinism
Given these concepts, it’s almost delineating what I am driving at. With the sheer volume of data in the field; from the inchoate phase to the MVP - one can, with enough patience, depict exactly how to build a software product. Not to undermine the nuances, although I don’t consider them relevant to this premise.
The core being that software engineering as a field is deterministic. You can conceptualize an idea and already have the discrete steps to achieve it: setting up authentication, architecting database tables, handling state etc. The entirety of it has been so thoroughly figured out - thanks to the good heart of engineers past - that a pastiche is now possible. We have mapped the territory so well that the map has become the territory.
***
This sort of determinism is not unique to software. There was a time when finance believed its edge was instinct. Trading floors were theaters of personality - men reading rooms, sensing momentum, making split-second decisions based on what they called "feel." The mythology was simple: the trader was the edge.
In 2000, Goldman Sachs’ U.S. cash equities desk had 600 of these edges. At the time, they were masters of the machine - 600 traders supported by two computers. But the traders, acting with a supercilious confidence, fed their logic into the systems. They taught the engineers how they thought, convinced their “instinct” was too complex to be codified.
By 2017, the inversion was complete. The desk went from 600 traders to just two humans and 200 computer engineers. The “feel” was compressed into 100 computers. The power migrated from the individual’s gut to the institution’s system.
This is our current coordinate. I run an AI system myself - Avae, an autonomous chief of staff - so I am not outside this inversion; I am a participant in it. We are moving from ten engineers to one engineer plus 100 agents.
In this shift, we often look to voices like Paul Graham, who has famously argued that “taste” is the final differentiator. The idea is that in a world where everyone can make anything, the only thing that matters is the ability to choose what to make. It’s the belief that taste is an objective, cultivated judgment - something that can be refined but not easily automated because it requires a “soul” for design. For many engineers, this debate around taste is the last human fortress; we tell ourselves that while the machine can execute the code, only we have the discernment to know if the result is truly “good.”
But if you look at it through the lens of determinism, taste starts to look less like a mystical quality and more like high-dimensional pattern matching. If Graham is right that taste can be learned and improved through exposure to “good” things, then it has a structure. And if it has a structure, it can be mapped. What we call “good taste” might just be the statistical convergence of human preference; a pattern that AI is becoming exceptionally good at replicating. If the machine can predict which architectural choice humans will find elegant with 90% accuracy, the mystique of the architect dies. We are watching the same power migration that happened on the Goldman floor, only this time, the engineers are the ones being compressed.
The Question of Value
If production becomes trivial, where does value thus, settle? If everyone can manifest software because the “how” has been solved and automated by agents, who consumes it? Cloudflare recently found itself in a digital war against AI agents - a villainous obstruction of progress to some, but perhaps a rational attempt to preserve the economic leverage of human creators.
There is a slight contempt in the observation. We are building the Ouroboros, and we are doing it with fervor. We are so busy being proud of the technology that we’ve stopped noticing that we are the ones feeding it.
Which leaves the question of the middle ground. If the shift from 600 traders to 2 was the triumph of automation, what do we call the space we are in now? Is it a transition, or is it a terminal velocity?