Are we augmented by AI yet?

An update about AI-assisted coding was long overdue. Here is my fresh take about it (until it changes again in 6 months). Actually, I started writing this in January, but couldn’t get it right, came back to it multiple times. Things are going so fast these days.

I’ve been doing software engineering for the past 15 years, and code has been part of my life for 25. I’ve seen enough shifts, hype cycles, abstractions, and “this changes everything” moments to be cautious when a new one arrives.

In July 2024, I wrote Brandolini’s Law in the GPT Era, and in August 2025, I wrote Are we replaced by AI yet?.

Reading these back, I feel how much things have changed since then, and at the same time, how what I wrote then still applies now, which is surprising to me given how fast things are moving.

Like me, you’ve probably been hearing/reading so much bullshit your ears/eyes are bleeding at times. Linkedin shitposts informing you that “PMs are in charge of development now” (god please no - love you guys but please no), or that agent swarms are inevitable (or whatever variation of it comes out every two weeks). Public figures with bold predictions. I’ve heard from a publicly traded company internally forcing opus 4.6 down the throat of SWEs: something along the lines of “if you’re not using $10k worth of token a month, you’re not trying hard enough. Manual coding will be over in 2 quarters.” What the actual fuck is happening? How much disconnected from reality are you when doing this? Or am I the one disconnected from reality?

And yet, I’ve been practicing it. Everyone noticed it, there’s been a major shift last year with opus 4.5 and gpt-5.2-codex. Everyone’s getting high on cheap tokens these days. This is a rat race on steroids.

When chatgpt became kind of usable, I felt ashamed asking a bot for help. I wouldn’t tell, and people around me would do the same. Now the narrative has dramatically changed: you are expected to do it. And that’s ok I guess, because it allows for faster delivery, and automated review is now due diligence before asking for review. You are now expected to master every aspect of the alien technology.

My latest work is a large internal set of features we had postponed for years. It was too much investment for internal tooling, so it never became a priority. With modern coding models, that changed. The result is roughly 80k lines of Go, plus a bit of TypeScript, much of it produced with AI assistance. It delivers real value for relatively little time investment, and without that tradeoff, we probably would not have built it at all.

But the tradeoff is real. I understand the system conceptually, yet I do not hold every implementation detail in my head the way I would have before. We are accumulating what I previously called semantic debt, and what others call cognitive debt. For now, I accept that trade.

We now have a full AI loop at work. Opencode runs in a scheduled CI job with various commands. It creates issues, plans, implements PRs, reviews them with another model. Are the issues good? Frankly, yes. It found interesting bugs in “my” main project. This is embarrassing, but boy I’m happy these are fixed. I may have read these files dozens of times, reasoned about it, and yet, with a few dollars worth of inference and twenty minutes, it found out what me and my reviewers did not catch for years. I am weak, fallible. There are so many things I can focus on, and AI buys me so much time.

So are we augmented by AI yet? I think we are, there’s no way denying it. Even the most skeptical people I know are using it now, even in highly demanding niches, such as C++ high frequency trading. Business-wise, this is a no-brainer, on the surface at least. By that I mean it is today, short-term, but there are consequences we might not even comprehend. Software engineering is such a complex, wide topic, and we are eroding what has taken decades to build up: the craft, the career ladder for juniors, and ultimately the passion for writing code.

We are creative people at heart, and this is a big hit. The question is no longer if AI can build things, it can to some extent, it is about what remains. I feel a growing split personality. On one hand I feel so productive using AI to spit out large quantities of code fast, with a good level of quality, and on the other hand, this goes against my vision of what software engineering should look like, or even, feel like.

I learned writing code without auto-completion. Hell, I learned to write machine code, this was one of my first encounters with this field. I learned to write code on paper, I swear, literally using a pen and paper: this was how we were evaluated during my studies. This feels antiquated and yet, this was a mental challenge I enjoyed. Writing quicksort with just my brain, me, a pen and a sheet of paper was so intense.

Actually, using AI to write code goes against a lot of what I learned the hard way. Planning, specs, and architecture were always a massive part of the job, of course. But implementation was where I used to spend the most time, and that time mattered. That was where edge cases surfaced, where abstractions met reality, where I internalized how the system really worked. Do you remember that time? It’s been months since I wrote anything complex by myself. I feel sad this time is over now. Maybe a lot of it is nostalgia though - I’m not sure yet.

A lot of the best writing on AI coding right now focuses on discipline: how to stay in the loop, how to review well, how not to drown in slop. I agree with all of that. But my discomfort is not only about quality.

I do not think the old way was pure or better in every respect. A lot of it was slower, more tedious, and not necessarily more correct. The tools really do unlock things that would otherwise not get built. But previous abstractions mostly removed mechanical effort. AI removes part of implementation itself, and implementation was where I used to discover the system, build a mental model, and earn my understanding. That is why this feels categorically different.

We traded this away willingly because the productivity gains were too obvious to ignore. And now I’m not sure I like the deal. This is not quite the job I signed up for. Building things is still satisfying, but the way we build them used to be part of the craft, of the pleasure, of how we learned. AI-assisted coding changes the relationship between the engineer and the code, and maybe, in time, the meaning of the job itself.