Article posted on: 2026-10-05 20:47
Last edited on: 2026-10-05 20:47
Written by: Sylvain Gauthier
In line with the entire cringey reactionary tech enthusiast blogosphere, time for my 2 cents about AI and particularly agentic programming that no one asked or care about.
That’s right, I’ve started playing around with it and – after an initial phase of skepticism and just silly contrarianism, I decided to give Deepseek 4.1 flash a go.
So at least this isn’t the “muh AI bad” sort of blog post, so if you’re interested in a somewhat balanced take on this otherwise completely unhinged blog, carry on.
As if this wasn’t debasing myself enough, I’m also using their node.js
harness.
Furthermore, in my 9-to-5 SWE job at a large network company, I am now churning out heaps of generated C++ slop because that seems to be the only way to keep my job, and frankly it’s not like I enjoyed writing that code by hand anyway.
A note on the harness. Yes, dsh (which you don’t have to use) is a local
node.js application, so all the prompting happens on a browser page. At least
it’s honest about what it’s trying to do. Reminder that claudecode and
codex are essentially web pages rendered in the terminal, built by web devs.
I’d rather just use a web page, and the introspection features and dashboards
are actually nicer to look at in a web page.
I’ve been working on a semi-serious game project on my free time, a space-mining sim with a RPG element. All in C90, everything made by hand including the physic and rendering engine. I started working on it in 2024 and started using AI to aid its development in early September 2026, so the project was already relatively large, which matters.
There are all sorts of opinions on what AI can/should do in software projects. Some people write the spec and let AI implement it, with the extreme being vibe-coders. Some like to use AI to refine high level development then implement things themselves.
The workflow that ended up working best for me is a hybrid: use AI to help drive every step of the way, but steer it towards your vision from top to bottom.
At the end of the day, I generally have a very precise idea of what the code should look like, and AI helps me get there in about half to one third of the time, but not without constant back and forth.
I often find myself laying out the high value ground work by hand, leaving some gaps for tedious/mechanical work that the AI then completes.
My policy is that I only merge code I could have written myself, and more often than not it requires heavy refactoring/rewriting from what the AI produces. It only works well for me because the AI already had about 20k SLOC written by hand to extrapolate from.
I’ve heard from a lot of people that AI agents would 10x their productivity. There is only one way for this to be remotely true: they’re merging stuff without reading it. Maybe for front-end garbage that works, but I now know for a fact that for performance sensitive code like a game engine, this is a recipe for disaster. I could write an entire article with all the stupid mistakes my agent did that would completely ruin performance.
Talking with my colleagues and reading discussions online, I find that three clear-cut camps emerge among programmers. Those who legitimately think AI will replace them and are panicking or at best being stoic about it, those who think it has value and can augment engineers but still have fundamental flaws that make human oversight non-optional, and those who flat-out reject it.
Let’s breakdown those three positions.
I don’t want to sound like an arrogant asshole but I find that this often comes from profoundly uncreative people. They may be great programmers but only when it comes to applying known templates.
AI has no agency, no intent, it only predicts the next words from the previous ones. It’s very good at making its output look plausible. Sometimes, more often than not, the best way to be plausible is to be correct, but in those corner cases where it’s not the case, things fall apart.
I was recently talking with one such colleague and something struck me.
My argument was that human thought process is not purely language, I’m able to visualize a tree traversal, the shape of the intersection of two rotated cubes and so on. His answer, “I’m not able to visualize graphs/trees, so I narrate in my thoughts what the traversal would be like with language”.
So I can see why people may think this way, if their own thought process literally looks like an AI reasoning, they may struggle to find an edge over AI.
I fall in this category after being a contrarian.
The thing that always come up when people talk about their work with AI is that it lacks judgement. That sparks that make you go “this is a hot path and needs to run fast or we’ll get 25 FPS”.
At work I debug network switches a lot, complex beasts with lots of advanced components interacting with each other in weird and non-documented ways. And that’s something that AI unequivocally and utterly sucks at. It focuses on irrelevant details. It hallucinates. It gets tunnel-visioned into some scary log errors that have nothing to do with the problem. And it completely misses the small detail that gives out the actual cause of the bug. Because it does not have a mental representation of the box.
It’s not even funny, it’s flat out annoying, I’ve lost track of people handing me over a switch in failed state and a long sloppy AI report they generated: “hey I’m not lazy look I’ve already looked into it (with AI)”. And guess what.
The quadrillion line pompous reddit-tier prose is never right, not even close.
Now the code-generating part, massive time saver indeed, but only in so far as you actually read and understand what it does. It’s a very fancy auto complete.
I respect this position, I won’t even say they’ll be left behind. I don’t think the physical action of writing code ever was a bottleneck. It’s still entirely possible to go from point A to point B without a GPS, and in fact it’s better for your brain. The same is true with code.
Another formulation I really like, from Casey Muratori, is that at some point you’re good enough, or have such high standards for your code, that the fastest, most information-dense prompt to get what you want is literally the code itself.
In fact, that’s often the case for me, at some point when I don’t trust AI, I’ll just write the damn thing myself. There are times when it just feels like the right thing to do. If you ever felt like this working with humans, it will be the same with AI, “ugh never mind I’ll just do it myself”.
AI has created an interesting and very real paradox. In many ways, it feels exactly like mentoring an intern or a new grad. Being a mentor for an intern right now, I’m very well positioned to observe this, the dynamic is exactly the same, the output also, good at low levels, but lacks broader judgement.
So now, we are in a situation where it takes multiple years of experience to gain a solid edge against AI, at least in the eye of management/executives, which turns senior engineers into hot commodities and junior into unemployable NEETs.
Yet what makes the synergy work so well between senior engineers and AI is precisely the fact that they learnt to code without it, so they can judge its output with a critical eye. How can juniors get to that point now that they have to compete with AI, I don’t know. My company forbids the use of any AI tools to interns and new hires which I think is the way to go, but is it enough?
I don’t think software engineering as a discipling will go away anytime soon. We just got a cool auto complete that speeds things up. Non-technical people have a great way to prototype things quickly as long as they don’t mind their users' password ending up on a public spreadsheet.
My company is still hiring new grads and interns like there is no tomorrow. Things are moving faster and are getting out of hands in some aspects, but humans are in the loop more than ever.
Here is a quick case study of two industries where originality/creativity matters eminently: video games and high frequency trading.
In high frequency trading, SWE are a hot commodity (at least in Sydney). I know from first hand account that they have all models with unlimited tokens yet never merge non human-reviewed code. How do you get an edge worth tens of millions of dollars over your competitors by prompting a public API? You can’t, you need creative people coming up with out-of-the-box novel ideas. Not even talking about the fact that a bug in those software can crash markets, disregard regulations and cost the firms hundred of million of dollars.
Video games are also interesting: lots of shit-eaters on twitter posting their fancy Minecraft clone slopped out in two prompts (and $1500). Not one actually playable or remotely fun. In fact, the best outcomes (and that’s saying a lot) are remakes, so not exactly original by definition. I’m an avid Project Zomboid player: the mod scene has exploded, with some vibe-coded modes being half decent, but again, the entire game logic those mods rely on was carefully crafted and balanced by humans.
At the end of the day, the regression to the mean is an inescapable horizon of those technologies. In other words, the slopiness of AI generated anything will always be inversely proportional to the amount of underlying human steering/hand-holding.