AI tools can generate functional, adequate, perfectly average code at a speed and cost that would have been unimaginable even five years ago. And like the outsourcing wave of the early 2000s, the economics are real and rational. Nobody is wrong for using these tools. The code they produce is often fine. It works. It passes tests. It might ship as-is.

Not the first time I've read this kind of statement and I always struggle to reconcile this with my personal experience. I'm seriously doubting that I'm just not a "good enough prompter". I know how to explain context from domain to tech and vice versa, that's like, a good 20% of my job. I'd say that AI tools are good at producing code that already exists.

The LLMs are an interface to a corpus of written material. They've never had a thought, a chat around the coffee machine, or any experience in the largest sense of the world. This is a hard barrier on any induction they may emulate.

We need the equivalent investment now. If average code is cheap, then the scarce resource is no longer the ability to produce it. The scarce resource is the ability to read it, to navigate it

You know what would help a lot with understanding the code one is working on? Writing it yourself without turning your brain off via AI.

But that's an insight the article somehow seems to be missing.

Meh, disagree with a lot of this.

AI tools can generate functional, adequate, perfectly average code

Not in my experience.

The outsourcing era taught us that the expensive part of software was never writing it. It was understanding it well enough to change it safely, to debug it under pressure, to explain to the next person why a particular decision was made at 2 a.m. on a Tuesday.

Since AI is adequate, just have AI change, debug, and explain it. You don't even need devs running the AI. Have AI generate intent. Just have AI scrape Twitter for people complaining about applications they wish existed, and have the AI make them. Let AI do market research. It's supposedly perfectly adequate.

As if needing to be able to understand code quickly wasn't already a problem before LLMs.

I was having this exact conversation with one of the archetects in my company - top down orders were to start integrating llm's into our engineering workflows - so be it, they sign my paychecks and selfishly i want those to keep flowing - but beyond the quality and slop concerns i had already raised and were disregarded, i was realizing that i was losing my intuition about the code i was releasing under my name - it takes me longer to answer questions and i cant just wing my answers based on the intuition built during development because i didnt build intuition during development when development was outsourced to the llm - thats a liability im trying to highlight to managment and their risk analysis and also to myself for risk reduction

Intuition isnt made in documentation (nor the slopdocs llms makes that no sane person will ever actually read) - its built by the effort of comprehension and theyll be no shortcut to that

But we've seen this pattern before. When code production gets cheap, the cost doesn't disappear. It migrates. It moves from creation to comprehension.

"AI" coding tools can offer some value. The problem is that they often generate tech debt with pay day loan level interest rates. What's made the interest rate so high is that now, not only do you have the actual code base tech debt. You also have a bunch of code that no one understands and the barrier to entry for software engineering has become so high that fewer younger people are actually learning how to be good programmers. Lots of organizations don't give a shit about their rapidly growing mountain of tech debt today but they're sure going to at the end of the week when the payment comes due.

What their "leadership" fails to understand is that any idiot can shell out code. I've seen lots of terrible programmers generate millions of lines of really shitty code that somehow, by the power of the dark Lord himself, manages to compile. That's not what a software engineer actually does though. Software engineers design operational systems with software. Writing code is a secondary function of that. There are currently no AI agents that can successfully design a software system with any degree of complexity because LLM's don't actually understand anything.

The profits also aggregate at the top. The problem is it’s short term, and the later cost isn’t funded by the top as it’s a one-way flow.

Record profits, new product release: they’re often followed immediately by layoffs so the next quarter’s metrics look extra good and they can cash in.

The few poor souls remaining hold the fort until hiring starts again and the cash out cycle begins once more.

Problem is the code is crap. Code without intent is crap. Code IS intent. LLMs are just statistical models with no understanding.

They are fancy copy paste machines. They have some minor "intelligence" but it's pretty much just the ability to associate like information. Their responses are nothing but compositions of there training data and local context.

It has some uses granted. But writing your code beyond auto complete or extremely simple and straight forward tasks isn't one of them. LLM generated code is not production ready code.