AI coding tools are everywhere, promising a revolution in software engineering. Some say they'll replace developers, or make us 10x more productive.
The truth?
AI is a fantastic prototyping tool, an accelerator of work, but it doesn't engineer.
(caption) On the left: headlines, capital, promises — but fragile infrastructure. On the right: fewer spotlights, more hard hats — engineers crafting a solid bridge, humans orchestrating machines, building systems that hold.
That’s the difference between hype and durable value.
AI excellence
AI excels at prototyping and exploration. Need to spin up a new design pattern? A Proof of Concept (POC) or throw-away toy prototype? AI is definitely your best friend. It'll generate scaffolding, draft APIs, and wire up a frontend+backend in hours or less. Perfect for commitment free experimentation. Example: I built a Sublime Assistant AI plugin with context functionality in less than an hour.
Writing a one-off script or a tool for your own use? Do it. Nobody is maintaining or scaling this, and generative AI's speed and flexibiltiy make it disposable code where enterprise concerns don't matter.
Low-intensity work where you're basically copy-pasting templates together? AI will automate it, allowing you to focus on where you add the most value.
AI failing
The trade-offs are the most important. As Linus Torvalds put it (unscripted conversation in a LTT video, around the 36:30-37:30 mark):
"Anybody who thinks that's a valid metric (lines of code, red.) is too stupid to work at a tech company. "
It was never about pure code output. Real engineering is about trading maintainability, scalability, flexibility, and future proofing. Decisions that make code last. AI just generates lines of codes - the least useful metric. It'll rewrite your entire system - with 80% accuracy - but it won't ask whether you should, and what trade-offs you are willing to make.
In my boardgame experiment I banned myself from coding to test generative AI's limits. It made a mistake in core logic (hexagonal adjacency) and could not self-correct - I had to have it visualize the logic, then tell it how to correct the mistake. I had to tell it what was wrong. Studies like the one from METR (source) found that, despite dev's confidence in a 20% speed up, using generative AI tools actually slowed them down by 19%.
Despite $30–40 billion in enterprise investment into GenAI, this report uncovers a surprising result in that 95% of organizations are getting zero return. (Source: MIT)
One explanation is a common occurence when debugging with AI. In a reasonably large (10-40%, anecdotally) amount of instances, it will confidently tell you where the problem lies. And when you tell it that didn't work, it will dig deeper. It keeps digging a hole. It never admits that it was wrong - and that the answer is probably on a forum somewhere already.
AI: Intelligent? Ask an influencer
AI proponents have been extremely optimistic, promising world-shattering revolutions in the way we work. Jobs phasing out in less than a year. Profits will be skyrocketing. As technology review put it:
"AI companies have presented each new product drop as a major breakthrough, reinforcing a widespread faith that this technology would just keep getting better."
And to be fair, they are getting better. Smarter, more efficient, better with handling context, tools, tasks, skills, RAG. For this optimistic view on AI and general review of news, I would suggest Nate B Jones.
AI definitely gets better at scoring well on benchmarks. In a well-known anecdote that I heard from a bloke in a pub - much like an LLM gets some information from a random reddit post: Early AIs were benchmarked for solving chess puzzles. As a result, they were fairly decent at chess. However, later models weren't trained on those benchmarks anymore and as a result they lost the ability to play chess almost overnight when the release cycle happened.
These considerations indicate that benchmark performance, even across many tasks, is not sufficient evidence of general intelligence. General intelligence would presumably entail strong benchmark performance, but available real-world evidence suggests that the core properties of general intelligence – flexibility, generality, and reliability – remain elusive. (Source: Rumors of AGI ... greatly exaggerated)
For a more grounded and realistic assessment of generative AI, I suggest Gary Marcus.
AI Heuristics
What do I think about generative AI? It's a tool. It's as much as a tool as a good framework is to python, or python was to C++. You gain something (easier work) and you lose something (static typing, proper memory management). It's engineering, so it's about tradeoffs.
I think AI is absolutely good for:
- Exploring patterns and proof of concepts
- Prototypes, throw-away tools
- Boilerplate and repetitive tasks
- Replicating patterns
I also think there's some things it shouldn't be used for. AI shouldn't decide; keep a human in the loop, ideally one with domain expertise. Don't vibe code your way into production - review, think about maintainability, scalability, security, governance; be an engineer. Production should last.
The main message
For me, AI is not merely a technique for rapid prototyping, but a way to focus on high-impact tasks where we can add value. Focus on engineering and architecture more, and less on implementation. It was never about typing code.
AI does not lower the cost of poor decisions, only the cost of producing code. As execution becomes cheaper, the trick is proper judgement.
The strategic move is not to maximize (low-value) output, but to redirect freed capacity towards architecture, clarity, testability, security, and long-term coherence.
Used carelessly, AI will only create chaos. Used deliberately, and it creates value — shifting low-value production to high-leverage thinking.