An experienced engineer who holds two degrees in hardware and systems has grown tired of hearing that AI has solved coding. His case, titled “Coding Is Not Solved,” has drawn notice online, and it warrants a second look.
The argument rests on a straightforward premise: artificial intelligence can produce serviceable code for modest undertakings, yet when dependability, safety, or responsibility come into play, it fails. The writer, who constructed his own harness powered by an LLM and has instructed the subject for years, does not oppose AI. His opposition is directed at the promotion surrounding it. He wishes to dispute the “brain-dead narrative” that asserts coding has been resolved and that engineering today amounts merely to a matter of preference.
The Author’s Case Against the Narrative
This writer has two engineering degrees and has spent the past four years actually working with AI and constructing AI systems. He calls himself an early adopter of LLM-powered coding tools rather than someone who doubts them. That distinction is important to keep in mind. He is making his case from hands-on experience, not from a position removed from it.
Three product types are listed that do not strictly need the code to be read.
- Personal software: scratching an itch, automation, DIY patches
- POC (proof of concept): demonstrating technical feasibility and product viability
- Weaponized AI: acknowledging the risk and deliberately pointing it at a target to cause harm
What ties these three together is a willingness to accept danger. The first pair can fall apart without causing real harm. The third instead takes the natural danger and makes it part of the plan.
The Accountability Problem
The core grievance here concerns responsibility: AI cannot be made to answer for its actions in the same manner as a person can. It cannot feel the weight of punishment, it cannot die, it cannot be subjected to consequences. The most severe action available against an AI system is merely to shut it down.
“If you’re toying around, LLMs do a great job,” he writes. But that is the problem. The people pushing the “coding is solved” narrative tend to be the ones who have nothing to show for it. Anthropic accidentally leaked Claude Code, which turned out to have many flaws, and their status page shows orange is the new green.
Why Coding Is Not Solved
The writer contends that coding remains one of the final domains where LLMs have yet to fully take over. The explanation given is that humans have constructed a feedback loop that feeds syntax and runtime errors back into the model, repeating the process until most errors are resolved or concealed. This approach works well for straightforward tasks.
The very system that falters over counting the Rs in “Raspberry” or proposing a trip to the car wash can still expose logical fallacies. These engines operate stochastically and probabilistically; the only route to logic is to encase them in traditional code, subject them to testing, and employ techniques such as chain-of-thought. The fundamental problem persists: these engines struggle with logic and scale. As input grows larger and the context window fills up, accuracy diminishes.
Who Claims LLM-Generated Software Is Good Enough
Five traits characterize those who assert LLM-generated software is sufficient, according to the author’s identification.
- Haven’t written code in ages
- Cannot spot if their code figuratively had six fingers
- Have a low bar for what good looks like
- Don’t care about quality or NFR
- Have difficulty understanding an S-curve
The writer includes a sixth characteristic: honesty. The machine produces superior code. Yet the claim that it performs just as well across an entire field is, in his view, absurd reasoning.
The Industries That Can’t Afford Failure
The author lists industries with low risk tolerance:
| Industry | Risk Tolerance |
|---|---|
| Healthcare | Low |
| Finance | Low |
| Automotive | Low |
| Defense | Low |
| Power plants | Low |
| Aviation | Low |
| Manufacturing | Low |
When an error carries a price tag measured in dollars, human lives, or legal penalties, there must be someone to answer for it. That responsibility is something AI simply cannot offer.
The Dunning-Kruger Effect
A bit of Dunning-Kruger is observed by the author, with those who fail to read the output feeling more sure of it than those who do. This difference in certainty becomes hazardous when the stakes are high.
What the Author Wants
The author is not trying to change anyone’s workflow or toolbox. He couldn’t care less about that. What he cares about is that the services he pays for — looking at you, Google and GitHub — are degrading with stupid bugs that could be avoided if reliability and accountability were prioritized over velocity.
The guidance for managers is plain: quit pressuring your otherwise smart developers to push AI onto every available surface and workflow. The technology carries real strength, and it stands as the largest shift in the industry in ages. Yet too much reliance on AI exists, and when it damages the customer, you bear the blame.
The Limits of AI
The author does not claim that AI offers no assistance whatsoever. The case being made is that the promotion surrounding it has grown beyond what the technology can actually deliver. LLMs serve as useful tools, and their abilities are growing along an S-curve. However, there comes a point where more costly models do not produce results that match the pace of the added expense.
This point rests on the author’s own body of work. He constructed his own LLM harness, instructed the subject, and created LLM-powered products through it. He speaks from direct experience rather than as a mere observer.
The Verdict
What makes the author’s case convincing is its specificity. Rather than taking on AI as a whole, he targets the particular assertion that AI has resolved coding, and supports his position with concrete instances of where LLMs fall short.
The Dunning-Kruger observation cuts especially close to the mark: those who do not read the output are more confident in it, and that is a pattern worth keeping an eye on.
Of all the problems tied to AI, holding it to account stands out as the hardest to resolve. The system has no physical form capable of suffering a prison sentence or bearing fines, which means there is nothing an authority can punish it with. As a result, AI can never face the consequences for its actions.
What the writer closes with is a plea for reason. The endless recycling of overblown claims from token vendors about what AI can do must stop. It is we, the users, who bear the cost.
Online attention has turned toward the author’s work, and disagreement over it remains unresolved. Yet the author deserves credit for pushing the entire field to acknowledge where the technology falls short.
Source material: “Coding Is Not Solved,” alexewerlof.com.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.