Episode 58, zero lines of code produced. The developer used Claude Code, Codex (OpenAI's code-generation model), DeepSeek, Doubao, and other tools to generate the admin-side configuration management module, but essentially discarded the output after review and spent a full day rewriting it by hand. The reason is not complicated: the system configuration supports dynamic forms (form fields added by users at runtime), involving 25 input components, and Go is a strongly-typed language (variable types must match strictly, without PHP's automatic conversion) — if any single link in the type-conversion chain breaks, the whole thing breaks.
What This Is
This is entry 58 of a personal blog series. The author, originally a PHP full-stack engineer (a developer who writes both frontend and backend code), aims to use AI assistance to learn Golang from scratch and build an open-source admin project called ai-go-admin. The task for this episode: implement the system configuration management page on the admin side, referencing an existing PHP version.
It sounds like AI's sweet spot: a clear reference, a clear target page, and an existing data model. But the actual result was that 100% of the AI-generated code was discarded. This "AI generates + humans scrap" loop, surfacing even on a task with a definitive reference implementation, is worth pausing on.
Industry View
One camp argues this illustrates exactly where AI coding stands today: it can run through a flow, but it can't bear the weight of complex business rules. Dynamic forms, type conversion, edge-case handling — these are precisely the most time-consuming parts of any engineering project, and AI currently performs unstably on tasks that require understanding contextual constraints.
The counterargument exists too: the developer's own familiarity with Go may be the real efficiency bottleneck. If the author were a seasoned Go developer, the AI-generates-plus-locally-tweak pattern might well have worked. In other words, this failure may not be AI's problem — it's the cost of the usage pattern "using AI to learn a new language". When someone uses AI to help them write an IELTS essay in English and gets it sent back for rewrite, that doesn't prove AI writing tools are useless.
A cooler judgment: across 58 episodes, AI remains effective at least at boilerplate code (highly repetitive template code) and function-level implementation — it just cannot replace deep thinking about the business model. Treating it as an "accelerator" rather than a "replacement" sets more reasonable expectations.
Impact on Regular People
For enterprise IT: If your team is evaluating AI coding tools, take this author's real test as reference: even on modules with a clear reference and well-defined scope, AI can still produce code that needs major rework. Discount your deployment expectations.
For individual careers: Non-developers shouldn't panic — the gap between AI coding progress and your replacement is not as close as you think; but developer peers should note: AI won't put you out of a job, but people who use AI may cut down your overtime.
For the consumer market: The "10x efficiency boost" and "a line of requirements generates a whole project" marketing claims from various AI coding assistants still read as sales talk based on current evidence. Consumers pay based on real scenarios, and enterprises purchase based on real ROI (return on investment) — that's closer to the truth.