1. Foreword Discussion: Why AI Coding Makes People Panic

Recently, I saw a question in a tech group that perfectly captures the technical anxiety most people feel in the AI era.

The question came from a product manager who had deeply adopted AI development early on. He could skillfully use Vibe Coding to create product prototypes, build everything with Next.js, and deploy automatically via Vercel. He could also independently implement login systems and connect to Supabase for databases.

He said he knew the entire AI Coding workflow, but hit a bottleneck in practical implementation. All his projects remained at the self-entertainment demo stage and had never landed a complex Agent product. Whenever he had a complete, complex product idea, he ultimately had to seek help from professional programmers to realize it.

He admitted he was panicked. Despite doing a lot of hands-on practice with AI and seemingly mastering AI coding skills, he could never break out of the demo level and could not independently deliver a commercial product.

His core dilemma is clear: he doesn't want to keep relying on programmers and wants to independently land products, but he knows his AI coding ability has obvious shortcomings and can't find the right direction for advancement. The discussion eventually produced two solutions:

The first is to completely transform into a developer, master the underlying technical know-how, write code and submit PRs personally, and complete the entire development and deployment independently. The obvious downside is that the development field encompasses massive knowledge systems like frontend, backend, operations, and architecture. The endless technical details will trap you completely, consuming enormous energy, and it easily breaks the development boundaries within a product R&D team, triggering collaboration conflicts. It's very hard for an ordinary person to persist long-term.

The second is not to dig deep into low-level coding or forcibly transform into a developer, but to position yourself as an evaluator, tester, and business decision-maker. Rely on AI's brute force, using ample tokens and AI computing power to achieve functional implementation, while independently maintaining a dedicated test suite, collecting industry experience, product logic, user feedback, and test issues to verify and optimize AI-generated output.

Given his role as a product manager, my advice to him was: prioritize the second path. There's no need to spend massive amounts of time gnawing on complex low-level development knowledge. Learn synchronously while AI solves problems, iterate quickly, build evaluation capabilities based on a test suite, and close the user loop. The cost-effectiveness and landing efficiency far exceed blindly diving deep into coding.

2. So, Is It Still Necessary to Learn Programming?

The core of this discussion actually corresponds to an ultimate question in today's tech circle: Cursor, Copilot, Devin and other AI coding tools have become standard, and AI can handle the vast majority of coding work—so do ordinary people and tech practitioners still need to painstakingly learn programming?

I've observed two extreme views on the market. On one side, newcomers think mastering Prompt techniques is enough to handle all development, and that traditional syntax, algorithms, and engineering logic don't need to be learned. On the other side, many senior practitioners cling to old-style development models, insist on writing code line by line, deliberately resist AI tools, and their development efficiency remains stagnant for years.

From the perspective of frontline AI engineering implementation, both views have pitfalls. AI has indeed overturned the traditional work model of manual coding and significantly lowered the barrier to producing code, but it has never dissolved the core engineering value of programming. The real change in the industry is the restructuring of programming's learning logic, R&D division of labor, and talent evaluation system.

AI will continue to replace execution-oriented personnel who only mechanically move bricks and mindlessly stack code, while continuously amplifying the core value of practitioners with abilities in problem decomposition, technical trade-offs, architecture design, and engineering delivery.

Everyone should ask themselves a question: why can many people use AI to write code but only make demos, never able to launch online and support long-term iteration? The essence is that most people confuse "code that runs" with "engineering that works," and misunderstand the true meaning of learning programming.

3. Most People Overestimate AI Coding Ability and Underestimate Engineering Complexity

I've interacted with many people from non-development backgrounds, especially those who attended two-day coding bootcamps, and they generally share the same problem: relying on AI to quickly make functional demos, they feel they've mastered complete development skills, but their projects forever stay at the demonstration stage and cannot be commercialized.

I've summarized the core problems of this group:

• Unable to discern whether the tech stack recommended by AI is reasonable; piling heavy architecture on small projects and using lightweight frameworks for large projects, causing resource waste or performance risks.
• Unable to accurately restore business and design requirements; only using vague terms like "better looking, smoother" to adjust AI output, without standardized acceptance and optimization logic.
• Lacking production-grade engineering thinking; not understanding permission control, data validation, logging, security risk control, and completely lacking the ability to deliver for launch.

Most of them just find it novel, feeling that things which previously required programmers to implement can now be done by themselves, and they think they are amazing.

Even engineers with basic development experience will have their shortcomings fully exposed in complex business scenarios. For simple CRUD and static page development, AI output quality is completely sufficient. But when encountering core business involving multi-table associations, cross-module linkage, and complex data flow, AI code generally has hidden problems: it runs, it works, but it's not standardized, hard to maintain, has hidden dangers, and doesn't support iteration.

More critically, many people can't see these problems. They can only passively reuse AI code, lacking the ability to troubleshoot vulnerabilities or optimize architecture. Some developers even use AI code directly without review, which ultimately leads to losing control over the project.

4. Real Implementation Case: AI Can Do Features, But Not Engineering

Let me give an example for comparison: a real enterprise implementation scenario, fully dissecting the core gap between AI autonomous development and human intervention, covering the entire process from requirement input, development, to verification.

Complete Scenario Information

• Scenario Role: Enterprise backend R&D engineer, responsible for developing the multi-dimensional table module of an internal collaboration platform.
• Core Goal: Build an enterprise-grade multi-dimensional table supporting multi-person collaborative editing, role-based permission differentiation, and message @ reminders, suitable for long-term business iteration and team-scale usage.
• Input: Standardized natural language requirements—replicate the core capabilities of Feishu Multi-dimensional Table, implementing multi-person online editing, differentiated role permissions, and message @ push reminders.
• Processing Flow: AI autonomously completes architecture design, technology selection, and code writing throughout the entire process, with no human architectural intervention, no business rule constraints, no engineering standard verification, only requiring the functions to run normally.
• Output and Verification: AI quickly completed full coding; basic functions like page opening, editing, and viewing worked normally, but deep engineering problems erupted comprehensively: missing data operation audit trails, coarse permission granularity unable to be tiered, incomplete message push chains, and data overwrite risks during multi-user concurrent editing, completely failing to meet production delivery standards.

From this case, the core difference can be clearly seen: AI's core capability is quickly stacking surface-level features, but it completely lacks commercial engineering implementation thinking.

Tools only mechanically match current requirements to complete development; they won't proactively consider long-term elements like business iteration, team maintenance, data security, and system stability. And these are precisely the core capabilities that programming learning truly needs to master.

These engineering aspects are the deep water; otherwise, why would Feishu need a team of thousands of people?

5. Comparison of Two AI Coding Modes

To clarify the value of learning programming, one must first understand the two mainstream agent coding modes currently available. The human-machine collaboration logic, implementation cost, complexity, and capability requirements differ completely between modes, directly determining an individual's irreplaceability.

Here is a comparison I made (you can swipe left and right)

Mode Type	Design Idea	Advantages	Disadvantages	Applicable Scenarios
Collaborative Agent	Human-machine collaboration, AI-assisted coding; humans are responsible for control, error correction, and standardization.	Replaces repetitive coding, improves human efficiency; fast iteration, flexible scenario adaptation.	Prone to hidden bugs, code redundancy, messy architecture; no independent delivery capability.	Routine business iteration, frontend building, rapid demo verification.
Knowledge Output Type	Relies on standard knowledge bases and rules; tool-led, high precision with zero error.	Stable and standardized output; no need for repeated debugging, extremely high efficiency in fixed scenarios.	Narrow boundaries, lack of flexibility; cannot handle complex or medium-to-large scale projects.	Fixed scripts, syntax correction, simple data cleaning.

All mainstream AI coding tools currently on the market belong to the Collaborative Agent category. After extensive business testing, the comprehensive output accuracy of such tools is only around 80%.

The remaining 20% of hidden architectural vulnerabilities, specification issues, business adaptation defects, and performance risks cannot be autonomously fixed by AI and must rely on manual identification and optimization. This leads to the core issue: In the AI era, what's truly valuable has never been typing code.

Conclusion: Abandon the Code Rat Race, Become an Engineering Controller in the AI Era

In the era without AI, the core barrier to programming was 'knowing how to write code,' and everyone competed on manual coding speed, syntax proficiency, and basic code volume. Now, with the popularization of AI coding, the barrier has completely shifted upward. The core barrier to programming has become judgment, review, decision-making, and architecture.

After reading this article, I hope everyone will no longer be trapped in the internal conflict of 'whether to painstakingly learn code,' nor be fixated on transforming into a developer, and certainly not blindly believe that AI can handle all projects with one click. The optimal growth path for ordinary people is: Let AI do the execution, let yourself do the decision-making.

In the next section, I will detail the five-layer core capability ladder for learning programming in the AI era. Welcome to like, follow, and share!

Welcome to follow my public account: 深入浅出AI