After two years of vibecoding, I’m back to writing by hand

Summary

The speaker discusses the challenges and realities of using AI-driven coding agents, highlighting how they can create deceptively good-looking but flawed code. While AI tools like Claude or CodeX can generate code that passes tests, they often result in hallucinated solutions that lack true understanding. The speaker emphasizes the importance of human responsibility in software development, stressing that AI cannot replace the critical human judgment required for complex and secure code. In the end, the speaker reassures developers that their jobs are safe and that AI is far from replacing the human touch in engineering.

Takeaways

• 😀 AI coding tools like Claude or Codeex can produce seemingly good code, but often it reveals flaws when reviewed long-term.
• 😀 Initial excitement about AI-generated code can lead to a false sense of achievement, as users may believe the AI has completed tasks on its own.
• 😀 AI can handle boilerplate code well but often struggles with more complex tasks, leading to failure or incomplete results.
• 😀 The user often ends up doing the heavy lifting, essentially doing the structural work while the AI just fills in parts of the code.
• 😀 As users refine their prompting and specifications, they realize that AI cannot always anticipate real-world production environments.
• 😀 Even with detailed, well-crafted specifications, the AI may produce code that looks functional but is ultimately a 'hallucination' of code—practical, but not sustainable or reliable.
• 😀 Breaking complex tasks into smaller pieces and refining the scope of requests still leads to code that may look good but fails under real-world inspection.
• 😀 Passing tests and looking at isolated diffs in code reviews can be misleading; code may appear fine at first but fall apart when examined in-depth.
• 😀 Software developers are ultimately responsible for ensuring code quality, and AI tools should not replace human judgment and accountability in the development process.
• 😀 While AI can assist in coding, it cannot replace the value and responsibility that comes with human oversight and decision-making in software engineering.
• 😀 The fear of AI replacing software development jobs is overblown; anyone who spends significant time with AI coding agents realizes that human developers are still essential.

Q & A

What is the main issue the speaker highlights about AI-generated code?

The speaker discusses how AI-generated code can initially look fine and pass tests, but upon deeper inspection, it often appears disorganized, unrefined, or even nonsensical. This is especially true after spending more time with the code, leading to the realization that much of it isn't actually solid or reliable.

How does the speaker describe the experience of using AI coding tools like Claude or Codeex?

The speaker compares the experience to a cycle of initial optimism followed by disillusionment. At first, AI coding tools may seem to work well, leading to a dopamine hit of validation. However, over time, the speaker realizes that the AI isn't truly generating effective solutions—it's mostly the user doing the heavy lifting and structuring the work.

What does the speaker mean by 'hallucinations' in AI-generated code?

The term 'hallucinations' refers to situations where the AI produces code that seems plausible or functional at first glance but is actually flawed, unreliable, or misleading upon closer inspection. Despite passing tests and appearing correct, it’s not truly quality code.

How does the speaker suggest improving the interaction with AI in coding tasks?

The speaker suggests breaking down large tasks into smaller, more specific components, writing detailed specifications, and providing clearer prompts. However, even with these improvements, the speaker notes that AI will still fall short of producing truly reliable or well-structured code.

Why does the speaker emphasize the importance of human responsibility in coding despite using AI tools?

The speaker stresses that human responsibility is crucial because AI lacks the ability to understand the full context and implications of code in a production environment. Even if AI-generated code passes tests, it’s ultimately up to humans to ensure that the code is safe, secure, and functional in real-world applications.

What is the main psychological effect of using AI in coding, according to the speaker?

The psychological effect is that the user often feels validated by small successes or improvements from the AI, receiving dopamine boosts. This can lead to the misconception that the AI has truly coded the solution, when in reality, the user has been the one guiding the AI through the process.

What does the speaker mean by 'the deltas look good in isolation'?

This phrase refers to how the individual changes (deltas) in the code may appear fine when viewed in isolation, such as in diffs or pull requests. However, when considering the entire codebase or system, these changes may not hold up or lead to a reliable, maintainable product.

How does the speaker compare AI-generated code to a 'psychedelic airplane'?

The speaker uses this metaphor to describe AI-generated code that looks functional or plausible at first but, upon closer inspection, is revealed to be unstable or fundamentally flawed. The 'psychedelic airplane' is an illusion, something that may seem correct but is far from it when scrutinized.

Why does the speaker disagree with the notion that AI will replace developers?

The speaker argues that AI can assist in coding, but it cannot replace developers. The role of developers involves judgment, oversight, and responsibility—tasks that AI is not capable of performing adequately. Developers are essential for ensuring the safety, reliability, and quality of code, which is why their jobs are not at risk.

What does the speaker mean by 'you owe yourself to read the code'?

This means that developers should not rely solely on automated tests or diffs to verify code. Instead, they should take the time to thoroughly review the actual code to ensure it meets the necessary standards, especially when shipping software that will be used by real customers or handle sensitive data.