TL;DR: The hype around "Vibe Coding" (building software without knowing how to code) has hit a reality check. UC Berkeley Professor Sarah Chasins argues that manual coding skills are more necessary, not less, in the GenAI era. Why? Because AI is excellent at the "Old," terrible at the "New," and dangerously deceptive about speed. A recent study shows that LLMs might actually make you 20% slower, even while you feel 20% faster.
James here, CEO of Mercury Technology Solutions. Hong Kong - February 1, 2026
There is a dangerous narrative spreading in the tech world right now: "I don't need to learn Python; I just need to learn English. The AI will do the rest." We call this "Vibe Coding."
But UC Berkeley Computer Science Professor Sarah Chasins recently dropped a truth bomb that every aspiring "No-Code" founder needs to hear: "Yes, you still need to be able to write code the 'old-fashioned' way in order to get programs out of these GenAI tools."
If you think you can skip the hard work of learning computer science fundamentals, you are setting yourself up for failure. Here are the three reasons why.
1. The Decomposition Deficit
The hardest part of engineering isn't syntax (writing the if/else statement). It is Decomposition. Decomposition is the ability to look at a massive, fuzzy, ambiguous business problem and break it down into thousands of tiny, answerable logic gates.
- The Reality: AI cannot do this for you.
- The Skill: AI is a tactical executor. It needs you to be the Strategic Architect.
If you haven't trained your brain to deconstruct problems (which you learn by coding manually), you won't even know what to ask the AI. You will give it vague prompts and get vague, broken garbage in return.
2. The "Pray and Deploy" Strategy
If you can't read code, how do you know the AI is right? You don't. You are just praying.
You are relying on "Blind Faith Deployment."
- Maybe you have 20 test cases.
- Maybe you have a friend check it.
- But mostly, you are guessing.
Professor Chasins points out that without manual coding literacy, you lose the ability to Verify. In enterprise software, deploying unverified code is negligence. You need to be able to read the Matrix, not just look at the woman in the red dress.
3. The Innovation Ceiling (The Mirror Trap)
AI is a mirror of the past. It is trained on existing GitHub repositories.
- Vibe Coding works if you are rebuilding something that has been built 1,000 times before (e.g., a To-Do List app, a basic CRM).
- Vibe Coding fails if you are building something truly novel.
If you are trying to innovate—to write a program that has never been written before—the AI will hallucinate because it has no training data to rely on. At that moment, you are on your own. You have to return to first principles and write the logic yourself.
4. The "Efficiency Hallucination"
This is the most fascinating (and frightening) part of Chasins' research. We often talk about AI "Hallucinating" facts. But it turns out, AI causes Humans to hallucinate efficiency.
A study cited by Chasins revealed a paradox:
- Perception: Users utilizing LLMs felt they were working 20% faster.
- Reality: Those same users were actually 20% slower than the manual control group.
The AI gives you a dopamine hit of "progress" (look at all this generated text!), masking the reality that you are spending more time debugging and fixing its mess than if you had just written it yourself.
Conclusion: Don't Be a Passenger
AI is a powerful engine. But an engine without a driver is just a noise machine. To survive in 2026, you cannot just be a "Viber." You must be an Engineer.
You don't need to write every line, but you must understand every line. The goal isn't to escape the code; it's to master it so you can command the machine effectively.
Mercury Technology Solutions: Accelerate Digitality.
