This Week's Term: Vibe Coding.
Vibe coding is software development where AI generates most or all of the code based on high-level descriptions or prompts - it's all about the vibe, not the implementation details. The term gained prominence in February 2025 when AI researcher Andrej Karpathy tweeted: "The craziest thing about developing software with LLMs is that you can forget that the code even exists. Just vibe with the product."
This represents a fundamental shift in who can build software. In the 2024 Y Combinator batch, roughly 25% of startups reported having 95% or more of their code written by AI. Non-technical founders are building full applications in hours using tools like base44, Lovable, Bolt.new, and Cursor. One founder created a complete event management application in just one hour using Lovable - a task that would have traditionally required weeks of development work and a technical co-founder.
It's important to note that vibe coding is not the same as professional AI-assisted development. As developer Simon Willison clarified in his widely-cited article, vibe coding means "building software with an LLM without reviewing the code it writes." It's the difference between using AI as a drafting tool with human review versus accepting generated code blindly. This distinction has real consequences - stories abound of applications that delete user databases, expose security vulnerabilities, or trap founders in "development hell" when they need to make changes beyond the AI's initial output.
The business stakes are high. The barrier to building software has dropped dramatically - anyone with a clear vision can now prototype rapidly. But the risk-reward calculation has shifted: speed to market comes with technical debt that may be invisible until it's expensive to fix. Microsoft predicts that by 2030, 95% of all code will be AI-generated - but the winners will be those who understand when to vibe and when to verify.
The parallel to traditional business applies: CEOs don't personally audit the books, but they spot-check numbers and ask pointed questions when something doesn't add up. Product managers don't write the code, but they use the product extensively and probe edge cases. With vibe coding, the same principle holds.
Your management framework:
Use it relentlessly - Be the power user of what you built. Click every button, test every flow, break it intentionally
Ask targeted questions - When something behaves unexpectedly, ask your AI tool: "Why does X happen when I do Y?" or "What happens if a user tries Z?"
Spot-check the logic - Request explanations of critical functions: "Explain how user data is stored" or "Walk me through what happens when payment fails"
Test with real scenarios - Don't just test happy paths. What happens with bad data, concurrent users, or system failures?
The decision framework: Use vibe coding for speed-to-market when the cost of being wrong is low. Require expert review when the cost of failure is high (user data, payments, security, scale).
But don't give up on taking vibe coding all the way to production, as exemplified in the video below from Anthropic, who accepted 22,000 lines of code in product without reviewing all of them. The main insights and stories are up to minute 16, and then there are some interesting questions and answers, which are a bit more technical.