•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
Vibe coding—where developers rely on AI to generate entire applications from vague prompts—may look straightforward in viral social media videos, but real-world evidence suggests it is far from the way experienced programmers actually work.
In a description attributed to Andrej Karpathy, vibe coding is characterized as “trusting the vibe completely, forgetting that code exists, and not reading diffs anymore.” Yet a new study from UC San Diego and Cornell found that none of the observed professional programmers were truly practicing this approach.
The study followed 112 professional programmers using AI agents in their real work. Researchers directly observed 13 programmers as they worked with AI and surveyed an additional 99, all with at least three years of professional experience, with some reporting up to 25 years.
The results showed that experienced developers did not “sit back” while AI produced production systems without oversight. Instead, they planned before writing prompts and used AI for tightly scoped tasks.
Rather than delegating broadly, programmers mapped out architecture, constraints, and edge cases first, then provided AI with specific instructions. They also reviewed code changes line by line, citing prior experience with the consequences of insufficient review.
Developers limited the scope of AI’s impact, delegated smaller tasks, and intervened when issues touched multiple systems. The study describes their approach as treating AI like a fast junior programmer that requires supervision, not like a senior engineer capable of working independently.
The study also points to concerning findings from other experiments involving AI in real workflows.
A key finding was that developers felt positively about AI agents only while they retained control. When control was loosened, the study reports that code quality collapsed—something experienced developers recognized from their own experience.
In the study, all 11 participants building new features carefully reviewed every change AI produced. Three additional participants did not review code directly but monitored program results closely and were prepared to intervene. One participant used AI only to explain the architecture of the codebase rather than to write code.
On prompts, developers emphasized clear context and detailed instructions, including techniques such as screenshots, file references, examples, and step-by-step thinking. Some also used an external AI system to improve prompt quality before sending it to the main AI.
Developers also applied user-focused rules to enforce project standards and adjusted AI behavior based on prior interactions.
Developers rated AI as helpful for speeding up simple, repetitive tasks and for creating basic code frameworks. They described AI as effective for writing tests, documentation, simple edits, and fixing small bugs.
As complexity increased, AI’s suitability declined. Developers avoided AI for business logic or tasks requiring deep domain knowledge, and the study reports that no one believed AI could replace human decision-making.
One software engineer with 20 years of experience said: “I have been a software engineer and data analyst for 20 years, and I would never go back to writing code by hand. That ship has sailed and is fine. But a warning for younger developers: it remains crucial to know what you’re doing. AI is great for creating demos for those who lack technical understanding, but to move beyond that and produce something close to production you need substantial supervision.”
While vibe coding can generate impressive social media clips, the study’s findings indicate it does not produce high-quality software in practice. The differentiator, according to the evidence presented, is not whether AI is used, but whether developers maintain rigorous review processes, tightly scoped task division, and a clear understanding of what AI can and cannot do.
Premium gym chains are entering a “golden era” that is ending or already in decline, as rising operating costs collide with shifting consumer preferences toward more flexible, community-based ways to exercise. Long-term memberships are shrinking, margins are pressured by higher rents and facility expenses, and competition from smaller, more personalized…
