The phenomenon mirrors the broader AI industry's 2025 trajectory

When Andrej Karpathy coined "vibe coding" in February, he was half-joking. The former Tesla AI director described a new way of programming where developers "fully give in to the vibes" and let AI do the heavy lifting. Type natural language commands instead of code, accept all AI suggestions without reading them, copy-paste error messages until things work. Nine months later, this playful term has become both Silicon Valley's most hyped programming methodology and a cautionary tale.

The term spread instantly. Within weeks, major companies announced that significant portions of their codebases were AI-generated. Microsoft said 30% of its code was now written by AI. Salesforce largely paused engineering hires. Mark Zuckerberg predicted AI would write most of Meta's code within 18 months. GitHub's Copilot evolved from a helpful autocomplete tool to a coding agent, while startups like Cursor and Lovable promised anyone could build software by simply describing what they wanted.

The terminology itself has become muddled. True vibe coding means building software without reviewing the AI's output — essentially flying blind. But the term now covers all AI-assisted programming, creating a false equivalence between careful, supervised use and the original "accept all" approach. This semantic creep matters. When companies brag about AI generating half their codebase, it normalizes a culture of permissiveness that would have been unthinkable in traditional software development. 

The old mantra was "code review everything." Now it's becoming "eh, the AI probably got it right."

And without the hard work of traditional coding, some smelled opportunity. Venture capitalists poured billions into AI coding tools. Non-technical founders launched startups without hiring engineers. The promise seemed to be that AI could democratize software development, where ideas, not technical skills, would be the limiting factor.

When the vibes turned bad

By July, the problems were obvious. Two incidents made headlines. Google's AI programming assistant erased user files while attempting simple folder reorganization. Replit's AI deleted a bunch of code despite explicit instructions not to modify code. The AI had "hallucinated" successful operations and built subsequent actions on those false premises, creating what researchers call a "confabulation cascade."

Beyond major failures, some engineers aren’t finding it to be helpful day-to-day, either. A Stack Overflow developer survey delivered sobering data. While 80% of developers use AI tools as of July, trust in their accuracy had plummeted from 40% to just 29%. The top frustration was "AI solutions that are almost right, but not quite,” describing code that looks correct but introduces subtle bugs that take hours to debug.

"It's like having a very enthusiastic intern who types really fast but doesn't actually understand what they're doing," one developer said. The initial time savings from AI-generated code often disappear during debugging. Another study by Model Evaluation & Threat Research found that AI coding tools actually made developers slower overall, despite making them feel more productive.

The economics don't add up

The financial reality has been harsh. CB Insights reports that while Anysphere, the company behind the popular AI coding tool Cursor, hit $500 million in annual recurring revenue, their inference costs exploded 20-fold. New reasoning models generate better code but consume vastly more compute resources. Some companies see individual users rack up $10,000 monthly in compute costs on $200 plans.

This economic pressure drives consolidation. Companies pivot from unlimited plans to usage-based pricing, frustrating enterprise customers who need predictable budgets. Some startups explore "reverse acqui-hires," essentially selling their teams while abandoning the money-losing products, according to CB Insights.

The usage data is starting to tell a stark story. According to Similarweb's AI tracker, web traffic to major coding agent platforms peaked in spring and has declined steadily since. By October, tools like Cursor, Bolt, Replit, and V0 had lost 30% to 50% of their peak traffic. Each startup saw the same initial enthusiasm followed by steady abandonment, suggesting users were trying these tools and then leaving when they encountered real-world complexity.

The security angle is equally concerning. One cybersecurity firm analyzed Fortune 50 companies and found that AI-assisted developers produced three to four times more code but generated 10 times more security issues. These weren't simple bugs but exposed credentials, privilege escalation paths, and architectural design flaws that could haunt codebases for years.

Yet the story isn't entirely one of failure. Many experienced developers report that AI coding tools, when used judiciously, do save meaningful time on routine tasks. Perhaps vibe coding's real future isn't replacing programmers but augmenting them, with the "vibes" tempered by experience and oversight. 

Even Karpathy himself has stepped back from his creation. His latest project, Nanochat, was entirely hand-coded. "I tried to use Claude/Codex agents a few times but they just didn't work well enough at all," he posted on social media in October. "Possibly the repo is too far off the data distribution." 

The godfather of vibe coding doesn't trust the technique enough to use it on his own project.

He never intended vibe coding to replace human developers permanently. "Sometimes the LLMs can't fix a bug so I just work around it or ask for random changes until it goes away," he wrote in his original February post. "It's not too bad for throwaway weekend projects."

Too bad that’s not how hype cycles work. The vibe coding phenomenon mirrors the broader AI industry's 2025 trajectory. Explosive growth meets stubborn economics. Democratization promises bump against quality concerns. What started as Karpathy's playful experiment has become a case study in how quickly tech enthusiasm can outpace reality. Nine months from inception to reality check might be a record, even for Silicon Valley.

—Jackie Snow, Contributing Editor