Vibes Check

If there's one person you'd expect to write every line of code by hand, it's Linus Torvalds. The creator of Linux and Git has spent decades building software where every instruction matters. So when he revealed he's using AI to generate code for his hobby projects, the internet noticed.

His latest side project, AudioNoise, creates random digital audio effects for a homemade guitar pedal. The Python visualizer? "Basically written by vibe-coding," Torvalds admits. He describes his process: "It started out as my typical 'google and do the monkey-see-monkey-do' kind of programming, but then I cut out the middle-man—me—and just used Google Antigravity."

This tracks with his earlier comments at Open Source Summit Asia: vibe coding is fine "as long as it's not used for anything that matters." The kernel remains hand-crafted. The guitar pedal effects? Let the AI handle it.

Anthropic dropped Claude Code 2.1.0 on January 7th with 1,096 commits—what they're calling their "biggest productivity upgrade yet." The headline features: skill hot-reloading, session teleportation via /teleport, and Chrome browser control directly from your terminal.

But the real story is what's under the hood. Claude Opus 4.5 now hits 80.9% accuracy on SWE-bench—the highest score any model has achieved on real GitHub issues. Background agent support means you can kick off long-running tasks with Ctrl+B and keep working while they execute. Memory efficiency improved 3x, addressing one of the tool's persistent complaints.

One Axios reviewer described the experience: "You just tell it to do something, and it works." They called it "an infinite vibe coding machine"—noting that earlier tools still required understanding the underlying architecture. Claude Code changed that by giving the AI full read/write access to files.

Business Insider surveyed 167 software engineers about their vibe coding experiences. The results split into four camps: 46.9% feel they're "keeping up" with AI tools, 17.9% feel ahead, 16.2% feel behind, and 17.5% have opted out entirely—citing insufficient advancement or lengthy learning curves.

The productivity claims vary wildly. Appfire Principal Solutions Architect Ed Gaile, 55, says AI tools have "doubled, if not tripled" his productivity through reduced context switching. But a METR study from July 2025 found experienced developers spent 20% more time reviewing AI outputs versus coding independently—ultimately reducing overall productivity.

One engineer captured the core tension: "I think the future is learning how to read code." Another warned that over-reliance poses career risks. The survey confirms what many suspect: vibe coding's value depends entirely on the developer using it, the codebase they're working in, and whether they can validate what the AI produces.

The term itself—defined as developers who "fully give in to the vibes, embrace exponentials, and forget that the code even exists"—remains contested. Is it a workflow optimization or a risky shortcut? The 167 engineers can't agree.

GitHub announced it will deprecate four AI models from Copilot on February 17, 2026: Claude Opus 4.1, Gemini 2.5 Pro, GPT-5, and GPT-5-Codex. Suggested replacements: Claude Opus 4.5, Gemini 3 Pro, GPT-5.2, and GPT-5.1-Codex.

The deprecation affects all Copilot experiences—Chat, inline edits, agent modes, and code completions. Enterprise administrators need to verify alternative models are enabled before the deadline.

This accelerates what analysts call "The Unbundled Era." The 2023-2024 Copilot Era was defined by convenience and bundling—one tool, one subscription. The 2026 landscape is defined by control, specialization, and arbitrage. Developers now mix and match: Cursor for exploratory work, Claude Code for large refactors, Copilot for completions.

The numbers are finally coming in, and they're complicated. According to Second Talent's research, 74% of developers report productivity increases when using vibe coding. But 63% have also spent more time debugging AI-generated code than they would have spent writing it themselves—at least once. About 45% of AI-generated code contains security flaws.

Tricentis VP Damien Wong argues in CDOTrends that the quality debate misses the point: "AI can accelerate productivity and open software creation to more people, but the value of that speed depends entirely on the quality of the code it produces." His solution? Intelligent testing embedded directly into development workflows—"from a reactive checkpoint into an always-on safety net."

The stakes are real. Research shows 61% of organizations rank software outages among their top business threats. JetBrains' Developer Ecosystem survey found 48% of developers prefer to "stay hands-on" for testing and code reviews, even when using AI for everything else.

One developer's description of inheriting a vibe-coded project went viral: "It was like a beautiful house built on a foundation of sand. Everything looked great until you tried to add a new room, and the whole thing started to crumble." The metaphor captures exactly why the debate persists—and why it can't be resolved with productivity percentages alone.