X Discussion: Vercel CEO Guillermo Rauch Says AI Will Rewrite All Human-Made Software, Sparking Debate on Formal Verification and Rewrite Risk

Core Background and Industry Pain Points

AI security scanning and auto-fix tools have shipped at a fast pace in recent weeks: OpenAI described an automated pipeline that found Linux kernel privilege-escalation exploits, and Vercel launched Security Review in public beta. On November 5 at 02:22 UTC, Vercel founder and CEO Guillermo Rauch posted on X: "We will all be learning in coming weeks & months, that most human-made software was really slow, had many security vulnerabilities, was under-specified and incorrect." He continued: "AI will rewrite it all. Many times." This is Rauch's personal opinion, not a Vercel announcement.

Architecture Highlights and Underlying Mechanisms

Rauch's picture goes beyond faster code generation to a full engineering loop: AI that thinks about failure modes humans could not, then proves, formally verifies, load tests and hardens software, turning memory-unsafe foundations into robust, optimized ones, and documents it with diagrams, animations and slide decks on request. He called it "the true birth of software engineering." About an hour earlier he had shared Vercel Security Review figures as supporting evidence.

Benchmarks and Measured Results

There are no new public benchmarks in this discussion; none were published. What can be cited is engagement: when fetched, the post had about 3,060 likes, 272 reposts, 681 bookmarks, 256 replies and ~99K views. The Vercel figures Rauch shared are vendor-reported: 6,817 vulnerabilities found in 24 hours, 580 critical or high, 1,411 already patched by Vercel Agent, with Security Review free until November 10.

Developer Playbook and Getting Started

Quote posts split clearly. Some say this is exactly why their team did a from-scratch AI rewrite, and one cites Claude-written machine-checked proofs surfacing bugs that years of fuzzing missed. Critic @genevievebenoit argues that rewriting under-specified code just ships the same wrong behavior faster, so the spec has to be written first; @403access says formal verification raises quality but bugs move up the abstraction level instead of disappearing. Practical advice: add tests and specs before letting AI rewrite a system, and keep people who know the business logic reviewing critical paths.