Software Is Dead? I Tested the Viral REA Repo | With and Without It, same score?

Software is dead? I tested the viral REA (Reverse Engineer Anything) repo—with and without the tool—and compared what AI could actually rebuild.

The result: 20/20 both ways, a real OBS reconstruction, and a payment demo that exposes the gap between a demo and dependable software.

REA is impressive. But did it do the engineering, or did the AI model? I opened the repo, followed one request through the tools, compared controlled tests across four AI companies, and separated REA from the unrelated PhotoCraft/ArtCraft story.

WHAT YOU'LL SEE

• A byte-for-byte painter rebuild: 20/20 with REA and 20/20 without it

• CLI-reported list-price cost estimates in this painter test: $1.09 with REA vs $1.25 without (~13% savings)

• The OBS noise-gate test: why trusting decompiler pseudocode can fail

• A mock payment batch of 1,000 intended orders: retries, idempotency, duplicate charges

• How I exported my own video timeline into DaVinci Resolve

• Five save-worthy cards: tools, test methods, prompts, build-vs-buy and what to learn

CHAPTERS

0:00 Cold open

0:32 What you'll take away

0:47 The fear, and two different stories

1:36 Why people actually care

1:52 What the compiler throws away

3:06 The test, and run one

3:53 Open the repo

4:43 What REA actually is

5:11 One question's journey

5:57 A real REA report, from my own test

6:38 Who did what

7:19 Now take REA away

7:47 Four AI companies, one gives up

8:13 OBS: a real program

9:09 REA's own playbook

9:34 People did this before

10:30 What REA's own examples show

11:39 Free your own data

12:34 Time, not cost

13:00 Same AI, two prompts

13:50 Experience can't be prompted

14:36 What to learn next

15:18 What the industry is seeing

15:45 Build, or buy?

16:29 The CEO's answer

TOOLS AND PRIMARY SOURCES

REA: https://github.com/morluto/rea | https://rea.tools

PhotoCraft: https://github.com/storytold/photocraft

Compiler Explorer: https://godbolt.org

Decompiler Explorer: https://dogbolt.org

GhidraMCP: https://github.com/LaurieWired/GhidraMCP

IDA Pro MCP: https://github.com/mrexodia/ida-pro-mcp

leanrea: https://github.com/aicodeking/leanrea

DeepWiki: https://deepwiki.com

Escape's security research: https://escape.tech/blog/methodology-how-we-discovered-vulnerabilities-apps-built-with-vibe-coding/

Tyler Denk's beehiiv post: https://x.com/denk_tweets/status/2110371582153748942

PAYMENT DEMO CLARIFICATION

The ten-order and one-lakh-order examples in the narration illustrate a small demo and a production scenario. The measured mock run used a batch of 1,000 intended orders. The beginner run recorded 1,182 charge events across 400 unique order IDs, all charged more than once. The experienced run recorded 1,000 charge events across 1,000 unique order IDs, with no repeats. An idempotency key prevents duplicate effects only when the server enforces it.

The test data shown are specific to our experiments; not a claim that all apps can be reconstructed. Our payment demo uses a mock server and is not payment-integration advice.

Subscribe for AI Reliability Engineering experiments: https://www.youtube.com/@Qualixar-AI

X: https://x.com/varunPbhardwaj

Instagram: https://www.instagram.com/varunpratapbhardwaj/

Research: https://qualixar.com

Personal site: https://varunpratap.com

This video uses AI-generated avatar clips of me.

#ReverseEngineering #AIEngineering #REA