Survey finds REA“Replicating any software” mostly means reproducing partial functionality

AT A GLANCE

51 public records show that REA more often completes small games or partial functionality, while complex commercial software remains difficult to reproduce in full.

Article

REA“Replicate any software” put to empirical test: 51 public attempts, with most successes limited to partial functionality

Beating AI newsflash: independent technical author quanglee reviewed 51 public usage records and found that so-called “successful replication” often merely meant reproducing a single feature.REA October 12 He published the findings: of these, 9 cases came from REA project contributors, all of whom achieved their respective goals; among the other 42 independent attempts, 20 were completed, 16 were partially completed, and 6 failed.

The issue lies in the standard for “completion.”20 Among the independent success cases, 9 targeted small games or lightweight applications, 6 reproduced only one feature of larger software, and 5 did not even disclose the target’s name. One relatively strong case was the puzzle game Untangle:Agent analyzed its 167 KB program and created a playable web version. Of 54 technical conclusions, 52 were entirely correct. However, a subsequent audit by the developer found that 10 function and field names did not originate from REA analysis results; the model may have memorized the original version’s publicly available source code.

More complex projects require substantial work.REA contributors reverse-engineered Imperishable Night, making 1107 of 1106 functions match the original machine code. This took approximately 2 weeks and built on prior manual research. Another person attempted to replicate the voice-input software Wispr Flow. Although they analyzed parts of the recording, transcription, and text-pasting workflow, they ultimately produced only a 64-line Python script. Features relied upon by the original—including cloud-based models, context processing, and real-time feedback—were not fully reproduced.REA can help Agent invoke tools such as Ghidra and IDA to analyze existing programs and reimplement some functionality. However, the capabilities of commercial software do not exist solely in its installation package: models, data, and complex business logic on servers cannot be directly reconstructed by decompiling the client.

Original link https://m.theblockbeats.info/flash/371701

Key points

Independent technical author quanglee reviewed 51 public usage records REA and published the findings on October 12.
51 Of the records, 9 came from REA project contributors, all of whom achieved their respective goals; 42 among the independent attempts, 20 were completed, 16 were partially completed, and 6 failed.
20 Among the independent success cases, 9 targeted small games or lightweight applications, 6 reproduced only one feature of larger software, and 5 did not disclose the target’s name.
In the Untangle case,Agent analyzed 167 KB program and created a playable web version; 54 of the technical conclusions, 52 were entirely correct.
A subsequent audit by the Untangle developer found that 10 function and field names did not originate from REA analysis results; the article said the model may have memorized the original version’s publicly available source code.
REA Contributors reverse-engineered Imperishable Night, making 1107 of 1106 functions match the original machine code. This took approximately 2 weeks and built on prior manual research.
The Wispr Flow replication attempt analyzed parts of the recording, transcription, and text-pasting workflow, ultimately producing only a 64-line Python script.
Features relied upon by the original Wispr Flow—including cloud-based models, context processing, and real-time feedback—were not fully reproduced.
REA can help Agent invoke tools such as Ghidra and IDA to analyze existing programs and reimplement some functionality.
The article notes that models, data, and complex business logic on servers cannot be directly reconstructed by decompiling the client.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

REA It can assist in analyzing existing programs and reimplementing some functionality, but “creating a similar feature” is not equivalent to “fully replicating an entire software product.” Public records show that small games, lightweight applications, and individual features are easier to complete; where cloud-based models, data, real-time services, or complex business logic are involved, analyzing only the client installation package is insufficient.

Why it matters to readers

These tests narrow the practical meaning of “successful replication”:REA it demonstrates capability in program analysis and partial feature reimplementation, but the available evidence does not support the claim that it can easily replicate any complete software product.

When assessing such projects, beginners should first consider the success criteria, scope of reproduction, human effort, and server-side dependencies, rather than relying solely on the phrase “successful replication.”

Risks and unknowns

The material is a retelling of quanglee's investigation and does not provide the investigation's original publication URL, a complete list of cases, or reproducible experiments.
5 instances classified as independent successes do not disclose the target names, so their scope of completion cannot be verified from the material.
In the Untangle case, the 10 names may have originated from model memory; the material only states that this is “possible,” and it has not been confirmed.

Survey finds REA“Replicating any software” mostly means reproducing partial functionality

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center