GPT-6 Astra: Results from 5 Challenging Tasks

The CodeWithHarry channel set five seemingly impossible tasks to verify GPT-6 Astra's capabilities.

Instead of theoretical debate, the aim was to measure actual work performance to differentiate AI's limits from human roles, and a report card with scores out of 100 was planned based on the test results.

GPT-6 Astra was instructed to copy adjustment layer effects from a specific Premiere project (template.prproj) and apply them to another project.

It accurately identified and replicated color correction values such as white balance, tint, saturation, and exposure, and was asked to apply an audio noise reduction effect (De-noise) at an 11% level.

The AI successfully completed this editing task in just 3 minutes.

This task was applied to the "archives and compressions.mp4" file and created as a new Premiere file.

While the 11% accuracy for audio effect application was 100%, overall productivity was rated low.

Due to these factors, it received a score of 63 out of 100.

GPT-6 Astra failed to integrate with Premiere Composer, causing an internal model error.

The AI arbitrarily removed silent sections, making the video sound robotic and unnatural, and exhibited poor irregular cut editing and black frame handling.

As a result, it received a low score of 33 out of 100.

To verify its coding ability, GPT-6 Astra was given the task of cloning the ilovepdf.com website.

It was required to include all functions of the original site and implement features using the latest version of Next.js.

A test was conducted to create a PDF editing website, assuming senior developer-level scalability and industry standard application.

Senior developer-level scalability and industry standard application were requested, and the entire project creation and local host operation were required to be verified.

The PDF split function also accurately performed page-by-page splitting, showing successful results in implementing core functionalities.

The PDF compression function was found to be inoperable after testing a 5.1MB file, as there was no change before and after compression.

The website development took a total of 52 minutes, which was considered too long for an AI to create such a website.

Functional errors and the long time taken acted as deduction factors, resulting in a score of 65 out of 100 for the coding level evaluation.

Overall, the failure to properly implement one of the core functions, the compression feature, negatively impacted the evaluation.

Full access was granted to develop a browser-based video editor to verify if AI could perform a task that would take professional editors several months.

The user instructed the AI to create an editor similar to Premiere Pro in every aspect, including shortcuts and interface, in a specific directory path, integrating dozens of features.

Full access was granted and approximately 1 hour of generation time was allowed to verify if AI could perform a task that would take professional editors several months.

The video editor interface successfully launched in a browser environment through the Next.js code generated by the AI.

It was confirmed that while not 100% identical to Premiere Pro's interface, it was implemented as a web app with basic editing functions.

Keyframe-based scaling and position value adjustment functions also worked correctly.

These features demonstrate that even a browser-based editor provides a significant level of editing capabilities.

This allows users to perform basic video editing tasks in a web environment.

The attempt to finally export the edited video as an MP4 file via the browser encoder was successful.

Actual rendering and the application of edited keyframe effects were confirmed in the JavaScript environment, proving the completeness of the AI-generated editor.

The browser-based video editor generated by AI successfully performed core video editing functions such as color correction, exposure, and saturation adjustment.

Despite being browser-based, it implemented advanced interfaces like full-screen mode, and showed astonishing speed and efficiency, generating results in just 49 minutes.

However, limitations were pointed out regarding practical operation, such as scaling and bug response issues that might arise in complex production environments.

The analysis suggests that while AI-generated code is efficient in the initial stages, human expertise is essential for long-term maintenance.

This functional implementation took only 49 minutes, which is considered a groundbreaking advancement compared to technology levels before 2021-2022.

GPT-6 Astra was instructed to collect video editing resources for the YouTube channel 'CodeWithHarry'.

It performed the task of downloading various media assets such as B-roll, images, PNGs, and SVGs from the internet and organizing them into folders.

Search optimization for file names (e.g., wash, click) was done automatically, and appropriate video sources were successfully collected from external sources like Mixkit.

This automation significantly boosted video editing efficiency by replacing previously manual tasks.

The materials collected by the AI were well-organized with search-friendly names.

Various sources such as coding, statistics, logos, and sound effects were secured, and the data was appropriately grouped by category, deemed practical for real use.

The resource collection task received a satisfactory evaluation for its time efficiency, noting a speed humanly impossible.

However, quality issues in data composition arose, such as the omission of some key logos.

Many unnecessary elements, including numerous irrelevant illustrations, were included in the results, leading to a score of 75 out of 100.

Overall, the slightly slow task execution time was the core factor in the final evaluation.