Almost two years ago, from May to July 2023, I took on my first full-stack development challenge. That said, I couldn't possibly create complex business applications, so I first challenged myself to develop our company's website (this page). A few months prior, ChatGPT had been released by OpenAI at the end of December 2022. When I attempted development, ChatGPT 3.5 Turbo was the latest version. For the first time in 20 years since my distant student days when I briefly dabbled in HTML only to quickly give up, I faced the command-line interface again. It was quite a reckless challenge of "learning while developing," asking GPT everything from system architecture to terminology, but I was able to build a corporate site with blog application functionality in about a month and a half.

However, on the other hand, the specifications of 3.5 Turbo at that time had significant limitations, such as priority access being cut off every 30 minutes to an hour, making AI unusable for more than 3 hours. I thought it would be impossible to continue like this indefinitely. After all, when error codes appeared, I could only copy and paste the entire text and have AI fix the code, and since the code generated by 3.5 Turbo was fairly often incorrect, I needed to break down tasks and issues into considerably detailed segments.

Two years later

With the release of GPT-5, I took on the security vulnerability responses and reconstruction of the management interface and functions that I had been concerned about but unable to address (specifically, renewing the end-of-support CKEditor 4 to version 5). Since I understood well from daily work that AI had evolved dramatically, I estimated that this reconstruction would probably take about one week (40 hours).

The result was...

Total 6 hours (3 hours × 2 days)

Although the appearance remained mostly unchanged, despite requiring quite fundamental modifications to the underlying mechanisms, it was completed in one-fifth of the estimated time. I can only be astounded by the speed of generative AI evolution. I've summarized several insights I gained from the implementation experience.
 

1. Overwhelming processing speed and capability:
   No matter how many files are loaded, analysis is almost instantaneous, generating accurate code in one shot. It's truly another dimension.
    

2. Communication through images is possible:
   By providing screenshot images of UI errors, it accurately grasps the situation and immediately presents corrective code.

    

3. Near real-time access to the latest information from S3 and Heroku manuals and tutorials:
   While 3.5 Turbo had a knowledge cutoff more than six months prior and hadn't learned the latest errors and bugs, GPT-5 has completely caught up.

    

4. Ability to assign complex tasks to AI agents:
   Utilizing agent functionality, it works autonomously while providing progress reports as needed, allowing me to spend most of my time on "issue segmentation."

    

5. AI learning accumulation:
   With vast historical data already learned, it comprehensively handles even related code changes accompanying a single modification.

Such evolution in just two years. What will happen in the next three years...? What can be said with certainty is that system development and service approaches will undoubtedly change fundamentally. So-called coding personnel will truly cease to exist, I believe. This has the potential to fundamentally change the structure of Japan's IT industry, which has been ridiculed as "IT manual labor" under a multi-layered subcontracting structure. Companies positioned in Tier 2 and 3 will likely face considerable difficulties.

Furthermore, SaaS businesses, which were originally intended as solutions to such legacy systems and have expanded their markets and led the past decade's market, will certainly be significantly impacted by AI. I would like to write about this separately at another time.
Nevertheless, I can only tremble at the speed of evolution.