Saturday, November 21, 2026

Can we rely on ChatGPT to do some honest work?

 Most AI's on the Internet, the web based agents like Copilot, Gemini, ChatGPT etc are good for simple chats and they excel Googling. Copilot once made a complete reservation in an Indian restaurant to my family including prepayment arragnment etc. to find out the restaurant refused the order that Copilot certified good and even gave links for the transaction. 

However, to rely on them for tasks that requires planning and building over a longer stretch requires a lot of vigilence from the human entity trying to use the AI for such a purpose.

I am using a basic, paid version of ChatGPT working on an idea of mine to build Screen Developer program for all kinds of frameworks, Web, Kivy, etc.

Here is a confession from ChatGPT of not doing an assigned task. Besides retention disorder, hallucination, and context rot, it has many other browser related problems using AI.

I just hit a hardwall in the project when I found that ChatGPT was going ahead with something quite different pulling out of as if thin air. Here is a confession:

You're right. I made a fundamental mistake, and I should not have encouraged you to continue testing on that basis.

We have spent months developing a control model through ScreenBuilder, ScreenContract, and IntegratedScreenDesigner. The controls in the screenshot are merely the palette and inventory implemented in this new Tkinter designer. I treated that inventory as though it represented the agreed to model. It does not.

There are several important distinctions I failed to respect:

  1.  the established control vocabulary and semantics.

  2. unique element IDs and names, independent of displayed text.

  3. including type-specific properties and geometry.

  4. the established interaction model.

  5. connections between controls and application data.

The vocabulary we had previously discussed included containers such as VERTICAL, HORIZONTAL, etc. That is a historical reference, not permission for me to declare a new canonical list without checking the actual specifications.

We should stop work on the next data-source step.

The correct task is to audit GenericScreenGenerator 1.xxx against the established ScreenBuilder/ScreenContract model. We need to establish exactly what has been carried forward, what has been simplified or replaced, and what is missing.