Persistent Failures in Processing Tasks with ChatGPT

Persistent Failures in Processing Tasks with ChatGPT

Dear OpenAI Support Team,

I have been working extensively with ChatGPT to process a dataset of ports, aiming to generate detailed descriptions, rankings, and classifications. However, after numerous attempts and countless hours, the results remain unsatisfactory, and I feel compelled to share the situation for technical review.

Key Issues:

  1. Task Inconsistency: Despite providing clear instructions and agreeing on a specific output format (a structured table with detailed descriptions, rankings, and notes for missing information or non-port locations), the outputs are frequently inconsistent.
  • Some outputs include generic placeholders (e.g., “Port_11, Description needed.”) or lack meaningful details.
  • Rankings and classifications (e.g., N/P: “*”, “X”) are often missing or incorrect.
  1. Degradation of Output Quality Over Time: While the initial results in a session are relatively good, the quality of the descriptions and data accuracy deteriorates rapidly, suggesting memory limitations or processing constraints.
  2. Excessive Iterations: I have spent countless hours across multiple sessions attempting to correct and refine the output. Each session starts afresh, ignoring prior progress or context, making the task laborious and inefficient.
  3. Formatting Issues: At times, the outputs are presented in overly complex formats (e.g., HTML tables) or improperly formatted for spreadsheet use. Simpler tabular outputs agreed upon earlier are not consistently provided.

Time Spent:

  • I have dedicated several hours across multiple days to this task, trying different approaches, prompts, and adjustments. Yet, the results remain incomplete and unreliable.

Summary of Expectations:

  • A table with Port Name, Description, Ranking, and N/P Notes, structured in a consistent and usable format.
  • Descriptions that are specific, relevant, and non-generic, focusing on the cultural, historical, or geographical significance of the location, particularly for recreational sailing.
  • Clear identification of non-port locations (N/P: “X”) and minimal use of placeholders for missing information (N/P: “*”).

Impact:

This ongoing frustration undermines my trust in ChatGPT for complex tasks. I recognize the potential of the system, but its inability to reliably deliver on this project is causing significant delays in my work.

I hope this feedback reaches the appropriate technical team for review and resolution. Please let me know how to proceed or if additional details are needed.

Thank you for your attention and assistance.

Best regards,

The pain of first movers… I feel you - sending lot’s of love and appreciation!

Thank’s on behalf of all the guys who just wait until the bug reports stop.

@edwinarbus maybe you might be able to help here?

I have been in love with GPT but I think these enormous amount of failures and limitattions are incongruent with that romantic idea of a real partner with superpowers and now I feel desappointed.

1 Like

Nah… it’s going somewhere… It’s not that there have been libraries and everything for that production ready like Apache Airflow or other stuff…

Beside all that - I am happy that ChatGPT exists. Building like 5-6 plugins per day while having time to be in this forum and suck up all the pain of people - like everyone is complaining and nobody really writes about what works fine.

Some things are interesting though and the combination is hillarious. :smile:

I still feel like my brain is on steroids and downers the same time haha…

Even had time to figure out a frontend to see what the stuff does in the background.

A digital interface titled "Systemic Functional Linguistics Graph" features sections for "Ideational Metafunction Assignment" and a text input box labeled “gib”. (Captioned by AI)

I mean seriously! This would have taken me at least two days to make it manually instead of just a couple hours… including the research for the theories…

Nah, ChatGPT ist cool! It is not the only fish in the water but It will be the last fish so I plan on staying…

However, it cannot handle 50 lines of a chart simultaneously. Your example likely demonstrates the type of work it could approach, but the real task requires managing large amounts of code in a tedious and detailed manner. From my frustrating experience, I’ve realized that I cannot rely on this agent in the way I most need it.

Additionally, I’ve wasted enormous amounts of time during iterations because it regenerates everything from scratch in every cycle, without the ability to modify or replace just a specific part of the output. ChatGPT itself has explained that it lacks the capability to isolate a portion of the overall result and work on it exclusively. For instance, the concept of dividing an image into a matrix to focus only on the necessary part is not something it can accomplish.

Using different instances or sessions of the chat isn’t a viable solution either, as each instance requires refining the results using the prompt ChatGPT itself generated. This approach to reinitialize a task while preserving most instructions is so time-consuming that it becomes impractical as a strategy for this kind of work.

I think you are expecting wrong what it can and what it should.

Instead of letting it create tons of boilerplate code you should build an architecture, make sure it works fine and then create tons of boilerplate code and templates for the model to use…

it can’t handle architecture - fine let’s not give it architecture work. make sure it can work on small chunks of code and organise agents to this workflow…