ChatGPT-4o: At the verge of AGI v.01 Alpha - Part 1

It’s truly impressive what OpenAI unveiled and made available on Monday, May 13th and we are still digesting it.

Back in November 2022, OpenAI introduced a chatbot powered by an advanced LLM, capable of generating text on various topics, both imaginary and historical. It was a game-changer.

Fast forward to May 2024, and OpenAI has launched a groundbreaking assistant. This isn’t just any chatbot—it’s an assistant that asks questions to understand tasks better, iterates for higher quality responses, and performs functions we’re still discovering.

Today’s article highlights one functionality that hasn’t received mainstream attention but has amazed us over the past few days.

One of the most celebrated capabilities unveiled last Monday was the deep analysis of spreadsheets. You can feed ChatGPT a complex spreadsheet, and it can analyze and plot the information within. That’s truly astonishing.

Here you have the official article by OpenAI.

What’s even more impressive, and less publicized, is that ChatGPT can create spreadsheets from scratch and return them as downloadable files and it can also interpret and execute tasks within these spreadsheets. Beyond that, it can access and manipulate your Google Sheets and Microsoft Excel spreadsheets directly.

Our Exercise

We undertook an exercise to process a spreadsheet containing two distinctly different worksheets. The task involved identifying and filling in any columns that could be interpreted as containing answers to questions.

Worksheet #1: This worksheet was relatively straightforward to handle. It had a clear structure with an easily identifiable column for questions. The task here was to generate detailed and informative answers based on the questions provided and populate the designated answer column accordingly.

Worksheet #2: This worksheet presented a greater challenge as it was in a different language and had a different structure. Despite the language barrier and structural differences, we successfully identified the relevant columns and filled them with detailed responses generated from the questions present in the worksheet.

The results were on point

Our Prompt

Our prompt was very straightforward. We clearly explained that we were attaching a spreadsheet and described the challenges involved. Specifically, we outlined the following:

1. Defined Expectations: We set clear expectations that for each worksheet, we needed to locate the answer column, identify corresponding questions, and generate detailed and informative answers based on those questions.
2. Procedure: We provided a step-by-step procedure to be followed. This included analyzing the columns for keywords indicating potential answer columns, extracting the questions, and populating the answer columns with detailed responses.
3. Escape Plan: We also defined an escape plan in case of any issues. This involved stopping the process after three unsuccessful attempts and explaining the precise reason for the problem, along with proposing alternative solutions.

I have a file containing multiple workbooks (e.g., Excel spreadsheets). 

For each workbook: 

Identify the Answer Column: Locate a column that likely contains answers to questions. Look for keywords like "Answer," "Solution," "Response," or similar. 

Extract Questions: Analyze the rows containing potential answers. Try to identify the corresponding questions within the same workbook. Questions might be in the previous row, a separate header row, or embedded within the answer itself. Generate Extensible 

Answers: Don't just copy short answers. Based on the context and potential questions, provide detailed and informative answers that are well-explained and potentially go beyond the core information. 

Please answer on the same document in the right column for the solution/answer that you might find for each row, do not show the questions or the answers here. 

Just include all in the document and make it downloadable. Ask me any additional question if you encounter any problem. Exit the process if you count 3 failed attempts to resolve the task.        

Other prompts

You can experiment with additional tasks to enhance the spreadsheet, such as adding dynamic columns. For example, generate a new column or update an existing one with external data from the internet, like the EUR to USD conversion rate.

Or update a column with external data from the internet, such as the EUR to USD conversion rate with today's conversion.

Pitfalls

Despite its impressive capabilities and usefulness, the system is still somewhat unstable. Sometimes, the code generation behind these adjustments crashes without disclosing the reason, making troubleshooting difficult.

We've also encountered instances where it claims the task is completed, but it only performs the action on one worksheet, or sometimes not at all. In these cases, you need to emphasize that it failed and ask it to redo the task.

Summary

Imagine sharing a spreadsheet with ChatGPT and assigning tasks like updating the USD currency conversion rate, swapping axes, reformatting columns, extending forecasts with new hypotheses, or adding a parallel business model with different assumptions for comparison.

This transforms ChatGPT into an "InternGPT," capable of performing diverse, complex tasks seamlessly. We hope that in a few days or weeks, direct access to Google Sheets and Microsoft Excel will be released to the general public, so we can evaluate deeper the smartness and task-reach.