ChatGPT for data analysis

This is a question that has been bugging me for a long time. How can an LLM that never gives me the same answer to the same question, do data analysis - where I expect 100% accuracy?

Here is an example of what I mean when I say ChatGPT will not give you the same answer twice -

Directionally the same but not exactly the same.

This is a feature in Generative AI. It generates text every time and so by default it will be different each time. How GenAI works is that it is predicting the next word and based on probabilities, coming up with the sentence.

That being said, when I give it a table with data and ask it to analyze it, I want deterministic results. If I want to know how many Tables were sold in the East, there is only 1 answer.

So how does ChatGPT's data analyst CustomGPT claim to do data analysis?

Even MIT has a course on this - https://mitsloanedtech.mit.edu/ai/tools/data-analysis/how-to-use-chatgpts-advanced-data-analysis-feature/

I actually asked ChatGPT to answer this and I finally understand how this works.

Here is a step by step -

1. User Interface: The user inputs natural language queries into the CustomGPT.
2. Natural Language Understanding (NLU): The LLM interprets the queries.
3. Task Delegation: The LLM generates and sends code to deterministic tools (like pandas, numpy).
4. Execution Environment: The generated code runs and processes data.
5. Results Processing: The output is converted back into text by the LLM.
6. User Interface: The final results are displayed to the user.

So the actual analytics is done using code and this is the piece I was missing.

Actually this is a great example of GenAI as an agent, where we combine the LLM with Tools to solve real problems and this is the future of GenAI and how we will unlock practical use cases.

That being said, where errors can creep in in this process -

• Natural Language Understanding (NLU): The LLM interprets the queries. Again since this is probabilistic there can be mistakes / misinterpretation here.
• Results Processing: The output is converted back into text by the LLM. Again since this is probabilistic there can be mistakes / misinterpretation here.

Examples of mistakes in the NLU -

For instance, if a user inputs the query, "Show me the sales growth," the model might misinterpret this in several ways:

Temporal Ambiguity: The model might not understand the specific time frame the user is interested in. It might show sales growth for the last month instead of the last quarter or year, depending on what it guesses the user means.

Metric Ambiguity: The term "sales growth" might be interpreted in different ways, such as absolute sales numbers, percentage growth, or growth rate. If the model assumes one interpretation without clear context, it could provide incorrect or irrelevant data.

Contextual Ambiguity: If the user's query lacks context and the model doesn't have enough prior conversation history, it might not know which product line or geographical region's sales growth to display if the company operates in multiple domains.

Examples of mistakes in Results Processing

An example of an error in results processing can occur when the LLM incorrectly interprets or formats the output from the deterministic tool, leading to misleading or incorrect information being presented to the user.

1. Misinterpretation of Data:

- The LLM might incorrectly summarize the data if it doesn't properly calculate the overall average from the monthly averages.

2. Formatting Issues:

- Errors in formatting the output can lead to misunderstandings. For example, misplacing decimal points or mislabeling units (e.g., thousands vs. millions).

3. Inaccurate Aggregation:

- The LLM might incorrectly aggregate the results, such as summing values instead of averaging them.

4. Context Loss:

- If the context is not maintained accurately, the LLM might mix up different time periods or datasets, leading to incorrect results.

Bottomline - Understand all this before you pop that chart into ChatGPT and ask it to do data analysis.

BTW, even with GPT-4o, the image generation from text sucks. I asked it to generate an architecture diagram to show the flow and it came up with this.