LLMs can do amazing things like writing, summarizing, and processing data in an unprecedented way. But how well can it detect patterns in our everyday life?
In simple terms: can foundation models like GPT detect patterns in unknown data (for example, housing prices or sales figures) without additional training or fine-tuning?
Could it, for instance, tell a salesperson which products should be sold less and which should be sold more based on the data?
There are lots of great tools and techniques for training models to predict data, and some can cluster your data. But you need to know how.
Wouldn't it be great if the LLM that everyone already uses could help with that, at least to get a glimpse of hidden patterns you might never have considered?
We did a test with three popular foundation models: OpenAI o3-mini, Claude Sonnet 3.5, and Gemini.
And here is what I did:
- I generated an algorithm based on a set of “hidden” pattern (basically just an algorithm) so we can easily judge the analysis results using the code as the source of truth.
- I used a typical example for housing prices (but with some hidden patterns, as you’ll see).
- I generated a dataset with this code.
- I asked the foundation models (GPT, Claude, Gemini) to find hidden patterns without any further instructions.
- I took the response and asked a separate judge to check the correctness of the results, based on the algorithm defined earlier.
Disclaimer: For large datasets and in enterprise settings, you would likely use special models and training algorithms. This experiment is mainly for exploration, not for professional use.
Having said that, let’s dive in.
1. Step #1: Generate the Code
Here are the patterns we have built into the generated datasets for 1000 entries of house prices.
airabbit.blog
Top comments (0)