Generative AI through large language models has made some remarkable progress across many areas. At their core, these models have learned from the vast amounts of human discourse, enabling us to summarize information, reason, create content, and perform many other tasks that once required significant human effort and collaboration.
Naturally, this raises an interesting question. Could these models have learned enough about the world of tabular data to help with processing it, without relying on traditional software systems and tools at all?
It is tempting for a couple of reasons. First, it could remove the need to fully understand your own data before seeking insights. You might be able to ask an LLM a question and get an answer by providing only a small portion of your data.
The idea is that with its vast prior knowledge, the large language model or LLM can often handle the task on its own. This approach is commonly called zero-shot or few-shot learning, where you share just a little data and rely on the model's prior knowledge to generate the answer for you. To be clear, here we are specifically talking about asking a question about tabular data.
Second, this could make tabular data far more accessible. Business analysts and decision makers could interact directly with generative AI tools in natural language to extract insights and answers from their data, without needing to write code or master complex software or complex user interfaces.
Since the second scenario can already be handled with an agent interacting with existing software systems and tools, we'll focus on the first question. Is it possible to use generative AI tools to handle data processing and modeling tasks without relying on those software tools, or even learning from your own data?
To explore this, we'll walk through an example, synthetic data generation, showing how this idea plays out in practice.
The bottom line is that large language models, the engines behind generative AI tools, are naturally not suited for these kind of tasks. They often struggle, sometimes producing what we call negative productivity. As the name suggests, these tools have primarily learned from human language, not tabular data, much of which remains locked away inside enterprises.