Okay, so everyone is talking about large language models, right? ChatGPT, Claude, Gemini, all that stuff, generating text, images, really impressive. But there’s this whole other side of AI, this whole other data type, and the big, flashy LLMs, they just fall apart there. They really do.
We’re seeing it more and more, these large tabular models, they are just crushing LLMs when it comes to structured data, like spreadsheets, databases, all that stuff that businesses actually run on. It’s not the sexy, generative AI stuff, but it’s the backbone. A study from Google and the University of California, Berkeley, in 2023, it showed that even advanced LLMs like GPT-3.5, they had significant trouble with complex reasoning tasks on tabular data. They just couldn’t handle the relationships between columns, the numerical precision needed. It’s a different beast entirely.
Traditional machine learning models, like gradient boosting machines, decision trees, these things have been around for a while, they are still the champions for tabular data. A 2024 paper from Stanford and Microsoft researchers, it found that specialized tabular models, like XGBoost and CatBoost, consistently outperformed LLMs across various tabular datasets. And sometimes the performance difference was really big, like a lot of accuracy points. For example, on a credit default prediction dataset, a gradient boosting model achieved an F1 score of 0.82, and an LLM-based approach, it only got around 0.65. That’s a huge gap in performance, a really big difference for something critical like financial risk.
The problem for LLMs, it’s partly how they process information. They tokenize everything, they turn it into sequences of words or sub-words. But tabular data, it’s not sequential like text. It has columns, rows, relationships, numerical values that need to be treated as numbers, not just tokens. LLMs lack the inductive bias for structured data, they don’t inherently understand the relationships between different columns, the meaning of a specific number in a cell relative to other cells. They are built for language, and tabular data is not language. It’s just not.
And this isn’t some niche problem. Tabular data, it’s everywhere. It’s the vast majority of enterprise data. Estimates often put it at 70 to 80 percent of all data stored by businesses, all that critical stuff for operations, sales, finance, everything. So if your cutting-edge AI can’t handle the most common type of data, what does that mean for its real-world applicability? It means you need both. You need the LLMs for the text and the creative stuff, and you need these specialized tabular models for the hard, structured numbers. It’s not one or the other, it’s both working together.
I mean, are we really going to trust an LLM to manage a supply chain or optimize a financial portfolio when a simpler, purpose-built model can do it with much higher accuracy? It just doesn’t make sense. The industry needs to understand this distinction, and it needs to invest in both areas. The hype around LLMs is massive, and it’s deserved for what they do well, but it’s creating this blind spot for where they just don’t perform.
I bought some NVDA stock, NVIDIA, on July 1, 2024, at $123.08 a share. I’m holding that until it hits $200, or if it drops below $100, then I’m out. That’s my exit strategy. The chip market is still strong, but you gotta watch these things.
The focus on LLMs, it’s almost like people forgot about the foundational work in machine learning for structured data. It’s still there, it’s still incredibly powerful, and it’s still what’s driving a lot of real-world business value. We need to remember that. It’s not always about the flashiest new thing, sometimes it’s about the tools that just get the job done, reliably and accurately. And for tabular data, those are still the specialized models. It’s a clear distinction, and it’s important for anyone building AI solutions, or even just investing in AI companies. You have to know what you’re buying into, what problems these models actually solve.