Nadella: Excel trains tiny GPTs

Spreadsheets are turning into model runtimes. That sounds absurd at first, but the basic point is real: Satya Nadella has been describing Excel not as a thin chat wrapper around a big model, but as a place where smaller, task-shaped AI systems can actually live and run. That matters because enterprise AI has a cost problem, a latency problem, and a data-governance problem. A giant remote model is powerful, but it is also expensive and blunt. ### What did Nadella actually say? The clearest version came in a late-2025 interview with Dwarkesh Patel and Dylan Patel, where Nadella talked about Excel as more than a UI for Copilot. He described Microsoft taking ideas and IP from the GPT family and pushing them into the “middle tier” of Office, so the product can work with the spreadsheet’s own structure instead of just chatting about it. That is the important distinction. (dwarkesh.com) A wrapper answers questions about a sheet. A native model participates in the sheet. ### Why is Excel a weirdly good place for this? Because Excel already is a programmable environment. Cells depend on other cells. Formulas recalculate. Tables carry structure. People encode business logic in grids every day, even when they do not think of themselves as programmers. Microsoft’s new `=COPILOT` function plugs directly into that calculation engine, so prompts can reference cells and ranges and update when the underlying data changes. Basically, Excel already has the scaffolding that AI systems need — inputs, state, dependencies, and outputs. (techcommunity.microsoft.com) ### Are “tiny GPTs” inside a spreadsheet literal? Probably not in the sci-fi sense people hear online. Nadella’s point is better understood as architectural, not as “every workbook secretly trains a frontier model.” The claim floating around is that Excel can express the core ingredients of transformer-style learning — embeddings, attention, stochastic gradient descent, next-token prediction — in spreadsheet form. That appears to come from demos and commentary around Excel’s computational flexibility, not from Microsoft saying ordinary users are training serious language models in cells at scale. So yes, the idea is real. But the hype version overstates it. (blockchain.news) ### Then what is the real product move? Microsoft is trying to make AI composable inside business software. The `=COPILOT` formula is the visible part. Underneath that, the strategy is to make Office apps into execution environments where models, agents, formulas, and enterprise data all interact in a tighter loop. Nadella has been arguing that agents will need tools, and Office remains one of the biggest tool surfaces in business. Excel matters here because it is where planning, forecasting, reconciliation, and exception handling already happen. (techcommunity.microsoft.com) ### Why do smaller models matter? Because most enterprise tasks do not need a giant general-purpose model every time. If the job is classifying comments, mapping categories, filling patterns, or reasoning over a known table, a narrower model can be cheaper and faster. It can also be easier to constrain. That is the appeal of “tiny” models — not that they beat top-end systems, but that they are good enough for repetitive spreadsheet work. This fits Microsoft’s broader push to embed AI across apps without making every interaction feel like a trip to a remote chatbot. ### What about privacy? Microsoft is leaning hard on that point. (techcommunity.microsoft.com) In its Excel documentation for the `=COPILOT` function, the company says data sent through the function is not used to train or improve the AI models and remains confidential for generating the requested output. That does not mean “everything runs locally on your laptop.” It means Microsoft knows enterprise buyers care about where spreadsheet data goes and how it is reused. (techcommunity.microsoft.com) ### What is the catch? The catch is that “inside Excel” can mean several different things — in-cell formulas, app-level orchestration, middle-tier services, or cloud-backed inference tightly bound to the workbook. Those are not the same. The viral version of the story collapses them together. Microsoft’s actual product language is more careful. ### Bottom line? The interesting part is not that Excel became a toy AI lab. (techcommunity.microsoft.com) It is that Microsoft wants spreadsheets to become a native surface for small, structured, business-specific intelligence. If that works, companies stop thinking about AI as one giant assistant and start treating it like another spreadsheet primitive — closer to a formula than a chatbot.