June 30, 2026
Spreadsheet wars just got weird
TabFM: A zero-shot foundation model for tabular data
Google says its new AI can read spreadsheets cold, and the comments are already fighting
TLDR: Google launched TabFM, a tool that promises to make predictions from spreadsheet-style business data without the usual long setup. Commenters were split between impressed and suspicious, with debate over Google’s scoring claims and side-eye about rival-company timing.
Google just rolled out TabFM, a new artificial intelligence tool meant to make predictions from ordinary business tables — think customer lists, fraud checks, and sales data — without the usual hours of tuning and setup. In plain English: Google is promising a world where you can toss in a spreadsheet-like dataset and get answers fast, no custom training marathon required. That alone was enough to get the community buzzing, but the real fireworks came from what people read between the lines.
The biggest reaction was a mix of “wow” and “hold on”. One camp sounded genuinely impressed, pointing out that earlier tools in this area were already strong and saying this could be a serious shake-up for one of the most common kinds of business AI. But the skeptics came in hot over the benchmarking, basically arguing that Google’s victory lap felt a little too tidy. Translation for non-experts: people aren’t just asking whether it works — they’re asking whether the way it was scored tells the full story.
Then came the gossip angle. Multiple commenters immediately connected the launch to SAP’s recent purchase of Prior Labs, with one person saying the timing of that sale now looks suspiciously perfect. And because no comment thread is complete without a chaos goblin, someone joked, “150,000 rows of data, where will I store it all?!” Suddenly, a serious product launch turned into a spicy mix of industry chess, trust issues, and spreadsheet humor. Classic internet.
Key Points
- •Google introduced TabFM as a zero-shot foundation model for tabular data integrated into BigQuery ML.
- •The article positions TabFM as a way to simplify tabular classification and regression by removing per-dataset model training, hyperparameter tuning, and complex feature engineering.
- •TabFM frames tabular prediction as an in-context learning problem, using historical training examples and target test rows together in a single prompt.
- •The model is designed for tables rather than one-dimensional text sequences, addressing the two-dimensional and order-invariant nature of tabular data.
- •Google says TabFM combines ideas from TabPFN and TabICL, including alternating row/column attention, row compression, and a Transformer-based in-context learning stage.