Gemini in @GoogleWorkspace just got a lot more helpful. To do this, we had to tailor the model to each specific product and their user needs.
Here's a little behind-the-scenes on how that came to life in @GoogleDocs, Sheets, Slides, and @GoogleDrive:
While most AI writing tools operate on plain text structured with Markdown, we’re teaching Gemini the underlying data model of @GoogleDocs. By understanding Docs natively, Gemini generates and edits documents with full-fidelity formatting, fonts, and styles. That’s why when you ask Gemini to "match the format," it duplicates the exact structure— even clearing out the old text so you get a high quality, ready-to-use template instantly.
Most AI presentation tools rely on their own rigid templates and structured styles, but ask for something custom and they break. With Gemini in Google Slides, we use a custom translation between Gemini and Slides that fully leverages Gemini’s visual creative abilities while creating editable slides. For example, with our new “enhance” feature, Gemini will be able to up-level the look and layout of your slide while matching the style and structure of your deck.
While we don't have favorites, the evolution of Gemini in Google Sheets might be our most impressive yet. Gemini in Google Sheets has achieved a state-of-the-art benchmark, achieving a 70.48% success rate on the full SpreadsheetBench dataset. This performance not only exceeds competitors but nears human expert ability.
We accomplished this by equipping Gemini with better verbalization and enhanced coding capabilities. With these, Gemini can now natively build complex models and dashboards, solve your most complex optimization problems, and verify its own work for expert-level precision.
Read more about Sheets here:
The formatting fidelity problem in document generation is attacked by targeting an app's native data model instead of Markdown, and the 70.48% SpreadsheetBench figure gives a public number to compare spreadsheet agents against.
Read the full thread on X ↗
