Descriptions:
Dylan Davis documents four AI capabilities that have recently crossed from unreliable to production-ready, using a concrete construction industry project as the primary reference point. Six months ago, extracting structured data from complex manufacturer PDFs — containing handwriting, irregular tables, and embedded images — required five weeks of engineering work and six different AI models to reach 99% accuracy. Running the same task against Gemini 2.0 Flash with a basic prompt produced perfect output on the first attempt.
Three additional capabilities are covered with before/after comparisons: image consistency (placing a product in a specific person’s hand with accurate facial features and clothing details, demonstrated using Nanobanana versus Nanobanana Pro), handwriting extraction (accuracy climbing from roughly 60% to near 100% with Gemini 2.0 Pro, which now understands character relationships and spatial layout rather than just isolated characters), and general multimodal document understanding across complex layouts.
Davis also introduces a systematic method for tracking capability improvements: an “AI wish list” — a maintained backlog of previously failed use cases that gets periodically retested as new models ship. He distinguishes between genuinely unsupported capabilities and cases where poor prompting is the real bottleneck, arguing that prompt quality, context specificity, and clear expectations must all be in place before concluding a model cannot perform a given task.
📺 Source: Dylan Davis · Published February 12, 2026
🏷️ Format: Review
