Four AI Capabilities That Crossed the Threshold in Six Months

Executive overview

AI capabilities cross a usability threshold unpredictably — tasks that failed months ago may now work out of the box. Missing that moment means competitors gain a lead while you play catch-up. The fix is a systematic retesting process tied to model releases and quarterly reviews.

The competitive edge is not adopting AI early — it is knowing when a failed test is worth retrying.

Three capabilities that crossed the threshold

• Complex PDF extraction: a project that took five weeks of engineering and six models six months ago now works in one prompt with Gemini Flash.
• Image consistency: generating images with a consistent face, product, or logo across outputs was nearly impossible; current models handle it accurately.
• Handwriting extraction: accuracy on cursive and poor handwriting has jumped from ~60% to ~100% with models like Gemini Pro.

Why tests fail for the wrong reason

• Vague prompts and missing context make AI appear incapable when it is not.
• Before writing off a capability, confirm the prompt has: specific task context, clear expectations, and concrete success criteria.
• Only after a high-quality prompt still fails can you judge the model's actual limit.

The AI wish list system

• Keep a running list of tasks AI cannot yet do: record the task, the date tested, and the result.
• Store it anywhere you will actually use — Google Doc, Apple Notes, or similar.
• This list is a retesting queue, not a graveyard.

When to retest

• New model release from Anthropic, OpenAI, or Google — test relevant wish-list items immediately.
• New feature release from any of the big three — test if it addresses your use case.
• Quarterly calendar reminder — review all items even without a major announcement, as models improve silently in the background.

Retesting protocol

• Run the wish-list task with a basic prompt against the new model or feature.
• If it works, remove it from the list.
• If it fails, improve the prompt: add context, sharpen expectations, check current best practices.
• Only if it still fails with a strong prompt does the item go back on the list with an updated date.