# Upstream Notes
> An AI research result can take months or years to become a setting in a public product.

## Entries
- [Expressive Speech Synthesis: From Research Preview to Consumer App](https://upstream-notes.pages.dev/expressive-speech-research-to-product/): Expressive speech synthesis can reach a public product before a research preview becomes widely available, but the appearance of a demonstration does not.
- [How to Read an AI Model Card: Seven Lines Product Scouts Should Prioritize](https://upstream-notes.pages.dev/read-ai-model-card-product-scout/): An AI model card is a starting document for estimating whether a model fits a product requirement, not a guarantee of deployment readiness.
- [Changelog Archaeology: Reading the Signs of Small-Group AI Feature Rollouts](https://upstream-notes.pages.dev/changelog-archaeology-ai-release-notes/): AI release notes are a useful record of product development, but their wording carries important distinctions.
- [From Benchmark Table to Product Feature: How to Judge the Distance](https://upstream-notes.pages.dev/benchmark-table-to-product-feature/): A strong benchmark result establishes that a model performs a particular task under specified evaluation conditions.
- [Why AI Research Demos Never Ship: A Field Guide to Productization Blockers](https://upstream-notes.pages.dev/why-ai-research-demos-never-ship/): An AI research demonstration proves that a capability can work under a particular set of conditions.
- [Character Consistency Across Photo, Voice and Video: How to Trace Research to Product](https://upstream-notes.pages.dev/character-consistency-photo-voice-video/): Consistent AI characters across photographs, voice, and video require more than a model that produces convincing individual outputs.

## About
- [About Upstream Notes](https://upstream-notes.pages.dev/about/): what the site covers and how entries are made

## Optional
- [RSS feed](https://upstream-notes.pages.dev/feed.xml)