The format is worth taking. The strings are worth cutting.
The most important thing Google released this year is not an AI model. It is a folder of text files.
On June 12, Google quietly published the Open Knowledge Format. No new model, no software to install, no account. Just a convention: plain text files, a few labels at the top of each, and ordinary links between them. That is the whole idea.
It matters because of what it removes. Every company wants to feed its own knowledge to AI, but that knowledge sits trapped in a wiki here, a vendor's platform there, each speaking its own language. For years the answer has been RAG: pile all your content into a search system and let the AI pull the relevant pieces back for every question. It works, but it is heavy machinery, vector databases and the rest, and it rediscovers the same knowledge from scratch every time. Andrej Karpathy, a founding member of OpenAI, put the alternative simply. Instead of re-searching the same documents forever, keep a living, linked set of notes that compounds and gets sharper the more you use it. An open format is what lets that library outlive any single tool.
If the argument sounds familiar, it should. In March I wrote that the end of vendor lock-in begins with open file formats, the week Germany made the Open Document Format the binding standard for its public sector. That was about your documents. This is the same argument one level up, about your knowledge, the context your AI reasons over before it answers.
And here it gets interesting, because the format itself is genuinely neutral. No SDK, no account, tied to no cloud. However, Google does not ship it alone. The format arrives bundled with the company's own kit: an enrichment agent that reads straight from BigQuery, a graph visualizer, and a Knowledge Catalog updated to serve the format to Google's own agents. All optional, all Apache-licensed, and every one of them points home.
The promise underneath all of this is that your knowledge lives inside the format: active, cross-linked, growing. Which is why one line in the spec stopped me. OKF makes broken links legal. A link to something that does not exist is "not malformed," it says, it may just be knowledge not yet written. Generous, and also the one rule that cannot tell a note not yet written from one quietly deleted. A format built to hold living knowledge quietly permits it to rot. Keeping it alive is the harder part, and the part I have built into the projects I work on. I am happy to share how in a follow-up.
What began with who can open your documents now reaches the knowledge your AI runs on, and the lesson has not changed: you only control what you can read without asking permission. That is what makes an open format worth having, and it is why the kit Google packed beside it is worth refusing. Google conceded the first half in public. The second half is on you. Taking the open road is easy. Leaving the on-ramp is the actual test.
