How AI finds the school email buried under 11 replies.
Old inbox search matched exact words — useless when the message you want says “permission slip due Friday” but you searched “field trip form.” New AI assistants match meaning. Fire a natural-language query into the 3D memory below and watch the right email light up.
Drag to rotate · nearer to your query = more relevant
Re: Re: Fwd: Class notice
“…just a reminder the permission slip needs to be signed and returned by Friday…”
How it works: embeddings
An AI model turns every email (and your question) into a list of numbers called an embedding — a point in a high-dimensional space where meaning is geometry. Texts about the same idea land near each other, even with zero shared words. “Permission slip due Friday” sits close to “when is the school thing due?” because the model learned they’re about the same real-world event.
Search then becomes distance: embed your question, find the nearest email points, return them ranked by closeness (usually cosine similarity). This is why you can ask in plain language and skip the exact subject line entirely. The 3D scatter above is a flattened cartoon of a space that really has hundreds of dimensions.
Old keyword search
Matches literal strings. “Field trip form” finds nothing if the email says “permission slip.” Fails on synonyms, typos, and rephrasing — exactly how humans actually write.
Semantic search
Matches concepts. Understands that slip = form, Friday = due date, class notice = school. Ranks by meaning, so the buried reply surfaces first.
The bigger idea: RAG and “the right problem”
Wrapping this in a chatbot gives you retrieval-augmented generation (RAG): the assistant embeds your question, retrieves the most relevant emails, then reads only those to write an answer — “Your daughter’s permission slip is due Friday; reply to Ms. Lee to confirm.” It’s grounded in your actual inbox, not the model’s imagination, which is what keeps it from making things up.
The tweet’s real point is about priorities. Summarising executive meetings was an easy, lucrative first target for AI. Digging a parent’s school email out of an 11-reply thread is unglamorous but life-shaped — and technically the same retrieval problem. Pointing capable models at ordinary domestic logistics, not just knowledge-work, is a genuine shift in who the tools are built for.