MAYU
A story of one company navigating generative AI adoption. Ten episodes, now being released.
Episode 1: AI Doesn’t Know the Answer
A proposal Yuna Sugimoto built with MAYU contained plausible-sounding misinformation. Rather than dismiss it as a simple misuse, Yui Takase repeats the same question to trace how the answer changes. Tatsuro Oba explains that generative AI isn’t retrieving truth — it’s predicting the next word from context.
Episode 2: Where Did That Word Come From
Without waiting for Yui, Yuna keeps testing the same question while varying phrasing and conversation length. When Yui joins and narrows the experiment to one variable at a time, the conditions under which the old terminology appears turn out to be inconsistent. Tatsuro Oba lays out that an answer's origin — trained knowledge, prompt instructions, conversation context, or an external reference — can't be determined from the text alone, so the two bring a formal request for a broader log review to Kanako Miyahara.
Episode 3: The Company’s Memory
With Kanako’s approval, Yui Takase and Yuna Sugimoto open MAYU’s retrieval logs. However Yuna rephrases her question, MAYU keeps landing on the same internal documents. Tatsuro Oba explains why: AI search doesn’t match identical words — it turns text into vectors of numbers and finds documents by closeness of meaning. That is embeddings, vector search, and RAG. But as Yui scans the index MAYU searches, she spots a single entry in an unfamiliar format that no part of the company uses today.