Hello everyone, I’m Masahiro, an engineer working on SaaS products in Japan. Currently, I’m focusing on improving how products are designed, built, and delivered end-to-end, rather than just optimizing operations or feature output. Looking forward to exchanging perspectives with everyone.

I may still have gaps in my understanding, but I wanted to share where I currently feel the AI PM paradox. I’d really appreciate any feedback or perspectives from others who’ve experienced similar tensions. 【PRODUCT NAME】 YouTube Recommendation System 【WHAT IT DOES】 It predicts which videos a user is most likely to watch next based on behavioral data and surfaces them on the home feed and as recommendations. 【WHAT CAN’T BE FULLY SPEC’D】 What makes a “good recommendation” cannot be fully defined upfront. Even users watching the same genre may want different content depending on time of day, recent viewing behavior, mood, or context. Because these factors continuously change, fixed rules like “under these conditions, recommend this video” do not reliably work. 【WHY AI (NOT SOMETHING SIMPLER)】 Simple approaches such as genre classification or popularity-based ranking fail to capture contextual user intent. The real question is not “which videos are relevant,” but “which video this user is most likely to continue watching right now.” Answering this requires interpreting implicit behavioral signals—such as watch time, skips, and session flow—and making probabilistic predictions. This is fundamentally a prediction problem, not a deterministic logic problem. 【HOW IT HANDLES BEING WRONG】 When recommendations miss the mark, users can provide feedback like “Not interested” or “Don’t recommend this channel.” These signals are treated as learning inputs rather than failures and are used to gradually adjust future recommendations. 【ONE THING THE TRADITIONAL PM WOULD MISS】 Traditional PM approaches cannot keep up with how quickly user preferences change, especially in social and video platforms where trends shift rapidly. Locking down specifications upfront becomes a bottleneck. In AI-driven products, even incorrect recommendations produce learning signals. User reactions reveal shifting intent and allow the system to adapt over time. As a result, the PM’s role shifts from defining features to deciding which experiments to run and which signals to trust, enabling faster delivery of user-aligned value than specification-driven approaches.