W19 - NotebookLM, DeepWiki, and AI Products

Recently I've been using information-organizing tools like NotebookLM and DeepWiki. NotebookLM is designed around the concept of a workbench and carries an IDE-like feel. Moving this to an "at work" scenario, if a workbench could freely integrate internal content sources such as dx, ones, code, and km, I believe productivity could multiply. Having studied the OODA loop, AI effectively accelerates that cycle, especially the Observe and Orient stages. It helps us extract and understand key information faster and build cognitive frameworks, enabling more frequent strategic adjustments and decisions. For tasks like writing a COE, reviewers only need to focus on root causes and lessons learned—the timeline and similar details would be auto-populated—so anyone can drill into an incident at any time; that is true case study work.

I recommend an episode of Lenny's Podcast featuring OpenAI's CPO—great product-perspective insights. If you want to save time, you can use NotebookLM. Three points left a strong impression on me.

1. The technology stack underpinning products is shifting. It is moving from the old rule-based “deterministic input—deterministic output” systems to today's model-based “probabilistic outputs.” Models are iterating very quickly—every couple of months machines can do things they couldn't before—so the model you use today will likely be the worst you ever use in the future. This requires constantly adjusting product thinking and cadence.

2. Evals are becoming a core competency for product managers. As Yao Shunyu described in the “LLM second half” perspective, it's not just about testing models but about evaluating how the entire technical system performs on specific problems. Product teams need to build and adapt models around real-world problems rather than waiting for models to be "perfect."

3. AI will not generate a single ultimate creative idea with one click; rather, it accelerates the process of bringing ideas to life. It can quickly provide multiple options for creators to test, combine, and refine. The real productivity gain comes from this iterative "human and AI co-creation" loop.