How I think through hard problems
Two end-to-end case studies — from the first signal something is wrong, through research, synthesis, and prototyping, to what actually changed.
The model was accurate.
The scientists stopped using it anyway.
Two months after launch, a generative molecule design tool was losing 60% of its weekly active researchers. The model's accuracy wasn't the problem. Scientists had no way to decide whether to trust a result — so they went back to PubMed.
Usage logs showed the drop wasn't random. It was consistent across all six research teams and accelerating. Three teams had stopped creating tool-related tickets entirely — no complaints filed, just absent. Leadership read this as a model accuracy problem. The ML engineers disagreed. I pulled the session logs to look before forming an opinion.
Week-8 tool usage as % of week-1 baseline, by team
I ran 9 contextual inquiry sessions — watching researchers do real tasks, not demos. Followed by a one-week diary study where 6 researchers logged every lookup they performed, whether they used the AI tool or went to PubMed, and why. 180 data points total.
"I check it, then I go to PubMed to make sure. At that point it's faster to just start with PubMed."
Riya, Senior Medicinal Chemist — diary study"My PI asked me to cite where that 87% confidence came from. I had no answer. That was the last time I used it for a formal recommendation."
Arjun, Computational Chemist — interviewGoal: verify that a hypothesis-first entry and an explicit confidence breakdown would be understood without explanation. Neither the workflow inversion nor the confidence labelling had been used before — both needed validation before engineering time.
Key finding: all 4 participants immediately understood the confidence distinction without prompting. "High data density" vs. "extrapolated" removed the most common question — "what does this number mean?" — entirely. Two participants said they would now use the result in a report.
How YouTube could add 30 minutes of meaningful watch time per user per week
YouTube's recommendation engine is extraordinary at getting people to keep watching. But "more time on platform" and "watch time that users feel good about" have started to diverge. This is a product exercise in closing that gap.
YouTube's recommendation system optimises for watch time and click-through rate. Both are reasonable proxies for interest. Neither captures whether a user felt the time was well spent. The gap between "total watch time" and "satisfied watch time" is the core problem.
The drop-off in satisfaction isn't random — it clusters around three failure modes: the session has no natural endpoint, recommendations optimise for adjacent click-through rather than depth, and the home feed doesn't adapt to session intent. All three are solvable at the product layer without changing the underlying model.
% of dissatisfied sessions citing each cause (illustrative survey, n=800)
"I came to watch one video about cooking. An hour later I was watching car review videos. I don't even drive."
User interview, Priya — student cohortThe concept: a session intent prompt at open (one tap, optional) shifts the recommendation mode for that session. "Go deep" surfaces same-topic, same-creator recommendations. "Explore" keeps current behaviour. "Wind down" surfaces shorter, lighter content and shows a session wrap at a natural break point.
The session wrap nudge ("good stopping point") is shown only once per session and only at a natural end — last video in a series, or after a topic cluster completes. It doesn't enforce a limit. It gives the stopping point that most sessions currently lack.