A road trip. A reward function.
A live reinforcement-learning experiment: combine helpfulness, safety, and truthfulness into one reward function.
Watch on YouTube
1:18:12
Watch on YouTube
1:18:12
The filters needed work.
I built a login.
A face-filter app, some coffees with strangers, and the feedback I didn’t act on. Notes from a project that failed.
Read the original retrospectiveThe warning was in the conversations.
I took paper mockups to a local coffee shop and offered to buy people a coffee for a few minutes of feedback. Those conversations raised problems with the idea. I kept building.
More features were not the answer.
Later, time went into login, password reset, and features around the app. The quality and speed of the filters needed that attention instead. They were the core product—and a recurring point of feedback.
The 2016 retrospective separates the failure into design, coding, and distribution. The Maskito source code was opened up alongside it.
More from the notebook
- Thirty apps in thirty daysA 2015 retrospective
- JonesAudio: from voice to cursorA macOS tool