1,638,093 people studied.
Each circle is one study. Area is its sample.
Pick any study for its result and paper.
1 dot = 40 people, at least one dot per study. Cohorts may overlap.
Not counted: · ·
What we’re studying.
These are plans. We’ll post results here as they come in.
When to speak, and when to stay quiet.
Protocol in design
- Why it's open
- In HeartSteps, prompts timed by context worked on day one and faded within a month. Nobody has tested prompts timed by what a person is visibly doing right then.
- How we'll test it
- A micro-randomized pilot. Each time HABIT notices a target moment, like an hour in a chair or a phone in bed after 23:00, it randomly either speaks or stays quiet. The quiet moments are the comparison.
- What we'll measure
- What the wearer does in the next 30 minutes, and whether that effect is still there in week six.
- What would count
- An effect that holds past the first month, at a rate of prompts people don't turn off.
HABIT speaksHABIT stays quiet
- 08:12, HABIT stays quiet
- 09:42, An hour in a chair, HABIT speaks
- 11:24, HABIT stays quiet
- 13:06, HABIT speaks
- 14:36, HABIT speaks
- 16:12, Another hour sitting, HABIT stays quiet
- 18:30, HABIT stays quiet
- 20:18, HABIT speaks
- 23:18, Phone in bed, HABIT speaks
How the pilot works, for one wearer on one day. Each moment is a coin flip, so the quiet moments show what would have happened anyway. Measuring change that lasts.
Planned, after the first pilot
- Why it's open
- Only 8% of the gym programs in the megastudy kept working after they ended, and new habits took a median of 66 days to settle. Most studies stop long before that.
- How we'll test it
- Follow pilot wearers for at least 12 weeks, measured against their own first two weeks. Then switch prompts off for a stretch and see what stays.
- What we'll measure
- Weekly counts of the behavior each person picked: hours sitting, screen time in bed, gym visits, meals.
- What would count
- Change that is still there with the prompts off.
Seeing everyday actions from a camera on your face.
Prototype hardware in development
- Why it's open
- Ego4D shows models can learn from eye level, using large datasets and large computers. Whether a small, low-power camera in an ordinary frame can do it all day is still open.
- How we'll test it
- Record prototype footage, with everyone's consent, of a short list of actions: picking up a phone, sitting down, eating, a set at the gym, lights out. Label it, then test models small enough to run on the glasses.
- What we'll measure
- Precision and recall for each action, false alarms per day, and hours of battery.
- What would count
- Accurate enough that wearers trust it, with few enough false alarms that nobody takes the glasses off.