5 ms·

Perhaps something like the Nest thermostat approach might be best: have something that learns your preferences based on time, day of the week, and other factors

Perhaps something like the Nest thermostat approach might be best: have something that learns your preferences based on time, day of the week, and other factors. Maybe it attempts to see what kind of mood you're in?

bondarchuk 6y ago
No, please, no. It will end up like youtube: spend one evening watching chess videos, and suddenly all your recommendations are chess videos. And afterwards on a meta level there is anxiety with everything you click on and how it might influence your filters.
- mncharity 6y ago
  Adding a feature to a system, like learning, may require additional user control to preserve empowerment. What if you had transparency and control over the interest vectors/graphs used, and management of the training set? A system can learn from reader actions, both about reader preferences, and about the content. Whether that's used for empowerment, or dystopia... hopefully we don't find that design choice is so forced, that the only way to avoid the latter is to neglect the former.
- paul_f 6y ago
  But I only want to read comments I agree with. :-)
- Isamu 6y ago
  > afterwards on a meta level there is anxiety with everything you click on and how it might influence your filters Yes I have so much hesitation where I am curious about someone’s fringe viewpoint but DO NOT want to be swamped with more of the same
mncharity 6y ago
Or a personal assistant? As there's potential for rich interactive refinement, as well as learned inference. As with Nest, there's an architectural privacy challenge. You'd like your pa to have insights like "oh, your interests on topic X align closely with other-person, so let's see, maybe you'd like thing too?". Ideally using local or trusted computes. But you don't in general want other-person's pa to know all your things. So perhaps an ecology of trusted visibility? Your pa can run parts of itself on a bonded site, with greater visibility than it would have locally, but with restrictions on what it can report back? But that's hairy and hard. So...?

bondarchuk 6y ago

No, please, no. It will end up like youtube: spend one evening watching chess videos, and suddenly all your recommendations are chess videos. And afterwards on a meta level there is anxiety with everything you click on and how it might influence your filters.

mncharity 6y ago
Adding a feature to a system, like learning, may require additional user control to preserve empowerment. What if you had transparency and control over the interest vectors/graphs used, and management of the training set? A system can learn from reader actions, both about reader preferences, and about the content. Whether that's used for empowerment, or dystopia... hopefully we don't find that design choice is so forced, that the only way to avoid the latter is to neglect the former.
paul_f 6y ago
But I only want to read comments I agree with. :-)
Isamu 6y ago
> afterwards on a meta level there is anxiety with everything you click on and how it might influence your filters Yes I have so much hesitation where I am curious about someone’s fringe viewpoint but DO NOT want to be swamped with more of the same

mncharity 6y ago

Or a personal assistant? As there's potential for rich interactive refinement, as well as learned inference. As with Nest, there's an architectural privacy challenge. You'd like your pa to have insights like "oh, your interests on topic X align closely with other-person, so let's see, maybe you'd like thing too?". Ideally using local or trusted computes. But you don't in general want other-person's pa to know all your things. So perhaps an ecology of trusted visibility? Your pa can run parts of itself on a bonded site, with greater visibility than it would have locally, but with restrictions on what it can report back? But that's hairy and hard. So...?