4 ms·

Thanks. Really insightful. I've never really considered the idea of declarative AI personal assistant interactions. The concept is pretty mind-blowing actually

Thanks. Really insightful. I've never really considered the idea of declarative AI personal assistant interactions. The concept is pretty mind-blowing actually. It's pretty damn obvious that's the way to go. i.e. You tell the AI the state of the world you desire and it handles it given an information set much much larger then you as an individual could possibly ingest. Rather than the typical action of "AI bot send flowers to my wife" you'd say "AI bot make my wife know that I love and care for her." Of what the "system" knows about your wife's preferences, it can then indicate the action it believe will maximize the likelihood of adjusting the world into the asked-for state. That could be flowers, it could be a table for two at a restaurant she'd mentioned in an Email 6 months ago that you've forgotten about. Not "AI bot, order me an Uber to location x", rather "AI bot, get me to location x ASAP."

yomritoyj 10y ago
Why not just "AI bot, give me the best life possible"? To be less flippant, declarative AI already exists where it is easy. An airline booking website already gives you the ways to getting to location x ASAP. But looking at the performance of current recommendation systems, and natural language processing systems, I doubt if in the near future we would have a bot that would be better than me in guessing whether my wife would prefer flowers or a dinner on a particular anniversary, unless it cheats by using my wife's private information, in which case she's going to hate me either way.
- miguelrochefort 10y ago
  > unless it cheats by using my wife's private information, in which case she's going to hate me either way The should be no such thing as private information. The fact that you're saying that access to it is "cheating" seems to indicate that access to it is the best outcome on a productivity point of view. Why makes things more difficult in the name of privacy when transparency makes everything easier?
miguelrochefort 10y ago
Exactly. One of the big problem in communication and reasoning is that people often skip the "why" and jump straight to a final solution. They immediately come up with a final/complete plan such as "I need to get my car and drive to the nearest Home Depot to purchase a hammer so that I can nail this painting to my bedroom wall". A proper AI assistant would challenge this demand/plan. - Do you really need to use your car to get there? Can't you use your bicycle? Can you wait 15 minutes, your neighbor is also going to Home Depot? Can you wait for tomorrow, you'll drive right next to one in your way to work? - Do you really need to go to Home Depot? You could find a hammer at Walmart, which is closer. You could ask your neighbor for one, he has many and he won't use them today. Maybe there's a Uber fo tools that you could use to rent it? - Do you really need a hammer? Wouldn't some other heavy object do the job? What about this nailgun you received as a gift last year? - Do you really need to nail that painting? Wouldn't it be a better idea to use some of those 3M wall tapes that support up to 30 pounds and can be removed without leaving a mark/hole? - Do you really want to hang that painting in your bedroom? Wouldn't it be better in the living room? Should we ask "Stack Overflow for home decoration"? By locking ourselves into final solutions, we miss all alternatives which in most cases are much superior. By having tools that do exactly as we ask, we can't get what we really need. We need tools that challenge us, ask "why" until they find the actual end and not just the means. Fortunately, the why/how tree axis is first class in my communication platform. You need flour and water to bake bread. Why do you need flour? To bake bread. How will you bake bread? With flour and water. In the ideal world, you simply describe what you want in terms of the highest "why" you can think of, and the system then let you discover the different branches of "how" you can do it. The only reason we do things in a certain order is because of state dependencies. Each task (make dough) needs a valid starting state (access to water and flour) to be completed as the ending state (have dough). You can combine these tasks like you would dominos or functions (and apply all the usual concepts like map, lift, reduce, curry, monoids, monads). Basically, people don't know what they want and tend to copy others and/or their last selves. In order to avoid that and lead to the most efficient solutions, we need a communication system that challenges people until they really understand what and why they want. I can easily imagine a person's complex set of preferences and interests to be reduced to a handful of high-level values/goals from which an infinity of possible plans/solutions/paths can be derived.

yomritoyj 10y ago

Why not just "AI bot, give me the best life possible"? To be less flippant, declarative AI already exists where it is easy. An airline booking website already gives you the ways to getting to location x ASAP. But looking at the performance of current recommendation systems, and natural language processing systems, I doubt if in the near future we would have a bot that would be better than me in guessing whether my wife would prefer flowers or a dinner on a particular anniversary, unless it cheats by using my wife's private information, in which case she's going to hate me either way.

miguelrochefort 10y ago
> unless it cheats by using my wife's private information, in which case she's going to hate me either way The should be no such thing as private information. The fact that you're saying that access to it is "cheating" seems to indicate that access to it is the best outcome on a productivity point of view. Why makes things more difficult in the name of privacy when transparency makes everything easier?

miguelrochefort 10y ago

Exactly. One of the big problem in communication and reasoning is that people often skip the "why" and jump straight to a final solution. They immediately come up with a final/complete plan such as "I need to get my car and drive to the nearest Home Depot to purchase a hammer so that I can nail this painting to my bedroom wall". A proper AI assistant would challenge this demand/plan. - Do you really need to use your car to get there? Can't you use your bicycle? Can you wait 15 minutes, your neighbor is also going to Home Depot? Can you wait for tomorrow, you'll drive right next to one in your way to work? - Do you really need to go to Home Depot? You could find a hammer at Walmart, which is closer. You could ask your neighbor for one, he has many and he won't use them today. Maybe there's a Uber fo tools that you could use to rent it? - Do you really need a hammer? Wouldn't some other heavy object do the job? What about this nailgun you received as a gift last year? - Do you really need to nail that painting? Wouldn't it be a better idea to use some of those 3M wall tapes that support up to 30 pounds and can be removed without leaving a mark/hole? - Do you really want to hang that painting in your bedroom? Wouldn't it be better in the living room? Should we ask "Stack Overflow for home decoration"? By locking ourselves into final solutions, we miss all alternatives which in most cases are much superior. By having tools that do exactly as we ask, we can't get what we really need. We need tools that challenge us, ask "why" until they find the actual end and not just the means. Fortunately, the why/how tree axis is first class in my communication platform. You need flour and water to bake bread. Why do you need flour? To bake bread. How will you bake bread? With flour and water. In the ideal world, you simply describe what you want in terms of the highest "why" you can think of, and the system then let you discover the different branches of "how" you can do it. The only reason we do things in a certain order is because of state dependencies. Each task (make dough) needs a valid starting state (access to water and flour) to be completed as the ending state (have dough). You can combine these tasks like you would dominos or functions (and apply all the usual concepts like map, lift, reduce, curry, monoids, monads). Basically, people don't know what they want and tend to copy others and/or their last selves. In order to avoid that and lead to the most efficient solutions, we need a communication system that challenges people until they really understand what and why they want. I can easily imagine a person's complex set of preferences and interests to be reduced to a handful of high-level values/goals from which an infinity of possible plans/solutions/paths can be derived.