4 ms·

Executive level meaning C suite and it's political and military equivalent The idea that an AI is going to decide we should first-strike another nation and top

Executive level meaning C suite and it's political and military equivalent The idea that an AI is going to decide we should first-strike another nation and top brass say "well the AI has never been wrong so let's do it" is laughable.

ajuc 5y ago
This is the Hollywood version of problems with AI. Real life equivalent: "AI changed the margin on humanitarian aid in country X by 28%", the end result would be war as well but the decision wouldn't be "war or peace" it would be "a little more cost-savings vs a little higher risk of problems". And it's not sci-fi, see 2008 crisis and its consequences. Bad model that gives short-term profits is almost always preferred over good model that reduces profits. Or, to put it in Gulag/private prisons terms: "on the advice of AI we decreased calory intake of our prisoners by 30%" sounds much better than "kill prisoners by overworking and strvation".
- helen___keller 5y ago
  > Or, to put it in Gulag/private prisons terms: "on the advice of AI we decreased calory intake of our prisoners by 30%" sounds much better than "kill prisoners by overworking and strvation". The executive decision made by a human "we should optimize prisoner calorie intake" is separate from the AI advice "30% is the amount of calorie reduction that prisoners can withstand" Right now AIs are good for optimizing and calculating things that we specify they should optimize and calculate. If we develop an AI that can independently decide to optimize human calorie intake without anyone doing work to make it an option, then we should be worried. Right now that is pure fiction. Pre-AI equivalent: These prisoners seem like they're eating too much, so let's cut their food portions. One third reduction seems reasonable.
  - ajuc 5y ago
    99.9% of decisions in real world isn't 0% or 100% but somewhere in between. The limit is there, the question is how big. The gulag example is a good one because with a different limit it would change from system of extermination into a civilized prison and it doesn't require any particular person to make this decision taking the full responsibility. It's just a change of some parameter by a few percents. > Pre-AI equivalent: These prisoners seem like they're eating too much, so let's cut their food portions. One third reduction seems reasonable. Bureaucracy and corporations are a type of AI, and they routinely make objectively evil decisions and people tend to follow. What computerized AI introduces is even stronger network effects, even easier monopolization, and even smaller accountability.

ajuc 5y ago

This is the Hollywood version of problems with AI. Real life equivalent: "AI changed the margin on humanitarian aid in country X by 28%", the end result would be war as well but the decision wouldn't be "war or peace" it would be "a little more cost-savings vs a little higher risk of problems". And it's not sci-fi, see 2008 crisis and its consequences. Bad model that gives short-term profits is almost always preferred over good model that reduces profits. Or, to put it in Gulag/private prisons terms: "on the advice of AI we decreased calory intake of our prisoners by 30%" sounds much better than "kill prisoners by overworking and strvation".

helen___keller 5y ago
> Or, to put it in Gulag/private prisons terms: "on the advice of AI we decreased calory intake of our prisoners by 30%" sounds much better than "kill prisoners by overworking and strvation". The executive decision made by a human "we should optimize prisoner calorie intake" is separate from the AI advice "30% is the amount of calorie reduction that prisoners can withstand" Right now AIs are good for optimizing and calculating things that we specify they should optimize and calculate. If we develop an AI that can independently decide to optimize human calorie intake without anyone doing work to make it an option, then we should be worried. Right now that is pure fiction. Pre-AI equivalent: These prisoners seem like they're eating too much, so let's cut their food portions. One third reduction seems reasonable.
- ajuc 5y ago
  99.9% of decisions in real world isn't 0% or 100% but somewhere in between. The limit is there, the question is how big. The gulag example is a good one because with a different limit it would change from system of extermination into a civilized prison and it doesn't require any particular person to make this decision taking the full responsibility. It's just a change of some parameter by a few percents. > Pre-AI equivalent: These prisoners seem like they're eating too much, so let's cut their food portions. One third reduction seems reasonable. Bureaucracy and corporations are a type of AI, and they routinely make objectively evil decisions and people tend to follow. What computerized AI introduces is even stronger network effects, even easier monopolization, and even smaller accountability.