Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

You Can Always Get What You Want, But Not What You Need (2016) [video]

youtube.com

Post reply on HN
Print view View original

11–20 of 20 posts

Re: You Can Always Get What You Want, But Not What You Need (2016) [video]

#11

Post by gonehome »

truculation wrote: post #7

Fascinating. Instead of banning money or fixing prices (did they try that in Russia?) we retain free markets but now with a new layer of government intervention. Not intervention by social spending or regulation but by buying and selling according to the dictats of an AI. And the utility function of this AI is set according to what exactly? My guess is it would have to be according to our values, not to specific goal…

This is the goal alignment problem that MIRI, FHI, and some safety groups at Open AI are working on. It seems like a hard problem to get right (with bad consequences if not aligned super intelligence happens first).

Re: You Can Always Get What You Want, But Not What You Need (2016) [video]

#12

Post by zokier »

truculation wrote: post #7

Fascinating. Instead of banning money or fixing prices (did they try that in Russia?) we retain free markets but now with a new layer of government intervention. Not intervention by social spending or regulation but by buying and selling according to the dictats of an AI. And the utility function of this AI is set according to what exactly? My guess is it would have to be according to our values, not to specific goal…

> but now with a new layer of government intervention

As far as I can tell, the government is the least active actor in this area, so curious choice of picking it here.

Re: You Can Always Get What You Want, But Not What You Need (2016) [video]

#14

Post by debt »

Peter Norvig’s book A modern approach to AI is pretty cool. I don’t know what they use now in 101 AI courses but I remember that book being very complete.

I still believe though that hard AI will only ever be possible if we ever make serious breakthroughs in the material sciences.

Re: You Can Always Get What You Want, But Not What You Need (2016) [video]

#15

Post by UncleEntity »

truculation wrote: post #7

Fascinating. Instead of banning money or fixing prices (did they try that in Russia?) we retain free markets but now with a new layer of government intervention. Not intervention by social spending or regulation but by buying and selling according to the dictats of an AI. And the utility function of this AI is set according to what exactly? My guess is it would have to be according to our values, not to specific goal…

> Not intervention by social spending or regulation but by buying and selling according to the dictats of an AI.

Which would still run up against The Economic Calculation Problem which has constrained all central planning proposals and (probably) always will.

Once they solve that then we can talk about an AI which can predict the outcomes of billions upon billions of individual decisions which are made by market actors to fulfill their wants and needs.

Re: You Can Always Get What You Want, But Not What You Need (2016) [video]

#17

Post by xapata »

arbie wrote: post #10
xapata wrote: post #3

Earlier quoted context omitted.

I thought that theory told us multicriteria optimization is intractable?

If that's the same as multivariate opt, there are billion-dollar industries that rely on it every day.

No, unless I've mistaken the phrase, they're not the same. I'm thinking of the problem when competing criteria can't be satisfactorily combined as inputs to a function with a scalar output. This comes up in ethical paradoxes, for example.

Re: You Can Always Get What You Want, But Not What You Need (2016) [video]

#18

Post by dikkechill »

To differentiate between needs and (economic) wants, I found the "fundamental human needs" framework quite useful: https://en.wikipedia.org/wiki/Fundamental_human_needs (as developed by Chilean Economist Manfred Max-Neef).

Basically: Subsistence, Protection, Affection, Understanding, Participation, Leisure, Creation, Identity and Freedom.

Re: You Can Always Get What You Want, But Not What You Need (2016) [video]

#19

Post by rmateus »

xapata wrote: post #3
rmateus wrote: post #2

Multicriteria value decision analysis theory and methods are able to provide the solution to model the need (utility). For instance, in reinforcement learning the payoff (goal) function is always a given without any depth consideration about its construction. It can be modeled by using the methods referred previously.

I thought that theory told us multicriteria optimization is intractable?

Multicriteria optimization is distinct from multicriteria value decision analysis. The former does not model the value/utility of the criteria, while the latter does. If you do not model value/utility (what you need) then you cannot solve the underlying multicriteria problem.

Re: You Can Always Get What You Want, But Not What You Need (2016) [video]

#20

Post by rmateus »

xapata wrote: post #17
arbie wrote: post #10

Earlier quoted context omitted.

If that's the same as multivariate opt, there are billion-dollar industries that rely on it every day.

No, unless I've mistaken the phrase, they're not the same. I'm thinking of the problem when competing criteria can't be satisfactorily combined as inputs to a function with a scalar output. This comes up in ethical paradoxes, for example.

Making value trade-offs between competing criteria ultimately boils down to hard ethical choices (not exactly paradoxes)
Post reply on HN