A radical notion ( ... or not). AI governors

[CultureCitizen]

Automation has been shedding employments since the beginning of the industrial revolution. Slowly, as time passes, some new jobs are created.

When agriculture was automated , many workers migrated to cities to work in factories.
As factories and warehouses continue to be automated workers migrated to the service sector, it is in part thanks to this that we have more health and education workers and other ammenities and low productivity jobs ( restaurants, hotels, bars, amusement parks, concert halls, movies , theaters), and a plethora of office jobs ( accountants, managers, clerks ).

Services are notoriously hard to automate and accelerated. Sure it would be nice to have the paperwork done in an instant, it would also be nice to have automated public transportation. But , unlike industry many services simply can't be accelerated, because it would ruin the service itself.

- Imagine going to a restaurant to a fancy dinner with the time dropped down from 1 hour to 15 minutes.
- Or a surgeon telling you you will have your gallblader removal surgery done in 15 minutes tops... maybe the doctor should take a bit more time.
- Or going to a concert with the music played at 3x speed for productivity's sake.
Most services can't be rushed, they can only be made more efficiently. The big surprise is that AI may take the creative jobs first and the hard stuff later ( artists may get replaced before car drivers).

Now, there is an area that usually gets overlooked: the government. Think of it: an AI has no interest in money, it can't be bribed. If trained properly it will try to optimize the well-being of the citizens and completely ignore all the lobbying efforts from special interest groups. And importantly it would make its best to keep the peace between countries. The greatest hurdles are the training and getting an economic theory that actually works.

"The only function of economic forecasting is to make astrology look respectable"
-John Kenneth Galbraith

 

[bretbernhoft]

There is a good deal of educated speculation about what an AI-led world government would look like. Much of the forecasts, especially in science fiction, are powerfully negative. But, as you mention, it doesn't have to be that way. We can build an Artificial Intelligence that is capable of restoring our planet, peoples and destinies. Great topic.

 

[CultureCitizen]

There is a good deal of educated speculation about what an AI-led world government would look like. Much of the forecasts, especially in science fiction, are powerfully negative. But, as you mention, it doesn't have to be that way. We can build an Artificial Intelligence that is capable of restoring our planet, peoples and destinies. Great topic.

It is a matter of training ... and simulating.
Self-driving cars have to go through extensive simulation before being released as beta products.
As more data comes ( recorded through the driver's cameras) the simulations improve and more scenarios become available and more complex simulations can be done.

So the first hurdle is the equivalent of a recording device for the "needs" of people. These may be as diverse as food, water, sanitation, security, education, and peace. We need training data, because of course , releasing an AI model without tons of data is a call for disaster.

 

[Swank]

I don't think it has ever been demonstrated that AI does anything better than people do, aside from manipulate people. So far, self driving cars are not a success. AI has not made any breakthrough discoveries in chemistry or physics. It is still just an untested concept.


Iain Banks wrote many books about benevolent AI government.

 

[Wayne Mack]

an AI has no interest in money, it can't be bribed. If trained properly it will try to optimize the well-being of the citizens and completely ignore all the lobbying efforts from special interest groups. And importantly it would make its best to keep the peace between countries. The greatest hurdles are the training and getting an economic theory that actually works.

The training is really the crux of the matter. The challenge is that we, as humans, have no idea how to properly train computer systems or ourselves. With machine learning, one can only judge the quality of the training by the results post-training. If the results are not as desired, there is no obvious correction to apply; one can only shut down the old AI system and train a new one in some different manner.

Even though there may be no explicit algorithm telling an AI to favor some characteristic, the AI will learn what is favored by its trainers. While technically, the AI cannot be bribed, if it is trained that 'Likes' are a positive, someone can cause the AI to favor something by deliberately giving it a lot of likes. In the same way that the biases built into an algorithm can generate unanticipated results, the biases in the training provided to a machine learning AI will do the same.

AI can be useful and augment human activities. Like various approaches in the past, though, AI is far from being able to take on decision making capabiliteis.

 

[CultureCitizen]

The training is really the crux of the matter. The challenge is that we, as humans, have no idea how to properly train computer systems or ourselves. With machine learning, one can only judge the quality of the training by the results post-training. If the results are not as desired, there is no obvious correction to apply; one can only shut down the old AI system and train a new one in some different manner.

Even though there may be no explicit algorithm telling an AI to favor some characteristic, the AI will learn what is favored by its trainers. While technically, the AI cannot be bribed, if it is trained that 'Likes' are a positive, someone can cause the AI to favor something by deliberately giving it a lot of likes. In the same way that the biases built into an algorithm can generate unanticipated results, the biases in the training provided to a machine learning AI will do the same.

AI can be useful and augment human activities. Like various approaches in the past, though, AI is far from being able to take on decision making capabiliteis.

It can be rewarded by a simulation ( much like a game) using reinforcement learning.
The better the AI gets at the "game" the greater the reward.
The problem is then creating a realistic simulation of a society and of the economy. The economy is the greatest challenge.

 

[CultureCitizen]

I don't think it has ever been demonstrated that AI does anything better than people do, aside from manipulate people. So far, self driving cars are not a success. AI has not made any breakthrough discoveries in chemistry or physics. It is still just an untested concept.


Iain Banks wrote many books about benevolent AI government.

It is the undefeated champion in Chess since 2005. It can translate as well as an average human. It has discovered new ways to optimize circuits and perform matrix operations. It can produce images with a description far faster than a human ( what a graphical designer does in 3 hours the AI can do in a couple of minutes). And it can compose music ( alas, not as good as very talented musicians, but probably as good as an average compositor).

 

[Stephen Palmer]

See also my novel The Autist.

 

[paranoid marvin]

The problem I think is in understanding humans, which computers can't do; or at least have to have their parameters set by humans (which kind of defeats the whole object).

So computer AI may decide that banning alcohol and smoking improves the health of the population and reduces health costs. It may also decide that the normal work day should be 10 hours long, 6 days per week - which most adult males are capable of. Regardless of all the additional benefits to the state, neither of these suggestions would go down well with the vast majority of the population.

The problem is that once the AI has been given a strict set of rules as to what it can and cannot do, it makes it little more use than what a human could do.

 

[CultureCitizen]

The problem I think is in understanding humans, which computers can't do; or at least have to have their parameters set by humans (which kind of defeats the whole object).

So computer AI may decide that banning alcohol and smoking improves the health of the population and reduces health costs. It may also decide that the normal work day should be 10 hours long, 6 days per week - which most adult males are capable of. Regardless of all the additional benefits to the state, neither of these suggestions would go down well with the vast majority of the population.

The problem is that once the AI has been given a strict set of rules as to what it can and cannot do, it makes it little more use than what a human could do.

No I don't think computers should "decide" what humans like, that is an individual decision.
Computer could "suggest" the amount of alcohol certain individual is too much ( or tobacco ).
They might also forbid someone from driving when intoxicated , because it may cause harm to others.
Regarding work... well , let me tell you the news there is a permanent oversupply of labour in the world. There are literally not enough jobs for people ( the unemployment index is a joke), in my country half of the country is either self-employed or employed in the informal market ( plumbers, housekeepers, street markets, car washers, you name it, there is a plethora of informal jobs).
And precisely the problem is AI might trigger a new wave of unemployment, a small blessing is this time it will star with the white collar workers.

 

[Elckerlyc]

I think it's way too early for this discussion. AI isn't developed enough today to do the job.
Government requires not just being able to make decision on the basis of available data, it also requires wisdom, perspective, correlation, understanding of human nature and knowledge of a plethora of more specific sciences that concerns society.
AI is very good with solving mathematical problems. Like chess and logistics or designing circuits. Governing requires more than tat.

Regarding work... well , let me tell you the news there is a permanent oversupply of labour in the world. There are literally not enough jobs for people ( the unemployment index is a joke), in my country half of the country is either self-employed or employed in the informal market ( plumbers, housekeepers, street markets, car washers, you name it, there is a plethora of informal jobs).

Regarding work... well , let me tell you the news there is a permanent oversupply of labour in the world. There are literally not enough jobs for people ( the unemployment index is a joke), in my country half of the country is either self-employed or employed in the informal market ( plumbers, housekeepers, street markets, car washers, you name it, there is a plethora of informal jobs).

This may be the case where you live. In many western countries the situation is exactly the opposite. We cannot find the people to do the jobs; technicians, health care, logistics, judges, you name it is, there is a massive shortage of available employees. To create jobs is one thing (but often failed), to create workers may be even trickier. Your AI government might suggest to produce more offspring (as a long term solution.)

 

[Christine Wheelwright]

Self-driving cars have to go through extensive simulation before being released as beta products.

Somewhat off topic, but apparently this is not the case. It amazes me that somehow Musk is allowed to launch a half-baked "full self driving" feature (that moves a two tonne object through a public space at speed) and basically say 'Here, give it a try! Let us know what you think! Careful though, it will kill someone if you don't pay attention!". Not to mention charge people money for it. Absolutely unbelievable. For various reasons (not all related to FSD) that guy is going to spend the next 15 years in court (or worse). Rant over.

 

[CultureCitizen]

I think it's way too early for this discussion. AI isn't developed enough today to do the job.
Government requires not just being able to make decision on the basis of available data, it also requires wisdom, perspective, correlation, understanding of human nature and knowledge of a plethora of more specific sciences that concerns society.
AI is very good with solving mathematical problems. Like chess and logistics or designing circuits. Governing requires more than tat.

Regarding work... well , let me tell you the news there is a permanent oversupply of labour in the world. There are literally not enough jobs for people ( the unemployment index is a joke), in my country half of the country is either self-employed or employed in the informal market ( plumbers, housekeepers, street markets, car washers, you name it, there is a plethora of informal jobs).

This may be the case where you live. In many western countries the situation is exactly the opposite. We cannot find the people to do the jobs; technicians, health care, logistics, judges, you name it is, there is a massive shortage of available employees. To create jobs is one thing (but often failed), to create workers may be even trickier. Your AI government might suggest to produce more offspring (as a long term solution.)

Correct, AI is currently not developed enough to do this job, just as it isn't developed enough to do most jobs; currently it can only replace some call center tasks, part of the musicians work and part of a graphical designer's tasks.

"

Government requires not just being able to make decision on the basis of available data, it also requires wisdom, perspective, correlation, understanding of human nature and knowledge of a plethora of more specific sciences that concerns society."

Judging by how most government act, I have my qualms about most public servants meeting the above requirements.

his may be the case where you live. In many western countries the situation is exactly the opposite. We cannot find the people to do the jobs; technicians, health care, logistics, judges, you name it is, there is a massive shortage of available employees. To create jobs is one thing (but often failed), to create workers may be even trickier. Your AI government might suggest to produce more offspring (as a long term solution.)

Yes, yet "western" countries only represent 10% of the world population , so it's hardly a represantitive sample of the overall situation.

 

[Wayne Mack]

Automating decision making has been a long term desire from well before computer use (think written checklists). One of the issues with these approaches is that they become frozen in time. They are not affected by new information, new conditions, or even old data that was unavailable or overlooked at the time. Newer AI techniques have the additional issue of limited recall. While things like decision trees are implicitly constrained by prior decision along a path, AIs are not and this leads to an AI session providing inconsistent or contradictory responses during a session. It also allows the AI to go off on strange paths quite divorced from an initial query.

In the case of playing chess and similar applications, realize that the rules of chess, the pieces, the playing board, tec. are frozen. There is also a relatively good understanding of the strength of a particular position. An AI wins not so much by choosing the best possible move, but by better avoiding less good moves. Change the rules of the game, though, and the AI will not adapt. Despite the name, AI does not display the ability to adapt to changing circumstances; in this aspect, it is not intelligent.

 

[CultureCitizen]

Automating decision making has been a long term desire from well before computer use (think written checklists). One of the issues with these approaches is that they become frozen in time. They are not affected by new information, new conditions, or even old data that was unavailable or overlooked at the time. Newer AI techniques have the additional issue of limited recall. While things like decision trees are implicitly constrained by prior decision along a path, AIs are not and this leads to an AI session providing inconsistent or contradictory responses during a session. It also allows the AI to go off on strange paths quite divorced from an initial query.

In the case of playing chess and similar applications, realize that the rules of chess, the pieces, the playing board, tec. are frozen. There is also a relatively good understanding of the strength of a particular position. An AI wins not so much by choosing the best possible move, but by better avoiding less good moves. Change the rules of the game, though, and the AI will not adapt. Despite the name, AI does not display the ability to adapt to changing circumstances; in this aspect, it is not intelligent.

That sounds like the notions previous to the deep nets of how a "coded" program worked.
Nowdays AI systems learn ( albeit slowly) by throwing them tons of data.
There are two techniques : online learning and batch learning.
With online learning the system keeps training as new data comes in , it is not the most usual technique and has a lot of challenges ( making sure the net performs correctly and has actually learned).
The other method, batch learning trains the net and expects it to be able to generalize with the examples it was given.
Even with batch learning you can always re-train the full net with the new data.

 

[Wayne Mack]

Nowdays AI systems learn ( albeit slowly) by throwing them tons of data.

In addition to getting lots of data, the systems are given a correct answer. When the answer is more binary, the systems tend to do better. As the answers become more probabilistic, then the systems do more poorly. As the correct answers are provided externally and not by the computer system itself, I argue that it is not intelligent. Intelligence would be indicated by the machine making its own determination of what a correct answer should be.

 

[CultureCitizen]

In addition to getting lots of data, the systems are given a correct answer. When the answer is more binary, the systems tend to do better. As the answers become more probabilistic, then the systems do more poorly. As the correct answers are provided externally and not by the computer system itself, I argue that it is not intelligent. Intelligence would be indicated by the machine making its own determination of what a correct answer should be.

No , not really. Only binary classification problems are "binary".
You are describing supervised learning but there are other techniques: unsuperrvised learning which is mostly used to cluster data according to similar traits and reinforcement learning (and genetic algorithms) in which simulations are used to train the agent.
Therefore the first step is to have a kind-of-accurate simulation of the world ( natural resources + human activities).
The simulation will be used to train the AI.

 

[Christine Wheelwright]

In the case of playing chess and similar applications, realize that the rules of chess, the pieces, the playing board, tec. are frozen. There is also a relatively good understanding of the strength of a particular position. An AI wins not so much by choosing the best possible move, but by better avoiding less good moves. Change the rules of the game, though, and the AI will not adapt. Despite the name, AI does not display the ability to adapt to changing circumstances; in this aspect, it is not intelligent.

OK, so this is something I really want to understand. Chess software has been around for decades. I think even in the 70s there were pretty good examples that would beat most people. But nobody ever claimed they were examples of AI. The programs simply evaluated possible moves and graded them according to a set of criteria, and then selected the best. They probably had some standard openings programmed in. That sort of thing. Just a typical computer program really.

Now, a proper AI would surely be able to play Chess after being fed the rules of chess alone (nothing else). Maybe let it look at some example games too (a hundred, say). But don't define methods of evaluating moves for effectiveness (it has to work that out for itself). Can anyone tell me; are we there yet, or is that still decades away?

 

[CultureCitizen]

OK, so this is something I really want to understand. Chess software has been around for decades. I think even in the 70s there were pretty good examples that would beat most people. But nobody ever claimed they were examples of AI. The programs simply evaluated possible moves and graded them according to a set of criteria, and then selected the best. They probably had some standard openings programmed in. That sort of thing. Just a typical computer program really.

Now, a proper AI would surely be able to play Chess after being fed the rules of chess alone (nothing else). Maybe let it look at some example games too (a hundred, say). But don't define methods of evaluating moves for effectiveness (it has to work that out for itself). Can anyone tell me; are we there yet, or is that still decades away?

AI is a very broad field and there are several techniques.
Those early AI chess players were indeed the baby steps of AI . Games are the second echelon for solving real-world problems ( the first are toy problems which are even simpler.
The rules were coded in those programs ( valid moves, gain conditions,etc).

With reinforcement learning the computer learns by doing: randomly at first, it moves pieces, if it made an illegal move it gets punished, if it looses the game it gets punished, if it makes a legal move it gets rewarded, if it wins it gets an even better reward.

Al the knowledge is gains from experience gets stored in the net as a markov-chain.
The markov chain basically says : given the current state ( of the board) , what action can I take to get to a state with a high reward, so it takes that action.

Slowly after millions of games it learns the rules and how to be a master in the game.

 

[Vertigo]

As the correct answers are provided externally and not by the computer system itself, I argue that it is not intelligent. Intelligence would be indicated by the machine making its own determination of what a correct answer should be.

But that is exactly how we humans learn! We are taught the right and wrong answers. As we develop we do start learning how to determine the right answers ourselves but initially we must be taught in a much more binary way.