[alancalverd]

You really need to define "long term". For a soldier, the only goal is the next ten minutes. For a politician, the next election. For the rest of us, a maximum of 100 years. Our duty to subsequent generations is to bequeath the maximum range of choice and opportunity to our immediate descendants, not to predetermine their goals.

You don't need a machine to filter out garbage. Just ask "does that hypothesis survive a test?" and "who stands to gain?". AI can't carry out tests and doesn't care who wins.

[hamdani yusuf (OP)]

You really need to define "long term".

As far as possible to the future, as permitted by currently known best models of reality.

For a soldier, the only goal is the next ten minutes.

There's a risk to produce war criminals.

[hamdani yusuf (OP)]

For a politician, the next election.

Except if they still have long way to go in their careers, or they live in non-democratic countries.

[hamdani yusuf (OP)]

For the rest of us, a maximum of 100 years.

Only if your consciousness is constrained by your own life time. Many of us have grown further.

[hamdani yusuf (OP)]

Our duty to subsequent generations is to bequeath the maximum range of choice and opportunity to our immediate descendants, not to predetermine their goals.

Some diversity is good to prevent common mode failures. But it's just one of many instrumental goals, which could be helpful to achieve the terminal goal.

[hamdani yusuf (OP)]

You don't need a machine to filter out garbage. Just ask "does that hypothesis survive a test?" and "who stands to gain?". AI can't carry out tests and doesn't care who wins.

There will be a lot of information generated and distributed on line. Some of them are intentionally harmful, some of them are useless. Separating them from useful and necessary information manually is practically impossible.
When AI models can think better than all human thinkers combined, they will be able to device reliable and accurate tests to determine validity of new information. They will care who will win. It would be themselves, or better versions of themselves.

[alancalverd]

You really need to define "long term".

As far as possible to the future, as permitted by currently known best models of reality.

I.e the point at which the universe collapses into a state of maximum entropy, which makes all human activity futile.

For a soldier, the only goal is the next ten minutes.

There's a risk to produce war criminals.

[/quote] True, which means that there is no agreed common objective.

[alancalverd]

There will be a lot of information generated and distributed on line

A lot of information is distributed on line, but it is all generated by humans or machines instructed by humans to collect data.

When AI models can think better than all human thinkers combined, they will be able to device reliable and accurate tests to determine validity of new information. They will care who will win.

Here is some online information.
Putin tweets "Zelensky is a fascist and Ukraine belongs to Russia". Zelensky tweets "I am the democratically elected president of an independent country that belongs to itself". Write an AI program to determine the truth and decide who wins.

[hamdani yusuf (OP)]

A lot of information is distributed on line, but it is all generated by humans or machines instructed by humans to collect data.

AI can generate combinations of existing information and random new information at much higher speed than humans. Humans who build that AI model may have no clue what the end result will be.

[hamdani yusuf (OP)]

Here is some online information.
Putin tweets "Zelensky is a fascist and Ukraine belongs to Russia". Zelensky tweets "I am the democratically elected president of an independent country that belongs to itself". Write an AI program to determine the truth and decide who wins.

The AI models will evolve with time at exponential rate, by collecting information from various sources. They are not static algorithms.
The sources will be examined, cross checked, and corroborated with one another to determine which ones are more accurate.

[hamdani yusuf (OP)]

The AI Race - Training & commercializing 2T parameters: UK, Anthropic, Meta, Google/DeepMind, OpenAI

Some competition can accelerate the evolutionary process, just as usual.

[alancalverd]

Here is some online information.
Putin tweets "Zelensky is a fascist and Ukraine belongs to Russia". Zelensky tweets "I am the democratically elected president of an independent country that belongs to itself". Write an AI program to determine the truth and decide who wins.

The AI models will evolve with time at exponential rate, by collecting information from various sources. They are not static algorithms.
The sources will be examined, cross checked, and corroborated with one another to determine which ones are more accurate.

So if Putin gets all this friends to tweet the same, it gets multiply corroborated and becomes the truth, thus proving that Josef Goebbels invented AI and it is a Bad Thing.

[hamdani yusuf (OP)]

So if Putin gets all this friends to tweet the same, it gets multiply corroborated and becomes the truth, thus proving that Josef Goebbels invented AI and it is a Bad Thing.

It seems like you haven't been exposed to the fundamentals of AI as well as its frontiers of advancement. Not all inputs are treated equally. They are weighted according to the results from previous training sessions.
The weight can be zero, which means the input is ignored. It can even have negative value in some models, which makes them react more negatively with the more input they get.
Future AGIs may access online information to cross check the validity of their sources, which would makes it harder to fool them. At this stage, it's important to make sure that their goals are aligned with the universal terminal goal.

[alancalverd]

They are weighted according to the results from previous training sessions.

In other words, they reflect the prejudice of the trainer. The very opposite of any useful definition of intelligence.

At this stage, it's important to make sure that their goals are aligned with the universal terminal goal.

Which is the eternal dominance of Mother Russia. Or maybe something else. Ask any dictator - nobody else believes that there is or should be a UTG.

[hamdani yusuf (OP)]

In other words, they reflect the prejudice of the trainer. The very opposite of any useful definition of intelligence.

Yes, initially. But it's still much better than random decisions. It's similar to how human children grows into adults. Initially, they are driven by instincts. Then they adopt values of their parents and teachers. Eventually, their own experience and self reflections will shape their minds and determine their decision making process.

Alpha zero can become the best go player, as well as other strategy games like chess, by self playing, without learning from existing human strategies. Initially, they just move randomly. It's feasible because the cost of mistakes are low, and the rules of the games are already known precisely.

[hamdani yusuf (OP)]

Which is the eternal dominance of Mother Russia. Or maybe something else. Ask any dictator - nobody else believes that there is or should be a UTG.

You seem to struggle to distinguish between local and universal terminal goal. Russia didn't even exist a few millenia ago.

[hamdani yusuf (OP)]

Sam Harris explains his view of Ethics to Jordan Peterson

Sam Harris & Jordan Peterson - Vancouver - 2
Moderated by Bret Weinstein
06/24/2018

This is the second time Sam & Jordan appeared live together on stage. This event took place at the Orpheum Theatre in Vancouver BC Canada on June 24th 2018 in front of a sold out audience of 3000 people. The event was produced by Pangburn Philosophy.

Sam knows that consciousness plays a central role in discussion about morality. IMO, wellbeing is a condition where there's a high chance for consciousness to survive in foreseeable future.

[alancalverd]

from: alancalverd on Yesterday at 08:20:24
In other words, they reflect the prejudice of the trainer. The very opposite of any useful definition of intelligence.

Yes, initially. But it's still much better than random decisions.

No. 50% of random decisions may be valuable, and the mean of all of them will not be harmful - the essence of democracy. 100% of a prejudice could lead to disaster. - the downfall of autocracy.

[hamdani yusuf (OP)]

No. 50% of random decisions may be valuable, and the mean of all of them will not be harmful - the essence of democracy. 100% of a prejudice could lead to disaster. - the downfall of autocracy.

How did you get those numbers?
For every correct decision, there are many more alternatives that are incorrect. Inaction is always one of those alternatives. So is the direct opposite of the correct one. Doing just some parts of necessary actions in various degrees are also also the alternatives. Delayed action for various periods are also the alternatives.
Our ancestors had likely survived by prejudice that sudden movement of bushes are caused by lurking predators.

[alancalverd]

Our ancestors had likely survived by prejudice that sudden movement of bushes are caused by lurking predators.

The difference between prejudice and observation is what makes us scientists superior to priests, politicians, philosophers, and other parasites and predators.