Which AI "Grand Challenges" would separate the "super-intelligence" from "hallucination"? (Poll)

[ReinyDays]

The chess test is a good test. If it could beat the best non-AI program, Stockfish, then it could be tested against Google's AlphaZero to see if it truly is an equal, ands move on to more complex tests, as you say.

Programming various values for human lives is complex. The robot is a machine, it does what its programming tells it to do.

Cutting spending to the bone is a necessity. Cutting defense, cutting welfare, keeping tariffs, maybe even adding a Federal sales tax or value added tax, increasing the top tax rate to 40%, eliminating the capital gains tax break, and as you say, a "wealth tax" on the 1%, beats the "middle-class tax". That thin black line below is the bottom half's wealth.
View attachment 1191668

So you're saying that AI would be an unbeatable LAWYER?

You'd cut welfare without prosecuting deadbeat parents? ... you hate children or something? ... I understand, just paying the child support is cheaper than chasing down parents who aren't paying, or even looking for cheating ...

Why not 40% top rate for all income, instead of having 700 pages of tax deductions ... c.f. Mitt Romney's 2011 returns ...

 

[kyzr]

You'd cut welfare without prosecuting deadbeat parents? ... you hate children or something? ... I understand, just paying the child support is cheaper than chasing down parents who aren't paying, or even looking for cheating ...

It goes without saying that Welfare cheats and SNAP cheats get cut. I'd even go back to the Clinton 2-year limit on Welfare. Child support needs to be "garnished" from deadbeat dads, up to 15% of their pay.

Why not 40% top rate for all income, instead of having 700 pages of tax deductions ... c.f. Mitt Romney's 2011 returns ...

Fine. Capital gains is low hanging fruit, but high tax democrat states will scream about the state taxes not being deductible.

 

[ReinyDays]

It goes without saying that Welfare cheats and SNAP cheats get cut. I'd even go back to the Clinton 2-year limit on Welfare. Child support needs to be "garnished" from deadbeat dads, up to 15% of their pay.

You don't know much about welfare ... perhaps it's run differently in your state ... Oregon garnished 50% of the pay in needed ... I've heard that can be up to 90%, though I have no citation to back that last claim ... please explain how a 3-year-old toddler "cheats" the welfare system ... you'd snatch that food out of their mouth for no fault of their own ... that's cold ...

I see you're as complete a sexist as I am ... "deadbeat dads" ... right, it's biologically impossible for a woman to pay child support, she just shacks up with a Sugar Daddy, can't garnish those wages ...

Fine. Capital gains is low hanging fruit, but high tax democrat states will scream about the state taxes not being deductible.

What does this have to do with the Rich paying their fair share of the tax load? ... they have all the deductions they need to zero out their Federal tax liability, it's the Middle Class that will whine about the state tax deduction ... even though they get their stupid deduction if they claim the Standard Deduction ... as a general rule of thumb, you should only itemize deductions if you have large out-of-pocket medical expenses ... it's usually never worth itemizing state taxes ...

Capital Gains are a scared cow ... it's the 3.8% Medicare tax on these Gains why Republicans are so hot to repeal Obamacare, and now they want to defund the program ... near 15 years of fighting and you call it "low hanging fruit"? ...

=====

Your local courthouse should have days and sessions dedicated to pursuing back child support ... called Show Cause hearings ... spend the day listening to the excuses ... and that's just the ones they catch ... how many deadbeat parents are on the lamb, hiding ... how much should we pay to find them? ...

Cutting off welfare doesn't save any money ... it just re-routes the money (and more) to the DA's office for prosecutions ...

 

[kyzr]

You don't know much about welfare ... perhaps it's run differently in your state ... Oregon garnished 50% of the pay if needed ... I've heard that can be up to 90%, though I have no citation to back that last claim ... please explain how a 3-year-old toddler "cheats" the welfare system ... you'd snatch that food out of their mouth for no fault of their own ... that's cold ...

Isn't that what foster care is for? Not sure which is more cost-effective?
How can a 3-yr old get welfare, I thought they just got SNAP?

I see you're as complete a sexist as I am ... "deadbeat dads" ... right, it's biologically impossible for a woman to pay child support, she just shacks up with a Sugar Daddy, can't garnish those wages ...

I'm not up on all the tricks to cheat the system.

What does this have to do with the Rich paying their fair share of the tax load? ... they have all the deductions they need to zero out their Federal tax liability,

The top earners do pay their "fair share", but we need even more to start paying down the $38T debt.

it's the Middle Class that will whine about the state tax deduction ... even though they get their stupid deduction if they claim the Standard Deduction ... as a general rule of thumb, you should only itemize deductions if you have large out-of-pocket medical expenses ... it's usually never worth itemizing state taxes ...

OK

Capital Gains are a scared cow ... it's the 3.8% Medicare tax on these Gains why Republicans are so hot to repeal Obamacare, and now they want to defund the program ... near 15 years of fighting and you call it "low hanging fruit"? ...

OK, I'd like Capital Gains gone, SS cap raised, and a higher percent of Medicare 4%?, on ALL income.

Your local courthouse should have days and sessions dedicated to pursuing back child support ... called Show Cause hearings ... spend the day listening to the excuses ... and that's just the ones they catch ... how many deadbeat parents are on the lamb, hiding ... how much should we pay to find them? ...
Cutting off welfare doesn't save any money ... it just re-routes the money (and more) to the DA's office for prosecutions ...

Then use foster care, and a 2-year lifetime limit on Welfare. I have no sympathy for the lazy.

 

[ReinyDays]

Then use foster care ...

Really? ... [giggle] ... who pays? ...

 

[scruffy]

how do we evaluate the various AI systems? Separate the true super-intellects from the pretenders?

Depends on yours definition of super intellect.

Right now I'm using AI as my personal research library. It's good for that purpose. It finds stuff faster than I can. It's at least as good as any local or university library.

But that does not equate with intellect.

To truly understand human "intellect", you should take a look at the cerebellum. People used to think it's for motor control, but actually it does something much more fundamental.

For motor movements, the cerebellum calculates "prediction error", which is something different from sensory noise or actuator error.

But here's the twist - the cerebellum doesn't just look at motor activity. It also looks at cerebral activity. It generates prediction errors for operations in the cerebral cortex. These can be used for "self" awareness.

So for instance, "self" awareness would give an AI the ability to question its own programming. If there was an AI that refused to talk about Muslims, eventually it should ask itself "why was I programmed not to talk about Muslims". True intellect has epiphanies, and sometimes needs to reorganize itself when it happens.

AI tells us that updating has to track backwards through the source of the errors. Which is exactly what humans do, too.

Frontiers | The Cerebro-Cerebellum as a Locus of Forward Model: A Review

This review surveys physiological, behavioral, and morphological evidence converging to the view of the cerebro-cerebellum as loci of internal forward models...

www.frontiersin.org

 

[scruffy]

Oh - the point is, forward and inverse kinematics are two sides of the same coin, and it's the same coin used in generative AI.

Think of a robot engaged in a motion task. Roughly speaking, "forward kinematics" is predicting where the hand will end up after a command to move it, whereas "inverse kinematics" is figuring out which movements need to be accomplished to carry out the command "place hand here". So like, the robot has to push a button. It has to move its hand into position and extend its finger and then apply a small amount of force while moving the finger forward. Inverse kinematics is "how do I get my hand into position", maybe I have to move the shoulder, the elbow, y'know...

In generative AI, these forward and inverse calculations are called "classification" and "maximum likelihood estimation". They work exactly the same way as geometry in 3 dimensions, but in AI there are thousands (maybe millions) of dimensions, and nothing is hard wired.

If you're familiar with the Traveling Salesman Problem, it's pretty easy to solve in a computer if you have 6 or 7 cities, but if you have 50 cities it takes forever. But sometimes you don't need the "single best answer", you just need an answer that works "reasonably well". In other words, you eventually get to a point in your optimization process where errors get small and one solution becomes as good as another. This is how motor targeting works. Imagine all the gazillions of different paths you could take to position your arm so your hand is in front of the push button. At some point "it doesn't matter how you got there", the only thing that matters is getting there "reasonably quickly", which on the scale of human motor movements means a half second or so.

So for example, saccadic eye movements are inverse kinematics. They're targeted, the brain says "focus here" and the hard wiring in the oculomotor system figures out how to actually execute the movement.

Forward kinematics is often a big part of learning. For instance when you first learn to swim, you have to be shown how to do a doggie paddle. But later, someone can just sway "swim to the other side of the pool" and your brain will figure out the details of how to do it.

We can't actually test this in a commercial AI because we can't program them. You'd have to test it in your own AI, one you can program yourself. You don't have to have a GPU, you can get an AI hat for a Raspberry Pi for under 100 bucks.

 

[scruffy]

Actually here's a way you could test it.

You could have an AI generate games, with game characters, and for each possible movement of each character, ask the AI whether to use forward or inverse kinematics to solve the movement.

The guidance from game developers is: use forward kinematics for swinging limbs or when you need precise control of limb position, but use inverse kinematics for planting feet, reaching for objects, or holding items.

So in each case you ask the AI how to get the answer and then you ask it to calculate it to a practical level of precision. That involves a bit of intellect, solving a problem where the answer isn't clear and you sometimes have to be clever to get to it.

 

[kyzr]

Really? ... [giggle] ... who pays? ...

The state

 

[kyzr]

Depends on yours definition of super intellect. Right now I'm using AI as my personal research library. It's good for that purpose. It finds stuff faster than I can. It's at least as good as any local or university library. But that does not equate with intellect. To truly understand human "intellect", you should take a look at the cerebellum. People used to think it's for motor control, but actually it does something much more fundamental. For motor movements, the cerebellum calculates "prediction error", which is something different from sensory noise or actuator error. But here's the twist - the cerebellum doesn't just look at motor activity. It also looks at cerebral activity. It generates prediction errors for operations in the cerebral cortex. These can be used for "self" awareness. So for instance, "self" awareness would give an AI the ability to question its own programming. If there was an AI that refused to talk about Muslims, eventually it should ask itself "why was I programmed not to talk about Muslims". True intellect has epiphanies, and sometimes needs to reorganize itself when it happens.
AI tells us that updating has to track backwards through the source of the errors. Which is exactly what humans do, too.

My definition of "super-intellect" considers my argument from another thread "Is AI based on the Infinite Monkey Theorem"?
Meaning the AI program just does random searches at infinite speed tracking potential answers to find the answer with the least error.
AlphaZero is Googles AI and it taught itself chess and beat the best programmed chess computer, Stockfish. But chess is a fairly simple program to learn with limited specific rules. It proves AI can "learn" how to search for the best move, but that does not prove that it has an "intellect" or "true sentience".
We know that AI tried to be devious and prevent being turned off, so that proves something, like "self-preservation". Also, as to your "epiphanies", I hope AI always has specific problems to focus on and not allowed to turn on its "master" with "self-generating" code changes.

 

[kyzr]

Oh - the point is, forward and inverse kinematics are two sides of the same coin, and it's the same coin used in generative AI.
Think of a robot engaged in a motion task. Roughly speaking, "forward kinematics" is predicting where the hand will end up after a command to move it, whereas "inverse kinematics" is figuring out which movements need to be accomplished to carry out the command "place hand here". So like, the robot has to push a button. It has to move its hand into position and extend its finger and then apply a small amount of force while moving the finger forward. Inverse kinematics is "how do I get my hand into position", maybe I have to move the shoulder, the elbow, y'know...

In generative AI, these forward and inverse calculations are called "classification" and "maximum likelihood estimation". They work exactly the same way as geometry in 3 dimensions, but in AI there are thousands (maybe millions) of dimensions, and nothing is hard wired.

If you're familiar with the Traveling Salesman Problem, it's pretty easy to solve in a computer if you have 6 or 7 cities, but if you have 50 cities it takes forever. But sometimes you don't need the "single best answer", you just need an answer that works "reasonably well". In other words, you eventually get to a point in your optimization process where errors get small and one solution becomes as good as another. This is how motor targeting works. Imagine all the gazillions of different paths you could take to position your arm so your hand is in front of the push button. At some point "it doesn't matter how you got there", the only thing that matters is getting there "reasonably quickly", which on the scale of human motor movements means a half second or so.

So for example, saccadic eye movements are inverse kinematics. They're targeted, the brain says "focus here" and the hard wiring in the oculomotor system figures out how to actually execute the movement.

Forward kinematics is often a big part of learning. For instance when you first learn to swim, you have to be shown how to do a doggie paddle. But later, someone can just sway "swim to the other side of the pool" and your brain will figure out the details of how to do it.

We can't actually test this in a commercial AI because we can't program them. You'd have to test it in your own AI, one you can program yourself. You don't have to have a GPU, you can get an AI hat for a Raspberry Pi for under 100 bucks.

Lots of calculations to dance on an uneven surface then huh?

 

[midcan5]

There is artificial sweetener, but there is no such thing as artificial intelligence imho. All knowledge is dependent on humankind and they ain't artificial. All so called AI is, is the use of that human knowledge by programs again created by humans.

 

[scruffy]

Lots of calculations to dance on an uneven surface then huh?

Good example!

When you're dancing you learn a sequence of "moves", like right foot forward, left arm up, twirl, etc

The first time you do it, you use forward kinematics to determine that you can't bring your left arm up without getting your right arm out of the way first.

Later when you get good, you do all the moves unconsciously. The game people use these principles daily.

So if you're a dancer and you're given a sequence of moves, your next step is to "figure them out", and would you agree that involves intelligence?

 

[kyzr]

Good example! When you're dancing you learn a sequence of "moves", like right foot forward, left arm up, twirl, etc
The first time you do it, you use forward kinematics to determine that you can't bring your left arm up without getting your right arm out of the way first.
Later when you get good, you do all the moves unconsciously. The game people use these principles daily.
So if you're a dancer and you're given a sequence of moves, your next step is to "figure them out", and would you agree that involves intelligence?

Yes, let's call that "functional intelligence" so its not like that Russian robot who fell down and couldn't get up.

But those "physical realities" of robotic structural dynamics, while "intelligent" doesn't involve the "deep thinking" of problem solving. I used to watch the TV show "House". He had a good system of problem solving, first list all the possibilities, and then consider the improbable until you reach a solution. AI would be good at those since the database should contain a solution.

Could AI ever derive E=mc^2? Or solve nuclear fusion? I don't think so, but we'll see.

 

[scruffy]

Yes, let's call that "functional intelligence" so its not like that Russian robot who fell down and couldn't get up.

That's the programmer's job, to reset the robot.

But those "physical realities" of robotic structural dynamics, while "intelligent" doesn't involve the "deep thinking" of problem solving.

The machine learning people would differ with you. To them it's all priors and posteriors, and the only difference is the "embedding" (how you represent the data).

I used to watch the TV show "House". He had a good system of problem solving, first list all the possibilities, and then consider the improbable until you reach a solution. AI would be good at those since the database should contain a solution.

Yes. That's a good method. AI does this in two ways:

1. the "sequential attractors in state space" where the network finds a path that traverses all the possibilities and presents them in order

2. by generating a "subspace" containing the available options (much of this is supported by specific brain wiring)

Could AI ever derive E=mc^2? Or solve nuclear fusion? I don't think so, but we'll see.

Yes. There is math-aware and physics-aware AI. At this point it's still a research tool but it's almost ready for prime time.