AIs gunning for our precious freelancers

[keenir]

It shouldn't take an AI to predict the response.

?
sorry, I was trying to quote one of the title characters from Wallace and Grommit, who is constantly inventing new machines.

.-------------------------------------------------.

i've been thinking...

I realized that Malloc is right: that AI is very much like airplanes. Both began as shoddy assemblages of doodles and sketches, then were made to do one thing...and then decades passed, wherein they got faster and faster and holy h e hockeysticks its fast...but ultimately neither can do things they weren't originally tested for in the beginning.

ie, the early tests in Paris, Kittyhawk, and southern England (among others - i think Brazil had test flights for airplane prototypes in that period as well; not sure about Australia) had the inventors dropping objects from their planes, carrying people, carrying objects, surveying landscapes, and even (this one's hearsay so far as i know) catching objects on the part of the plane or the pilot.

...which is pretty much what planes do now, all these decades later. (and before you object, I would wager that at least Kittyhawk saw a water landing or two)

So Malloc is correct in the train of reasoning/logic train...but not in the ultimate conclusion. There's probably a word for that, but I can't think of it; sorry.

[malloc]

I realized that Malloc is right: that AI is very much like airplanes. Both began as shoddy assemblages of doodles and sketches, then were made to do one thing...and then decades passed, wherein they got faster and faster and holy h e hockeysticks its fast...but ultimately neither can do things they weren't originally tested for in the beginning.

Sure but that would mean that generative AI will become even more efficient and competent and writing and drawing than now. Considering how many writers and artists are struggling to retain employment now, future iterations of generative AI may well kill their professions entirely. Meanwhile other forms of AI are rapidly taking over everything from biochemistry to taxi driving to therapy even.

This is a reference back a ways, but until we have robots that can make coffee in an arbitrary American home, we are not going to be replaced by AGI. And we are so far from simply having robots make coffee for us in our homes that it is not funny.

The absence of coffee-making robots seems driven more by lack of interest than practical obstacles. Presumably if engineers and investors really wanted a robot that could operate any coffee maker, they would throw their effort and resources behind it. That said, a robot specialized for this one task presumably wouldn't qualify as general intelligence.

[Travis B.]

This is a reference back a ways, but until we have robots that can make coffee in an arbitrary American home, we are not going to be replaced by AGI. And we are so far from simply having robots make coffee for us in our homes that it is not funny.

The absence of coffee-making robots seems driven more by lack of interest than practical obstacles. Presumably if engineers and investors really wanted a robot that could operate any coffee maker, they would throw their effort and resources behind it. That said, a robot specialized for this one task presumably wouldn't qualify as general intelligence.

It is not merely lack of interest. There are many fundamental AI problems involved in coffee-making robots:

• Coffee-making robots need to know about the layout of any given American house, and how to navigate them without fail.
• Coffee-making robots need to know where the many places people may store their coffee, coffee filters, coffee cups, etc., and be able to access them without fail.
• Coffee-making robots need to know how to recognize and operate a variety of equipment, including electric drip coffee makers, espresso machines, pod (e.g. Keurig) coffee makers, coffee grinders, etc. with varying designs and UI's, and do so without fail.
• Coffee-making robots need to be able to find sources of potable water and use them to dispense water into the available coffee maker without fail.
• Coffee-making robots need to be able to distinguish pre-ground with whole-bean coffee and grind coffee need be without fail.
• Coffee-making robots need to be able to operate a coffee pot and use it to dispense coffee into a coffee cup without fail.
• Coffee-making robots need to be able to clean up after themselves and properly put away things such as coffee and coffee grinders back into where they were originally stored without fail.

We are not close to solving even parts of this problem, much the less the whole shebang. The closest we have gotten is a robot that can take a pod, insert it in a pre-positioned pod coffee maker, and press the start button, which is nowhere close to a true solution to the coffee test.

[alice]

This is a reference back a ways, but until we have robots that can make coffee in an arbitrary American home, we are not going to be replaced by AGI. And we are so far from simply having robots make coffee for us in our homes that it is not funny.

The absence of coffee-making robots seems driven more by lack of interest than practical obstacles. Presumably if engineers and investors really wanted a robot that could operate any coffee maker, they would throw their effort and resources behind it. That said, a robot specialized for this one task presumably wouldn't qualify as general intelligence.

It is not merely lack of interest. There are many fundamental AI problems involved in coffee-making robots:

• Coffee-making robots need to know about the layout of any given American house, and how to navigate them without fail.
• Coffee-making robots need to know where the many places people may store their coffee, coffee filters, coffee cups, etc., and be able to access them without fail.
• Coffee-making robots need to know how to recognize and operate a variety of equipment, including electric drip coffee makers, espresso machines, pod (e.g. Keurig) coffee makers, coffee grinders, etc. with varying designs and UI's, and do so without fail.
• Coffee-making robots need to be able to find sources of potable water and use them to dispense water into the available coffee maker without fail.
• Coffee-making robots need to be able to distinguish pre-ground with whole-bean coffee and grind coffee need be without fail.
• Coffee-making robots need to be able to operate a coffee pot and use it to dispense coffee into a coffee cup without fail.
• Coffee-making robots need to be able to clean up after themselves and properly put away things such as coffee and coffee grinders back into where they were originally stored without fail.

We are not close to solving even parts of this problem, much the less the whole shebang. The closest we have gotten is a robot that can take a pod, insert it in a pre-positioned pod coffee maker, and press the start button, which is nowhere close to a true solution to the coffee test.

Yebbut technology is advancing so fast that these problems surely won't take that much longer to solve innit?

[malloc]

We are not close to solving even parts of this problem, much the less the whole shebang. The closest we have gotten is a robot that can take a pod, insert it in a pre-positioned pod coffee maker, and press the start button, which is nowhere close to a true solution to the coffee test.

That would imply that brewing coffee is even more intellectually demanding than folding proteins and advanced mathematics, which seems counterintuitive to say the least. Even the dullest flat-earth antivaxxer MAGA can brew coffee but few humans can fold proteins or take arbitrary roots of complex numbers. It seems obvious to me that the same techniques used to identify tumors years in advance from tiny specks in X-rays would also allow robots to identify coffee and coffee accessories or navigate a house.

[zompist]

We are not close to solving even parts of this problem, much the less the whole shebang. The closest we have gotten is a robot that can take a pod, insert it in a pre-positioned pod coffee maker, and press the start button, which is nowhere close to a true solution to the coffee test.

That would imply that brewing coffee is even more intellectually demanding than folding proteins and advanced mathematics, which seems counterintuitive to say the least.

Good learning opportunity: intuition is often wrong.

For someone terrified of artificial general intelligence, you seem to have no understanding of the "general" part. It precisely means being good at a wide range of physical and mental tasks, not just "advanced mathematics". It precisely includes the sort of disconnected and difficult tasks that Travis lists.

We've been over this a lot, but a lot of what look like "intelligent" tasks, like playing chess or finding roots, turn out to be solvable with rather simple brute-force methods, available to the computer and not to humans only because the computer is stupid but fast.

Navigating a 3-D environment, like a human house, is in fact a tremendously difficult problem, or rather a set of problems, starting with understanding vision itself. If you think that's easy, why not enlighten us with a general algorithm for how an array of pixels can be turned into a 3-D model of the world?

[malloc]

For someone terrified of artificial general intelligence, you seem to have no understanding of the "general" part. It precisely means being good at a wide range of physical and mental tasks, not just "advanced mathematics". It precisely includes the sort of disconnected and difficult tasks that Travis lists.

We already have AI models capable of handling a wide range of physical and mental tasks, everything from drawing original artwork to folding proteins. We could probably come quite close to AGI simply by folding all those AIs into one program and introducing a master AI to choose the appropriate model for a given task.

We've been over this a lot, but a lot of what look like "intelligent" tasks, like playing chess or finding roots, turn out to be solvable with rather simple brute-force methods, available to the computer and not to humans only because the computer is stupid but fast.

People used to consider chess and mathematics examples of highly sophisticated intelligence and only stopped when computers mastered those. Looking at the history of AI, it often feels like people are moving the goalposts to avoid granting computers the mantle of intelligence. If you traveled back to 1950 and told Alan Turing all the things computers have achieved, he would undoubtedly conclude that AI had succeeded beyond his wildest dreams.

Navigating a 3-D environment, like a human house, is in fact a tremendously difficult problem, or rather a set of problems, starting with understanding vision itself. If you think that's easy, why not enlighten us with a general algorithm for how an array of pixels can be turned into a 3-D model of the world?

That really is a question for experts in computer science and cognitive psychology. Nonetheless we know that navigating a 3-D environment requires little intelligence simply because even the least intelligent animals like insects achieve this with ease. Are you seriously suggesting that a housefly with mere thousands of neurons is more intelligent than a data center with trillions of logic gates comprising its processing units?

[zompist]

For someone terrified of artificial general intelligence, you seem to have no understanding of the "general" part. It precisely means being good at a wide range of physical and mental tasks, not just "advanced mathematics". It precisely includes the sort of disconnected and difficult tasks that Travis lists.

We already have AI models capable of handling a wide range of physical and mental tasks, everything from drawing original artwork to folding proteins.

So your example of physical tasks are two purely symbolic tasks. Do you know what "physical" means? Or do you thnk that AlphaFold is physically manipulating proteins?

Sometimes you sound like an AI yourself, in that you are utterly confident of your statements even when you know nothing about the subject. Really, stop talking about protein folding until you can explain in your own words why it's a difficult problem.

People used to consider chess and mathematics examples of highly sophisticated intelligence and only stopped when computers mastered those. Looking at the history of AI, it often feels like people are moving the goalposts to avoid granting computers the mantle of intelligence. If you traveled back to 1950 and told Alan Turing all the things computers have achieved, he would undoubtedly conclude that AI had succeeded beyond his wildest dreams.

And I've been criticizing Turing for thirty years. Your conception of AI seems to come right out of the 1950s. AI researchers like Terry Winograd and Roger Schank were talking about real-world knowledge back in the early 1970s.

Yeah, people got fixated on chess, and that was short-sighted. Vision turned out to be a much harder problem.

Navigating a 3-D environment, like a human house, is in fact a tremendously difficult problem, or rather a set of problems, starting with understanding vision itself. If you think that's easy, why not enlighten us with a general algorithm for how an array of pixels can be turned into a 3-D model of the world?

That really is a question for experts in computer science and cognitive psychology.

Name three who've addressed the problem and support your claim that it's extremely simple.

Nonetheless we know that navigating a 3-D environment requires little intelligence simply because even the least intelligent animals like insects achieve this with ease. Are you seriously suggesting that a housefly with mere thousands of neurons is more intelligent than a data center with trillions of logic gates comprising its processing units?

Dude, I spent my career as a programmer, I know how computers work way more than you do. You can't fool me by equating neurons with "logic gates". Do you know what a logic gate does or even how complex its "computation" is? The basic processor in the computer on your desk is absolutely brain-dead simple— you can learn to program it at the assembly level in a one-quarter course. Computers are fast, not intelligent.

A fruit fly has 140,000 neurons with 50,000,000 connections between them. Housefly brains haven't been mapped yet but they're estimated at 250,000 neurons.

That data center can't navigate an arbitrary room as well as a housefly. How do we know? Because "self-driving cars" can't navigate an arbitrary street without killing people, and that's with human observers doing a lot of the work. E.g. the company Cruise was found to use 1.5 workers per vehicle. If your advanced math is as bad as you say it is, that's 0.5 human workers more than are needed for a human to drive a regular car.

(I also have to note that we don't know much about how flies navigate the world. Their needs are certainly simpler than even a simple mammal's. A shrew, which I talked about recently, has 52 million neurons.)

[malloc]

Dude, I spent my career as a programmer, I know how computers work way more than you do. You can't fool me by equating neurons with "logic gates". Do you know what a logic gate does or even how complex its "computation" is? The basic processor in the computer on your desk is absolutely brain-dead simple— you can learn to program it at the assembly level in a one-quarter course. Computers are fast, not intelligent.

But neurons are nothing if not logic gates in their overall function. The main difference is composition, tubes of fat full of electrolytes versus sandwiched semiconductors. That and neurons are considerably less precise and reliable owing to their crude organic composition.

That data center can't navigate an arbitrary room as well as a housefly. How do we know? Because "self-driving cars" can't navigate an arbitrary street without killing people, and that's with human observers doing a lot of the work. E.g. the company Cruise was found to use 1.5 workers per vehicle. If your advanced math is as bad as you say it is, that's 0.5 human workers more than are needed for a human to drive a regular car.

Perhaps the fault lies with the programmers rather than the hardware. The data center has a million times the computational power of a human brain, given probably quadrillions of logic gates versus only fifteen billion neurons and likewise a trillion times the compute of an insect brain. It clearly has the resources to drive a car or direct a robot around a kitchen, just not the software yet.

[zompist]

But neurons are nothing if not logic gates in their overall function.

Jeez louise they are not, in much the same sense that an integer is not a database.

[malloc]

But neurons are nothing if not logic gates in their overall function.

Jeez louise they are not, in much the same sense that an integer is not a database.

They clearly are though, in that they both take inputs in the form of electrical impulses and produce an impulse in response. Certainly the technical nature of these impulses differs between the two, along with the "truth tables" available to each. Nonetheless there is no disputing the analogy between neurons and logic gates. It sounds like you object on the grounds that neurons allow many more inputs than logic gates, thousands rather than one or two. While that does complicate the analogy, it hardly changes the basic comparison. Each neuron takes a finite number of inputs and produces one output (with inputs and outputs equivalent to the "true" of logic gates).

[bradrn]

It sounds like you object on the grounds that neurons allow many more inputs than logic gates, thousands rather than one or two. While that does complicate the analogy, it hardly changes the basic comparison. Each neuron takes a finite number of inputs and produces one output (with inputs and outputs equivalent to the "true" of logic gates).

You’re forgetting that neurons accept continuous inputs while logical gates are always binary. The better comparison to natural neurons would be artificial neurons — though even those are greatly oversimplified compared to the behaviour of physical neurons.

[zompist]

But neurons are nothing if not logic gates in their overall function.

Jeez louise they are not, in much the same sense that an integer is not a database.

They clearly are though, in that they both take inputs in the form of electrical impulses and produce an impulse in response.

I can play that game too: logic gates and neurons are just the same because they're both made of atoms!

Or I can play your never-look-things-up game: neurons are made of only thousands of atoms, while logic gates are made of trillions!

Or your apples-and-oranges game: a logic gate is a minimal circuit that does just one binary operation, while neuron behavior are determined by 3 billion genetic base pairs!

You don't know what you're talking about and show no interest in learning.

[zompist]

Each neuron takes a finite number of inputs and produces one output (with inputs and outputs equivalent to the "true" of logic gates).

Can't let this one stand either... no, the inputs and outputs of neurons are not true/false. They are frequencies.

[WeepingElf]

Each neuron takes a finite number of inputs and produces one output (with inputs and outputs equivalent to the "true" of logic gates).

Can't let this one stand either... no, the inputs and outputs of neurons are not true/false. They are frequencies.

Yes - neurons are somewhat similar to logic gates - but more complex.

[bradrn]

Each neuron takes a finite number of inputs and produces one output (with inputs and outputs equivalent to the "true" of logic gates).

Can't let this one stand either... no, the inputs and outputs of neurons are not true/false. They are frequencies.

Actually, to my understanding malloc is correct on this one — neurons produce an action potential, which is an electric pulse. The action potential is always of maximum magnitude, called the all-or-none law. (The whole pulse train has a frequency, of course.)

[zompist]

Each neuron takes a finite number of inputs and produces one output (with inputs and outputs equivalent to the "true" of logic gates).

Can't let this one stand either... no, the inputs and outputs of neurons are not true/false. They are frequencies.

Actually, to my understanding malloc is correct on this one — neurons produce an action potential, which is an electric pulse. The action potential is always of maximum magnitude, called the all-or-none law. (The whole pulse train has a frequency, of course.)

Sigh. Yes, a neuron fires. Then it fires again. Then it fires again. Then it fires again. Then it fires again. Then it fires again. What's that? A frequency.

The process is explained in Hardin's Color for Philosophers... unfortunately packed up right now or I'd give more details.

As it seems I need to ram this point into the ground: a neuron is a pretty complex little device. Its inputs and outputs are, at the least, equivalent to integers, it's connected to an average of 7000 other neurons, and there's a lot of other stuff going on (e.g. neurotransmitters). It is absolutely not equivalent to a logic gate, which is the minimal operator acting on and outputting strictly on/off signals. malloc wants to scare himself by inflating the power of computers and underestimating the complexity of brains, and one trick he's using is comparing things that are not comparable; another is grossly undercounting neurons. And these errors, of several orders of magnitude each, add up, so he's spouting absurdities.

[rotting bones]

While human brains are more complex and specialized than our AI models, more "complex" outputs aren't necessarily more useful for reasoning. Neurons in a neural network deliberately convert numerical outputs into booleans. It has been theoretically shown there is no gain from having long chains of neurons without this step.

[bradrn]

As it seems I need to ram this point into the ground: a neuron is a pretty complex little device. Its inputs and outputs are, at the least, equivalent to integers, it's connected to an average of 7000 other neurons, and there's a lot of other stuff going on (e.g. neurotransmitters). It is absolutely not equivalent to a logic gate, which is the minimal operator acting on and outputting strictly on/off signals. malloc wants to scare himself by inflating the power of computers and underestimating the complexity of brains, and one trick he's using is comparing things that are not comparable; another is grossly undercounting neurons. And these errors, of several orders of magnitude each, add up, so he's spouting absurdities.

To be clear, I agree with all this.

Sigh. Yes, a neuron fires. Then it fires again. Then it fires again. Then it fires again. Then it fires again. Then it fires again. What's that? A frequency.

To a first approximation, a neuron does a weighted sum of its inputs and produces an output of a constant magnitude whenever the result is above some threshold. This is simply repeated on every input pulse it receives; to my understanding it does no frequency measurement.

(So where do the repeated pulses come from? From the sensory neurons which receive input from the outside world… which is probably why your book discusses this. But those are a different type of neuron to those we're discussing.)

[rotting bones]

Correction: I shouldn't have said "convert to boolean". What has been shown in the theory of neural networks is that introducing non-linearity at the end of each artificial neuron is important to prevent the whole network from being reduced to one system of linear equations.

Traditionally, the nonlinear activation function at the end of each neuron was a sigmoid function, which bounds the output between 0 and 1. tanh was also used, which similarly caps the output from -1 to +1.

These days, it has been found that the much cheaper ReLU function actually performs better than the sigmoid or tanh. All it does is set negative outputs to 0. That's it. The positive values are allowed to grow. Removing negative values prevents them from canceling out the positives, which allows signals to propagate.

None of these activation functions output a constant magnitude.