Right now there is a big debate over whether AI is like a brain, or like an algorithm. I think that this is a lot like debating whether planes are more like birds, or like blimps. I’ll be arguing pro-bird & pro-brain.
Just to ground the analogy, In the late 1800s the Wright brothers spent a lot of time studying birds. They helped develop simple models of lift to explain their flight, they built wind tunnels in their lab to test and refine their models, they created new types of gliders based on their findings, and eventually they created the plane—a flying machine with wings.
Obviously bird wings have major differences from plane wings. Bird wings have feathers, they fold in the middle, they can flap. Inside they are made of meat and bone. Early aeronauts could have come up with a new word for plane wings, but instead they borrowed the word “wing” from birds, and I think for good reason.
Imagine you had just witnessed the Wright brothers fly, and now you’re traveling around explaining what you saw. You could say they made a flying machine, however blimps had already been around for about 50 years. Maybe you could call it a faster/smaller flying machine, but people would likely get confused trying to imagine a faster/smaller blimp.
Instead, you would probably say “No, this flying machine is different! Instead of a balloon this flying machine has wings”. And immediately people would recognize that you are not talking about some new type of blimp.
II.
If you ask most smart non-neuroscientists what is going on in the brain, you will usually get an idea of a big complex interconnected web of neurons that fire into each other, creating a cascade that somehow processes information. This web of neurons continually updates itself via experience, with connections growing stronger or weaker over time as you learn.
This is also a great simplified description of how artificial neural networks work. Which shouldn’t be too surprising—artificial neural networks were largely developed as a joint effort between cognitive psychologists and computer scientists in the 50s and 60s to try and model the brain.
Note that we still don’t really know how the brain works. The Wright brothers didn’t really understand aerodynamics either. It’s one thing to build something cool that works, but it takes a long time to develop a comprehensive theory of how something really works.
The path to understanding flight looked something like this
Get a rough intuition by studying bird wings
Form this rough intuition into a crude, inaccurate model of flight
Build a crude flying machine and study it in a lab
Gradually improve your flying machine and theoretical model of flight along with it
Eventually create a model of flight good enough to explain how birds fly
I think the path to understanding intelligence will look like this
Get a rough intuition by studying animal brains
Form this rough intuition into a crude, inaccurate model of intelligence
Build a crude artificial intelligence and study it in a lab
Gradually improve your AI and theoretical model of intelligence ← (YOU ARE HERE)
Eventually create a model of intelligence good enough to explain animal brains
III.
Up until the 2010s, artificial neural networks kinda sucked. Yann LeCun (head of Meta’s AI lab) is famous for building the first convolutional neural network back in the 80s that could read zip codes for the post office. Meanwhile regular hand crafted algorithmic “AI” was doing cool things like beating grandmasters at chess.
(In the 1880s the Wright brothers were experimenting with kites while the first Zeppelins were being built.)
People saying “AI works like the brain” back then caused a lot of confusion and turned the phrase into an intellectual faux-pas. People would assume you meant “Chess AI works like the brain” and anyone who knew anything about chess AI would correct you and rightfully say that a hand crafted tree search algorithm doesn’t really work anything like the brain.
Today this causes confusion in the other direction. People continue to confidently state that ChatGPT works nothing like a brain, it is just a fancy computer algorithm. In the same way blimps are fancy balloons.
The metaphors we use to understand new things end up being really important—they are the starting points that we build our understanding off of. I don’t think there’s any getting around it either, Bayesians always need priors, so it’s important to pick a good starting place.
When I think blimp I think slow, massive balloons that are tough to maneuver. Maybe useful for sight-seeing, but pretty impractical as a method of rapid transportation. I could never imagine a F15 starting from an intuition of a blimp. There are some obvious ways that planes are like blimps—they’re man made and they hold people. They don’t have feathers. But those facts seem obvious enough to not need a metaphor to understand—the hard question is how planes avoid falling out of the air.
When I think of algorithms I think of a hard coded set of rules, incapable of nuance, or art. Things like thought or emotion seem like obvious dead-end impossibilities. It’s no surprise then that so many assume that AI art is just some type of fancy database lookup—creating a collage of images on the fly. How else could they work? Art is done by brains, not algorithms.
When I tell people they are often surprised to hear that neural networks can run offline, and even more surprised to hear the only information they have access to is stored in the connection weights of the neural network.
The most famous algorithm is long division. Are we really sure that’s the best starting intuition for understanding AI?
…and as lawmakers start to pass legislation on AI, how much of that will be based on their starting intuition?
IV.
In some sense artificial neural networks are still algorithms, after all everything on a computer is eventually compiled into assembly. If you see an algorithm as a hundred billion lines of “manipulate bit X in register Y” then sure, ChatGPT is an algorithm.
But that framing doesn’t have much to do with the intuition we have when we think of algorithms. Our intuition on what algorithms can and can’t do is based on our experience with regular code—rules written by people—not an amorphous mass of billions of weights that are gradually trained from example.
Personally, I don’t think the super low-level implementation matters too much for anything other than speed. Companies are constantly developing new processors with new instructions to run neural networks faster and faster. Most phones now have a specialized neural processing unit to run neural networks faster than a CPU or GPU. I think it’s quite likely that one day we’ll have mechanical neurons that are completely optimized for the task, and maybe those will end up looking a lot like biological neurons. But this game of swapping out hardware is more about changing speed, not function.
This brings us into the idea of substrate independence, which is a whole article in itself, but I’ll leave a good description from Max Tegmark
Alan Turing famously proved that computations are substrate-independent: There’s a vast variety of different computer architectures that are “universal” in the sense that they can all perform the exact same computations. So if you’re a conscious superintelligent character in a future computer game, you’d have no way of knowing whether you ran on a desktop, a tablet or a phone, because you would be substrate-independent.
Nor could you tell whether the logic gates of the computer were made of transistors, optical circuits or other hardware, or even what the fundamental laws of physics were. Because of this substrate-independence, shrewd engineers have been able to repeatedly replace the technologies inside our computers with dramatically better ones without changing the software, making computation twice as cheap roughly every couple of years for over a century, cutting the computer cost a whopping million million million times since my grandmothers were born. It’s precisely this substrate-independence of computation that implies that artificial intelligence is possible: Intelligence doesn’t require flesh, blood or carbon atoms.
A common response I will hear, especially from people who have studied neuroscience, is that when you get deep down into it artificial neural networks like ChatGPT don’t really resemble brains much at all.
Biological neurons are far more complicated than artificial neurons. Artificial neural networks are divided into layers whereas brains have nothing of the sort. The pattern of connection you see in the brain is completely different from what you see in an artificial neural network. Loads of things modern AI uses like ReLU functions and dot product attention and batch normalization have no biological equivalent. Even backpropagation, the foundational algorithm behind how artificial neural networks learn, probably isn’t going on in the brain.
This is all absolutely correct, but should be taken with a grain of salt.
Hinton has developed something like 50 different learning algorithms that are biologically plausible, but they all kinda work like backpropagation but worse, so we stuck with backpropagation. Researchers have made more complicated neurons that better resemble biological neurons, but it is faster and works better if you just add extra simple neurons, so we do that instead. Spiking neural networks have connection patterns more similar to what you see in the brain, but they learn slower and are tougher to work with than regular layered neural networks, so we use layered neural networks instead.
I bet the Wright brothers experimented with gluing feathers onto their gliders, but eventually decided it wasn’t worth the effort.
Now, feathers are beautifully evolved and extremely cool, but the fundamental thing that mattered is the wing, or more formally the airfoil. An airfoil causes air above it to move quickly at low pressure, and air below it to move slowly at high pressure. This pressure differential produces lift, the upward force that keeps your plane in the air. Below is a comparison of different airfoils from wikipedia, some man made and some biological.
Early aeronauts were able to tell that there was something special about wings even before they had a comprehensive theory of aerodynamics, and I think we can guess that there is something very special about neural networks, biological or otherwise, even before we have a comprehensive theory of intelligence.
If someone who had never seen a plane before asked me what a plane was, I’d say it’s like a mechanical bird. When someone asks me what a neural network is, I usually hesitate a little and say ‘it’s complicated’ because I don’t want to seem weird. But I should really just say it’s like a computerized brain.
Brains, Planes, Blimps, and Algorithms
I.
Right now there is a big debate over whether AI is like a brain, or like an algorithm. I think that this is a lot like debating whether planes are more like birds, or like blimps. I’ll be arguing pro-bird & pro-brain.
Just to ground the analogy, In the late 1800s the Wright brothers spent a lot of time studying birds. They helped develop simple models of lift to explain their flight, they built wind tunnels in their lab to test and refine their models, they created new types of gliders based on their findings, and eventually they created the plane—a flying machine with wings.
Obviously bird wings have major differences from plane wings. Bird wings have feathers, they fold in the middle, they can flap. Inside they are made of meat and bone. Early aeronauts could have come up with a new word for plane wings, but instead they borrowed the word “wing” from birds, and I think for good reason.
Imagine you had just witnessed the Wright brothers fly, and now you’re traveling around explaining what you saw. You could say they made a flying machine, however blimps had already been around for about 50 years. Maybe you could call it a faster/smaller flying machine, but people would likely get confused trying to imagine a faster/smaller blimp.
Instead, you would probably say “No, this flying machine is different! Instead of a balloon this flying machine has wings”. And immediately people would recognize that you are not talking about some new type of blimp.
II.
If you ask most smart non-neuroscientists what is going on in the brain, you will usually get an idea of a big complex interconnected web of neurons that fire into each other, creating a cascade that somehow processes information. This web of neurons continually updates itself via experience, with connections growing stronger or weaker over time as you learn.
This is also a great simplified description of how artificial neural networks work. Which shouldn’t be too surprising—artificial neural networks were largely developed as a joint effort between cognitive psychologists and computer scientists in the 50s and 60s to try and model the brain.
Note that we still don’t really know how the brain works. The Wright brothers didn’t really understand aerodynamics either. It’s one thing to build something cool that works, but it takes a long time to develop a comprehensive theory of how something really works.
The path to understanding flight looked something like this
Get a rough intuition by studying bird wings
Form this rough intuition into a crude, inaccurate model of flight
Build a crude flying machine and study it in a lab
Gradually improve your flying machine and theoretical model of flight along with it
Eventually create a model of flight good enough to explain how birds fly
I think the path to understanding intelligence will look like this
Get a rough intuition by studying animal brains
Form this rough intuition into a crude, inaccurate model of intelligence
Build a crude artificial intelligence and study it in a lab
Gradually improve your AI and theoretical model of intelligence ← (YOU ARE HERE)
Eventually create a model of intelligence good enough to explain animal brains
III.
Up until the 2010s, artificial neural networks kinda sucked. Yann LeCun (head of Meta’s AI lab) is famous for building the first convolutional neural network back in the 80s that could read zip codes for the post office. Meanwhile regular hand crafted algorithmic “AI” was doing cool things like beating grandmasters at chess.
(In the 1880s the Wright brothers were experimenting with kites while the first Zeppelins were being built.)
People saying “AI works like the brain” back then caused a lot of confusion and turned the phrase into an intellectual faux-pas. People would assume you meant “Chess AI works like the brain” and anyone who knew anything about chess AI would correct you and rightfully say that a hand crafted tree search algorithm doesn’t really work anything like the brain.
Today this causes confusion in the other direction. People continue to confidently state that ChatGPT works nothing like a brain, it is just a fancy computer algorithm. In the same way blimps are fancy balloons.
The metaphors we use to understand new things end up being really important—they are the starting points that we build our understanding off of. I don’t think there’s any getting around it either, Bayesians always need priors, so it’s important to pick a good starting place.
When I think blimp I think slow, massive balloons that are tough to maneuver. Maybe useful for sight-seeing, but pretty impractical as a method of rapid transportation. I could never imagine a F15 starting from an intuition of a blimp. There are some obvious ways that planes are like blimps—they’re man made and they hold people. They don’t have feathers. But those facts seem obvious enough to not need a metaphor to understand—the hard question is how planes avoid falling out of the air.
When I think of algorithms I think of a hard coded set of rules, incapable of nuance, or art. Things like thought or emotion seem like obvious dead-end impossibilities. It’s no surprise then that so many assume that AI art is just some type of fancy database lookup—creating a collage of images on the fly. How else could they work? Art is done by brains, not algorithms.
When I tell people they are often surprised to hear that neural networks can run offline, and even more surprised to hear the only information they have access to is stored in the connection weights of the neural network.
The most famous algorithm is long division. Are we really sure that’s the best starting intuition for understanding AI?
…and as lawmakers start to pass legislation on AI, how much of that will be based on their starting intuition?
IV.
In some sense artificial neural networks are still algorithms, after all everything on a computer is eventually compiled into assembly. If you see an algorithm as a hundred billion lines of “manipulate bit X in register Y” then sure, ChatGPT is an algorithm.
But that framing doesn’t have much to do with the intuition we have when we think of algorithms. Our intuition on what algorithms can and can’t do is based on our experience with regular code—rules written by people—not an amorphous mass of billions of weights that are gradually trained from example.
Personally, I don’t think the super low-level implementation matters too much for anything other than speed. Companies are constantly developing new processors with new instructions to run neural networks faster and faster. Most phones now have a specialized neural processing unit to run neural networks faster than a CPU or GPU. I think it’s quite likely that one day we’ll have mechanical neurons that are completely optimized for the task, and maybe those will end up looking a lot like biological neurons. But this game of swapping out hardware is more about changing speed, not function.
This brings us into the idea of substrate independence, which is a whole article in itself, but I’ll leave a good description from Max Tegmark
(full article @ https://www.edge.org/response-detail/27126 IMO it’s worth a read!)
V.
A common response I will hear, especially from people who have studied neuroscience, is that when you get deep down into it artificial neural networks like ChatGPT don’t really resemble brains much at all.
Biological neurons are far more complicated than artificial neurons. Artificial neural networks are divided into layers whereas brains have nothing of the sort. The pattern of connection you see in the brain is completely different from what you see in an artificial neural network. Loads of things modern AI uses like ReLU functions and dot product attention and batch normalization have no biological equivalent. Even backpropagation, the foundational algorithm behind how artificial neural networks learn, probably isn’t going on in the brain.
This is all absolutely correct, but should be taken with a grain of salt.
Hinton has developed something like 50 different learning algorithms that are biologically plausible, but they all kinda work like backpropagation but worse, so we stuck with backpropagation. Researchers have made more complicated neurons that better resemble biological neurons, but it is faster and works better if you just add extra simple neurons, so we do that instead. Spiking neural networks have connection patterns more similar to what you see in the brain, but they learn slower and are tougher to work with than regular layered neural networks, so we use layered neural networks instead.
I bet the Wright brothers experimented with gluing feathers onto their gliders, but eventually decided it wasn’t worth the effort.
Now, feathers are beautifully evolved and extremely cool, but the fundamental thing that mattered is the wing, or more formally the airfoil. An airfoil causes air above it to move quickly at low pressure, and air below it to move slowly at high pressure. This pressure differential produces lift, the upward force that keeps your plane in the air. Below is a comparison of different airfoils from wikipedia, some man made and some biological.
Early aeronauts were able to tell that there was something special about wings even before they had a comprehensive theory of aerodynamics, and I think we can guess that there is something very special about neural networks, biological or otherwise, even before we have a comprehensive theory of intelligence.
If someone who had never seen a plane before asked me what a plane was, I’d say it’s like a mechanical bird. When someone asks me what a neural network is, I usually hesitate a little and say ‘it’s complicated’ because I don’t want to seem weird. But I should really just say it’s like a computerized brain.