5 thoughts on “The Biggest Problem With AI (For Now)”
The biggest problem with neural networks is a flawed model. It’s based on Minsky and Papert’s perceptron model, which they had made out of discrete electronic components like resistors and capacitors and potentiometers; they determined a single layer was insufficient, and so a second layer was needed, fully connected to the previous layer, as well as a backpropagation of error signals to modify the weights. So the layer had to be simulated in a computer, using time domain multiplexing. That neural net model was based on what they could do with computers of that time. Our deep learning nets are just lots and lots of layers, throwing lots more computing power at the same model.
But that model of a neuron is a gross simplification of actual neurons. It treats them all as having some activation level, some analog output, all the time. Real neurons don’t act like that. Some emit a pulse train, which can be considered an analog signal, but some only emit a single pulse, acting more like a digital signal. Not all neurons are connected to all others in a previous “layer”, and real neurons aren’t really arranged in layers anyhow.
Most neurons are quiet most of the time, and if charge builds up but doesn’t exceed a firing threshold, that charge eventually dissipates. Instead of all the neurons having some activation level, there are about 8 “trains of thought” racing around the brain, with active neurons activating the downstream neurons.
And the model doesn’t consider glial cells at all, which have recently been found to communicate chemically with neurons and each other, both getting triggered by proximal neurons’ firing and by each other, and modifying neuron response. They provide also proximity information to neurons, enabling dendritic growth and the creation of synapses. Completely missing from the model, because we didn’t know 50 years ago that glial cells did anything but act as glue (hence the name) keeping neurons in place.
As long as we’re using that model, what we’re creating is just fancy statistics.
Not to worry, as I understand it, modern AIs are under the training philosophy of Hans Delbruck, Scientist and Saint, or is that Abby Someone? I can’t remember.
At the risk of committing heresy, I remain completely unconvinced that these massive neural network, deep learning models have any connection with intelligence. They are undeniably slick and when looking at superficial things like syntax and grammar from something like ChatGPT, it’s hard not to be impressed. Yet, when its output is compared to reality, example after example of complete fabrications are exposed.
The issue that should give the AI boosters pause is that the designers of these systems seem unable to correct this behavior. My understanding is that it is impossible to work back from the output of one of these systems to determine where it went wrong. While it’s marginally amusing when a lazy lawyer gets caught for relying on ChatGPT to construct a brief and it cited completely made up cases for support, it is more troubling when considering whether one of these systems piloting a self driving car might find a reason to drive into a brick wall at high speed.
If it is a ‘black box’ then how do they test it for accuracy? It would be like saying that the airline pilot ‘hasn’t crashed yet but don’t worry be happy’. And how do they check for ‘reproducibility’?
This sounds like all that hoopla there was about Chaos Theory decades ago. An interesting discipline to be sure …but does it prove anything?
Elon seems to agree with Mr. Motley. For awhile, he seemed content to wait for the government to gin up some sort of AI development oversight. Now, he seems to have decided that, as with cars and rockets, the way to avoid AI disasters that the oligarch hubrists of Silicon Valley might inflict on humanity is to beat them at their own game by developing better and benign AI first. As with all of his labors, I wish him well in this endeavor.
The biggest problem with neural networks is a flawed model. It’s based on Minsky and Papert’s perceptron model, which they had made out of discrete electronic components like resistors and capacitors and potentiometers; they determined a single layer was insufficient, and so a second layer was needed, fully connected to the previous layer, as well as a backpropagation of error signals to modify the weights. So the layer had to be simulated in a computer, using time domain multiplexing. That neural net model was based on what they could do with computers of that time. Our deep learning nets are just lots and lots of layers, throwing lots more computing power at the same model.
But that model of a neuron is a gross simplification of actual neurons. It treats them all as having some activation level, some analog output, all the time. Real neurons don’t act like that. Some emit a pulse train, which can be considered an analog signal, but some only emit a single pulse, acting more like a digital signal. Not all neurons are connected to all others in a previous “layer”, and real neurons aren’t really arranged in layers anyhow.
Most neurons are quiet most of the time, and if charge builds up but doesn’t exceed a firing threshold, that charge eventually dissipates. Instead of all the neurons having some activation level, there are about 8 “trains of thought” racing around the brain, with active neurons activating the downstream neurons.
And the model doesn’t consider glial cells at all, which have recently been found to communicate chemically with neurons and each other, both getting triggered by proximal neurons’ firing and by each other, and modifying neuron response. They provide also proximity information to neurons, enabling dendritic growth and the creation of synapses. Completely missing from the model, because we didn’t know 50 years ago that glial cells did anything but act as glue (hence the name) keeping neurons in place.
As long as we’re using that model, what we’re creating is just fancy statistics.
Not to worry, as I understand it, modern AIs are under the training philosophy of Hans Delbruck, Scientist and Saint, or is that Abby Someone? I can’t remember.
At the risk of committing heresy, I remain completely unconvinced that these massive neural network, deep learning models have any connection with intelligence. They are undeniably slick and when looking at superficial things like syntax and grammar from something like ChatGPT, it’s hard not to be impressed. Yet, when its output is compared to reality, example after example of complete fabrications are exposed.
The issue that should give the AI boosters pause is that the designers of these systems seem unable to correct this behavior. My understanding is that it is impossible to work back from the output of one of these systems to determine where it went wrong. While it’s marginally amusing when a lazy lawyer gets caught for relying on ChatGPT to construct a brief and it cited completely made up cases for support, it is more troubling when considering whether one of these systems piloting a self driving car might find a reason to drive into a brick wall at high speed.
If it is a ‘black box’ then how do they test it for accuracy? It would be like saying that the airline pilot ‘hasn’t crashed yet but don’t worry be happy’. And how do they check for ‘reproducibility’?
This sounds like all that hoopla there was about Chaos Theory decades ago. An interesting discipline to be sure …but does it prove anything?
Elon seems to agree with Mr. Motley. For awhile, he seemed content to wait for the government to gin up some sort of AI development oversight. Now, he seems to have decided that, as with cars and rockets, the way to avoid AI disasters that the oligarch hubrists of Silicon Valley might inflict on humanity is to beat them at their own game by developing better and benign AI first. As with all of his labors, I wish him well in this endeavor.