5 ms·

Well, yeah. Because "machine AI" has NOTHING much in common with how our biological brains work. And we aren't smart enough to know what intelligence really i

Well, yeah. Because "machine AI" has NOTHING much in common with how our biological brains work. And we aren't smart enough to know what intelligence really is when we can't even define it for ourselves or in animal models. And 99.999% of everyone working on machine AI has never taken a biology class let alone a class related to anatomy, neurology or experimental psychology so it's nothing more than "flinging shit on the wall and hoping it sticks" in terms of odds of success! Not that that would help because academia as it exists today frowns upon "getting out of your lane" or "challenging orthodoxy" so "knowledge hybridization" of two distinct silos is strictly forbidden. Basically the methodology of AI today is NO DIFFERENT than AI 1.0 from the 1960s and 1970s which was based on the assumption that all intelligence was merely predicate calculus and a fact store. The scientific and economic model was for that AI (and is still the model for AI today!) is nothing more than the Garden Gnome Business Plan: 1. Create a cute-but-sellable singular heuristic technique misnamed and misinterpreted as "intelligence" in the small 2. Sell the idea to implement the same thing 1000x, 1 000 000x, etc. in parallel 3. ???? 4. Success! We now how "strong AI" (which never comes because step #3 is bullshit and faith-based at best; fraud at worst) The problem is that's also IDENTICAL to the plan to create a 747 jet by putting all the parts into a shipping hold and shaking with the expectation that you'll have a fully-formed 747 pop out when you open it. Evolution is far smarter than us and has tested all the combinations) that take us generations to check. Evolution might well have taken longer per test but it's had a longer time. The best hope is to slavishly copy nature paying extreme attention to how nature pulls it off. That's NEVER BEEN DONE with AI!! So it's a VERY EASY technology to short in the long run because the fundamentals of assumptions and methodology are always such Epic Fail.

ramoz 6y ago
I think that's an intriguing point - and I also feel we generalize "artificial general intelligence" to the point & in such a way where humans have yet to even achieve that level of intelligence. How can we build smart systems if we don't know what smart even looks like and human benchmarks turn out inefficient for machines.
eutropia 6y ago
You start with a riff on the naturalistic fallacy "nature did it so it must be the right way" and conclude that all efforts to create AI not informed by neurobiology are doomed to fail (as doomed to fail as randomly assembling components and hoping for a jet as the output)? Further, I'd wager more than half of AI researchers are at least surface-level familiar with brain science, not (as you claim) less than 1 in a million (are there even a million AI researchers?). There's significant work between computational neuroscience, mathematics, philosophy, computer science, etc, etc in the field. Many smart people are giving it their best effort to understand different pieces of the puzzle from many different viewpoints and angles; FAANG corporations might be among the most visible, but their AI is necessarily profit driven and close to the ground, relevant to currently tractable problems (amenable to 'mere statistics'). And in fact, slavishly copying nature is something which has long been on the AI back-burner, but we're on the order of at least a decade from being able to create a computer system with enough transistors to do so. Not sure what else to say, really.
bglazer 6y ago
You're incorrect that neuroscientists and computational researchers don't collaborate. In fact, "computational neuroscience" departments have existed at major universities for a number of years. DeepMind was a spin-off of the Gatsby Computational Neuroscience Unit at UCL, and they continue to be influential in both neuroscience and (more famously) in artificial intelligence research. Here's an article from 2020 by authors at DeepMind and Geoff Hinton about potential biologically plausible mechanisms for backpropagation in the brain [0]. So, you can see that the fields actually influence each other in both directions. Computational researchers often propose ideas that neuroscientists then attempt to understand in biological systems. This same principle is true at all levels of computational and biological abstraction, from simulating individual neurons to machine learning and common sense. For the latter, research by Josh Tenenbaum might interest you. The reason this works is that there seem to be fundamental principles underlying information processing. Brains and CPU's are both systems that manipulate and store information, albeit in very different ways. Hell, even single cells and slime molds are capable of rudimentary decision making. So, the point is that we don't need to copy the brain. Instead, we just need to understand the principles of information well enough to build machines that can efficiently manipulate information and intelligence will arise out of that. Information theory and (by extension) statistics are the fields that deal most closely with this question, which is why they're used heavily by both neuroscientists and ML researchers. A rough analogy is that we don't build airplanes that flap their wings to fly. Instead, we understand aerodynamics well enough to generate lift and thrust through other mechanisms that evolution can't find. Like jet engines. Also, "slavishly" copying nature is insanely difficult. Biological neurons and brain tissue are extremely complex and poorly understood. Much of this complexity is likely incidental to information processing and would only hamper our efforts to build intelligent machines. Like, do we want our computer brain to get multiple sclerosis if the simulated neurons demyelinate? [0] https://www.nature.com/articles/s41583-020-0277-3 https://www.nature.com/articles/s41583-020-0277-3

ramoz 6y ago

I think that's an intriguing point - and I also feel we generalize "artificial general intelligence" to the point & in such a way where humans have yet to even achieve that level of intelligence. How can we build smart systems if we don't know what smart even looks like and human benchmarks turn out inefficient for machines.

eutropia 6y ago

You start with a riff on the naturalistic fallacy "nature did it so it must be the right way" and conclude that all efforts to create AI not informed by neurobiology are doomed to fail (as doomed to fail as randomly assembling components and hoping for a jet as the output)? Further, I'd wager more than half of AI researchers are at least surface-level familiar with brain science, not (as you claim) less than 1 in a million (are there even a million AI researchers?). There's significant work between computational neuroscience, mathematics, philosophy, computer science, etc, etc in the field. Many smart people are giving it their best effort to understand different pieces of the puzzle from many different viewpoints and angles; FAANG corporations might be among the most visible, but their AI is necessarily profit driven and close to the ground, relevant to currently tractable problems (amenable to 'mere statistics'). And in fact, slavishly copying nature is something which has long been on the AI back-burner, but we're on the order of at least a decade from being able to create a computer system with enough transistors to do so. Not sure what else to say, really.

bglazer 6y ago

You're incorrect that neuroscientists and computational researchers don't collaborate. In fact, "computational neuroscience" departments have existed at major universities for a number of years. DeepMind was a spin-off of the Gatsby Computational Neuroscience Unit at UCL, and they continue to be influential in both neuroscience and (more famously) in artificial intelligence research. Here's an article from 2020 by authors at DeepMind and Geoff Hinton about potential biologically plausible mechanisms for backpropagation in the brain [0]. So, you can see that the fields actually influence each other in both directions. Computational researchers often propose ideas that neuroscientists then attempt to understand in biological systems. This same principle is true at all levels of computational and biological abstraction, from simulating individual neurons to machine learning and common sense. For the latter, research by Josh Tenenbaum might interest you. The reason this works is that there seem to be fundamental principles underlying information processing. Brains and CPU's are both systems that manipulate and store information, albeit in very different ways. Hell, even single cells and slime molds are capable of rudimentary decision making. So, the point is that we don't need to copy the brain. Instead, we just need to understand the principles of information well enough to build machines that can efficiently manipulate information and intelligence will arise out of that. Information theory and (by extension) statistics are the fields that deal most closely with this question, which is why they're used heavily by both neuroscientists and ML researchers. A rough analogy is that we don't build airplanes that flap their wings to fly. Instead, we understand aerodynamics well enough to generate lift and thrust through other mechanisms that evolution can't find. Like jet engines. Also, "slavishly" copying nature is insanely difficult. Biological neurons and brain tissue are extremely complex and poorly understood. Much of this complexity is likely incidental to information processing and would only hamper our efforts to build intelligent machines. Like, do we want our computer brain to get multiple sclerosis if the simulated neurons demyelinate? [0] https://www.nature.com/articles/s41583-020-0277-3 https://www.nature.com/articles/s41583-020-0277-3