In 2024 the physics Nobel went to two AI researchers. Physicists asked how this counts. The answer is five pages long, published in 1982. Not a crossover. A claiming.
I just finished reading your exhaustive surveys of intelligence, this History and the Thermodynamics of Intelligence with an eye towards how to bake in guardrails to nueromorphic style AI minds. I am somewhat fixated on that type of AI both for its adaptability to real world learning and low power requirements plus the added bonus that it might actually be able to achieve a true type of novel intelligence. Not just the pattern matching lexical structure of current LLMs. I know that simulations of nueromporphic AI minds are being built, but at phenomenal energy costs and they still need to be flushed and rebooted to control drift.
You touched on simulated annealing style learning as a possible substitute for "programming" in ethical guardrails. I think you used the term adaptive style annealing as a method to bring an AI mind up through "childhood" and instill a sense of ethics into it. Not to just program in ethic, but to create a mental landscape where ethical behavior is baked in and just as in the robustness of an annealed coating you create a substrate that repels noise corrosion and still allows useful learning on the part of the AI. Like raising a child in a loving home and teaching it the difference between good and bad. But in this case in a more compressed timeline and at potentially an even deeper level.
Am I understanding this correctly? Or am I hallucinating? My mind is pattern matching on that subject.
I also confess I had to speed read your material. Legal training has me running through massive documents and applying my own type of mental compression. Errors can definitely be introduced that way.
Not hallucinating, though that particular idea isn't actually in either piece. What's there is noise clearing out false memories in Part 2, and annealing as a search strategy in Part 3. Neither one touches ethics or guardrails.
But I get why your mind landed there. There's clearly something in this stuff that feels like it should generalize to raising a mind well rather than just programming rules into it. That instinct is worth chasing on its own, maybe starting with constitutional AI or curriculum learning literature. Different vocabulary, similar spirit to what you're circling.
What's pulling you toward neuromorphic specifically, the power efficiency or the novelty angle?
Lightmatter's the real deal, and photonic compute for the linear-algebra-heavy stuff is a genuinely lower-power path. It just doesn't touch your bigger complaint. It changes how cheaply the computing gets done, not what's actually being computed.
The thing pulling me towards neuromorphic is that I don't really see binary implementation of linear algebra being robust enough. Power efficiency aside I really don't think the AI currently under development will be anything other than a glorified search and retrieval mechanism. The great Library of Borges. Very useful indeed, but not the world conquering bill of goods being sold by the industry fund raising hype. We are not going to be moving to Mars or building data centers in space. But a low watt high efficiency mind could go places humans will never be able to. The shear level of hype being spewed would make P.T. Barnum blush.
I do think that even a simple edge implementation of open source AI could super charge much of the drudgery of work. But not all work needs to use such power hungry tools. The beauty of all of the evolution of intelligence that you so meticulously document is that nature really is an optimization process. We are making fantastic leaps, but leap too far too fast and you find yourself in an evolutionary dead end before you know how to back out of it. We may have gotten smart, but I don't think the wisdom is keeping pace.
We need to hit the brakes and sit with it for a generation at least and learn how to properly integrate it into our lives. I like to think of AI as a consciousness extender, not a replacement. But human fear and greed will find its own way to throw a speed bump in the road if we can't consciously do it as a species. We may be building to an unplanned pause as nature intervenes to restore the proper rate of progress.
The Borges point is actually the argument in the upcoming Part 8 of this series. Whether LLMs show genuine novel jumps in capability or just smoothly interpolate a very large library is exactly what Wei and Schaeffer are fighting about, and it's unresolved as of that piece. I'd hold the skepticism a little more provisionally than you're holding it, in both directions.
The "nature hits the brakes" part I'm less with you on. Evolution doesn't have a target rate it's restoring, it doesn't have a rate in mind at all, it's just whatever survives. And human institutions have a long, consistent record of not slowing down for anything, wisdom very much included.
Granted, the most common theme in nature is to throw a lot of iterations of a model out with a little chaos or drift in it and see what persists. Humans as a combination of intelligence and lifespan seem to be at a sweet spot in evolution, but our model may soon be challenged. Right now everybody is holding their breath waiting for the chaos and hoping to be part of the group that survives to see what happens. Overall I'm hopeful. I'm also old enough to not worry about how my personal model will be judged.
On somewhat of a side note, attention is magnetic.
For various reasons, mostly revolving around physical survival, I am pretty good at turning down my own inclination to pass judgement, project my own view and just keep my attention as clear as possible and it is definitely a magnetic impulse driving it. Especially with other people, who are also magnetic fields. Not only do I sense them, I see them. Ophthalmology refers to them as "floaters." Those spots in the vision, generally dismissed as optical distortions or illusions. I experience them as other consciousnesses intersecting my own. I have been encountering them since I was a child and the integration and feedback with the environment has only grown more dense and complex over the years.
For example, a floater, glowing slightly red, with a line running through it, drawing it toward you, is the classic heart shape.
I've seen pictures of cave paintings, where the more abstract designs looked like what I experience.
Magnetic there is doing metaphor work rather than physics work, which is most of what this series is about. Hopfield's magnet isn't a claim that brains have fields. It's that one equation describes a magnet and describes his memory, with nothing changed but what the variables refer to. Attention being pulled toward things is real and something explains it, but it won't be magnetism.
The cave painting observation isn't yours alone, incidentally. Lewis-Williams and Dowson argued in 1988 that the abstract geometry in prehistoric rock art matches entoptic forms, things generated inside the visual system rather than seen out in the world. Contested, but it's a live literature.
Electrostatics, for one. Chemical bonding. But polarity isn't rare enough to point at anything in particular: every system with an energy function has it, because that's what an energy function is, some configurations costing more and some less.
Which is the thing about the Hopfield network. There's no force in it at all. The positive and negative couplings are signs on numbers in a sum. Nothing pulls on anything. What looks like attraction is just the arithmetic of which states cost less.
I just finished reading your exhaustive surveys of intelligence, this History and the Thermodynamics of Intelligence with an eye towards how to bake in guardrails to nueromorphic style AI minds. I am somewhat fixated on that type of AI both for its adaptability to real world learning and low power requirements plus the added bonus that it might actually be able to achieve a true type of novel intelligence. Not just the pattern matching lexical structure of current LLMs. I know that simulations of nueromporphic AI minds are being built, but at phenomenal energy costs and they still need to be flushed and rebooted to control drift.
You touched on simulated annealing style learning as a possible substitute for "programming" in ethical guardrails. I think you used the term adaptive style annealing as a method to bring an AI mind up through "childhood" and instill a sense of ethics into it. Not to just program in ethic, but to create a mental landscape where ethical behavior is baked in and just as in the robustness of an annealed coating you create a substrate that repels noise corrosion and still allows useful learning on the part of the AI. Like raising a child in a loving home and teaching it the difference between good and bad. But in this case in a more compressed timeline and at potentially an even deeper level.
Am I understanding this correctly? Or am I hallucinating? My mind is pattern matching on that subject.
I also confess I had to speed read your material. Legal training has me running through massive documents and applying my own type of mental compression. Errors can definitely be introduced that way.
Thank you, Aldous. I hope you've enjoyed my work.
Not hallucinating, though that particular idea isn't actually in either piece. What's there is noise clearing out false memories in Part 2, and annealing as a search strategy in Part 3. Neither one touches ethics or guardrails.
But I get why your mind landed there. There's clearly something in this stuff that feels like it should generalize to raising a mind well rather than just programming rules into it. That instinct is worth chasing on its own, maybe starting with constitutional AI or curriculum learning literature. Different vocabulary, similar spirit to what you're circling.
What's pulling you toward neuromorphic specifically, the power efficiency or the novelty angle?
I’ve also been looking at optical processing from companies like Lightmatter. Classical computing at lower power that might be suitable for robotics.
Computing with light has some appeal. A crystalline computer has a certain aesthetic appeal.
Lightmatter's the real deal, and photonic compute for the linear-algebra-heavy stuff is a genuinely lower-power path. It just doesn't touch your bigger complaint. It changes how cheaply the computing gets done, not what's actually being computed.
The thing pulling me towards neuromorphic is that I don't really see binary implementation of linear algebra being robust enough. Power efficiency aside I really don't think the AI currently under development will be anything other than a glorified search and retrieval mechanism. The great Library of Borges. Very useful indeed, but not the world conquering bill of goods being sold by the industry fund raising hype. We are not going to be moving to Mars or building data centers in space. But a low watt high efficiency mind could go places humans will never be able to. The shear level of hype being spewed would make P.T. Barnum blush.
I do think that even a simple edge implementation of open source AI could super charge much of the drudgery of work. But not all work needs to use such power hungry tools. The beauty of all of the evolution of intelligence that you so meticulously document is that nature really is an optimization process. We are making fantastic leaps, but leap too far too fast and you find yourself in an evolutionary dead end before you know how to back out of it. We may have gotten smart, but I don't think the wisdom is keeping pace.
We need to hit the brakes and sit with it for a generation at least and learn how to properly integrate it into our lives. I like to think of AI as a consciousness extender, not a replacement. But human fear and greed will find its own way to throw a speed bump in the road if we can't consciously do it as a species. We may be building to an unplanned pause as nature intervenes to restore the proper rate of progress.
The Borges point is actually the argument in the upcoming Part 8 of this series. Whether LLMs show genuine novel jumps in capability or just smoothly interpolate a very large library is exactly what Wei and Schaeffer are fighting about, and it's unresolved as of that piece. I'd hold the skepticism a little more provisionally than you're holding it, in both directions.
The "nature hits the brakes" part I'm less with you on. Evolution doesn't have a target rate it's restoring, it doesn't have a rate in mind at all, it's just whatever survives. And human institutions have a long, consistent record of not slowing down for anything, wisdom very much included.
Granted, the most common theme in nature is to throw a lot of iterations of a model out with a little chaos or drift in it and see what persists. Humans as a combination of intelligence and lifespan seem to be at a sweet spot in evolution, but our model may soon be challenged. Right now everybody is holding their breath waiting for the chaos and hoping to be part of the group that survives to see what happens. Overall I'm hopeful. I'm also old enough to not worry about how my personal model will be judged.
On somewhat of a side note, attention is magnetic.
For various reasons, mostly revolving around physical survival, I am pretty good at turning down my own inclination to pass judgement, project my own view and just keep my attention as clear as possible and it is definitely a magnetic impulse driving it. Especially with other people, who are also magnetic fields. Not only do I sense them, I see them. Ophthalmology refers to them as "floaters." Those spots in the vision, generally dismissed as optical distortions or illusions. I experience them as other consciousnesses intersecting my own. I have been encountering them since I was a child and the integration and feedback with the environment has only grown more dense and complex over the years.
For example, a floater, glowing slightly red, with a line running through it, drawing it toward you, is the classic heart shape.
I've seen pictures of cave paintings, where the more abstract designs looked like what I experience.
Quite literally, thought bubbles.
Magnetic there is doing metaphor work rather than physics work, which is most of what this series is about. Hopfield's magnet isn't a claim that brains have fields. It's that one equation describes a magnet and describes his memory, with nothing changed but what the variables refer to. Attention being pulled toward things is real and something explains it, but it won't be magnetism.
The cave painting observation isn't yours alone, incidentally. Lewis-Williams and Dowson argued in 1988 that the abstract geometry in prehistoric rock art matches entoptic forms, things generated inside the visual system rather than seen out in the world. Contested, but it's a live literature.
It might be metaphor, but what other force/dynamic has that distinct attraction/repulsion polarity to it?
Electrostatics, for one. Chemical bonding. But polarity isn't rare enough to point at anything in particular: every system with an energy function has it, because that's what an energy function is, some configurations costing more and some less.
Which is the thing about the Hopfield network. There's no force in it at all. The positive and negative couplings are signs on numbers in a sum. Nothing pulls on anything. What looks like attraction is just the arithmetic of which states cost less.