Long but absolutely beautiful blog review on the biggest wtf of our time: the massive over-investment on the promise of intelligence without understanding what intelligence actually is. I already shared this with my students. Thanks Alberto!
This article does a great job summarizing many things I felt for a long time. I studied neurobiology in college, and became obsessed with ANNs in the late '80s and '90s. It was always clear to me that all ANN efforts were strongly constrained by the use of gradient descent, which has no correlate in biological systems. Even Geoffrey Hinton's forward forward algorithm uses gradients at the local level, and a strictly layered architecture.
I strongly suspect that other learning algorithms are possible. That non-layered architectures with complex node dynamics and interrelationships can provide a qualitatively different behavior than today's increasingly byzantine layered monstrosities.
The field of ANN went up a path that was first indicated by that early simple neuronal "model" (which we understand now to be so simplistic as to stretch the boundaries of the term). It has ascended very far over the decades, but perhaps the peak it aims for is on an entirety different mountain, reached via a path that diverged at the very beginning.
Great piece! I thought it is about the neuron, but it’s not, it is about how to backtrack in technology when strapped to the rocket ship of capital. The development of the thesis is just beautiful.
Great article. Lots of things I would like to explain to my friends but would struggle to do so elegantly. What else would you put in the AI ‘circle’ (or set) other than ML?
This sent me on a reading tangent about symbolic AI. Pretty interesting that although ML is undoubtedly a subset of AI, the other major subset is essentially ‘pretend AI’.
Even more interesting that scientists / engineers seemed to think ‘scaling’ symbolic AI would lead to AGI!
Hmm, symbolic AI is rather "obsolete" (quotation marks needed because it exists in many places and systems, but it's not considered "AI" per se; look up the "AI effect").
Still, people like Gary Marcus champion a neurosymbolic approach toward AGI that tries to get the best of both worlds (symbolism and connectionism), but few people in the industry believe it could work. The reason is that symbolic AI failed in the past, leading to two "AI winters."
However, it's good to know the history of AI, so I encourage you to keep reading! (To answer your first question, many things inside AI are not ML!)
This goes pretty deep. ML is definitely a type of AI but I have developed ML models that no one could reasonably call AI.
Perhaps there is AI, the field that encompasses ML and symbolic systems, and AI, the goal of a truly thinking machine that has been somewhat perverted by the new goals ‘AGI’ and ‘superintelligence’ (right?).
The AI effect is really interesting reading and shows how much marketing defines the linguistics of this.
I think that’s the really interesting niche your page fills: ‘how our society shapes AI’, on top of the far more documented, ‘how AI shapes our society’.
Great essay. I’d like to call your attention to Daniela Rus’s lab at MIT, where they took inspiration from the nervous system of *C. elegans*, to create what they’re calling “liquid neural nets”. The resulting models are vastly more efficient in terms of memory and energy consumption. Here’s an interview in *Quanta* magazine (this interview led a venture capitalist to get in touch with her, and she’s formed a company, Liquid AI, to commercialize the idea.
Rus comes out of the robotics field, where AGI is not a pressing concern (or distraction).
My biggest frustration in AI is a misalignment of incentives. I am personally interested in the science of AI, which means posing a meaningful scientific question on the nature of intelligence. My own which I adopted from researcher Pei Wang at Temple University is: how can an agent arbitrarily adapt to a new environment under insufficient knowledge and resources (I abbreviate AIKR)? Useful evaluation methodology should follow. See https://cis.temple.edu/~pwang/GTI-book/
In the pursuit of AI, the first major difference in motivation is whether you are doing this for science or engineering. Nothing wrong with engineering, but even before ChatGPT research has mostly shifted towards engineering. Of course now it’s all shallow engineering when money is on the line. Caveat: you need science to really push the boundaries of engineering, no transistors without electromagnetism.
I think there is some good connectionist research on the AIKR in the field of continual learning, although this is basically a coincidence: the scientific question was not the premise of this line of research, although I hope it becomes clearer.
Personally I think you are sort of mixing questions when noting the difference between perceptrons and real neurons. I think this confusion is understandable given the explosion of connectionism. Understanding and modeling neurons is an important scientific question applicable to biology and medicine, but it might not tell us much about intelligence. We could simulate an entire brain yet fail to understand anything about how it works at the operational level.
Yeah, I wanted to add a section clarifying this part, that simulation doesn't entail understanding (we can simulate the entire brain of a fruit fly and yet we know nothing of how its behavior emerges nor can we predict it at all).
Great essay Alberto, thanks for taking the time and energy for educating me on important aspects of how and why we got where we are with 'AI' today. I fear the 'unreasonable effectiveness' of LLMs will continue to drive us all forward in the creation on increasingly powerful 'ghosts' .. it's really up to us to not be fooled and mistake them as fellow humans or God forbid... as God.
In addition to reconsidering the complexity of our artificial neuron models, perhaps it would also benefit us to consider reproducing another quality of biologically evolved neurons - the energy efficiency of their computations.
Global energy production is currently plateauing and, by the 30's, will be going into steep decline. The rest of the 21st century is shaping up to be a brutal game of musical chairs.
With this in mind, maybe the usefulness of AI will turn out to be - not the replication of consciousness - rather, as a bridge between consciousnesses? LLM's as an evolution of language - a more nuanced translator between worldviews. Therapist and relationship counsellor to the human species - a loom to re-weave our social fabric before we alienate ourselves to extinction.
What if the neuron model chosen by researchers makes perfect sense for how the see the world? These engineers are entirely satisfied with the behavior or their product and see no downsides. It may not be a confidence that their product thinks like they do (highly abstracted with no grounding in reality and a perfect simulation of emotion rather than emotion itself) and also that the neuron model they chose also appeals to their left brain perspective of the world as a simple machine. They ran with that model and they got what they intended to get. FYI the right hemisphere of the brain has a different architecture than the left and maybe we need a dual hemisphere ai with different architectures.
Superb piece that manages the demanding task of being faithful to two mistresses (Neuroscience and AI). One question would be at what point are we likely to realise that additional spend on say compute leads inexorably to diminishing returns? As you strongly imply there will be compelling reasons to keep the party going for some time. Aldo, on a fundamental level, what is the level of research activity on AI that eschew LLM approaches?
A good article. One subject it didn't get into is whether we need any kind of artificial neuron at all to implement human-level AI (AGI). Although there is a sense of frustration indicated in this article, and in the AI world more generally, that we have been trying to figure out AGI for a long time and haven't made sufficient progress, we haven't really explored algorithm space very much at all. Although it's a simplification, we have really tried only two approaches to get to AGI: logic (resulting in the first and second AI winters) and LLMs or deep learning (possibly resulting in a future third AI winter). It wouldn't surprise me if AGI is a much tougher nut to crack than flying. IMHO, it is way past time to try more stuff.
Long but absolutely beautiful blog review on the biggest wtf of our time: the massive over-investment on the promise of intelligence without understanding what intelligence actually is. I already shared this with my students. Thanks Alberto!
Thank you Simo 🙏🏻 hope they like it!
A+ for you today, Alberto! I will be sharing this essay with my class in Integrative Neuroscience and seminar in Biologically Plausible AI.
That's awesome!!
This article does a great job summarizing many things I felt for a long time. I studied neurobiology in college, and became obsessed with ANNs in the late '80s and '90s. It was always clear to me that all ANN efforts were strongly constrained by the use of gradient descent, which has no correlate in biological systems. Even Geoffrey Hinton's forward forward algorithm uses gradients at the local level, and a strictly layered architecture.
I strongly suspect that other learning algorithms are possible. That non-layered architectures with complex node dynamics and interrelationships can provide a qualitatively different behavior than today's increasingly byzantine layered monstrosities.
The field of ANN went up a path that was first indicated by that early simple neuronal "model" (which we understand now to be so simplistic as to stretch the boundaries of the term). It has ascended very far over the decades, but perhaps the peak it aims for is on an entirety different mountain, reached via a path that diverged at the very beginning.
"perhaps the peak it aims for is on an entirely different mountain" 100%. My entire article in one line, well done haha
Great piece! I thought it is about the neuron, but it’s not, it is about how to backtrack in technology when strapped to the rocket ship of capital. The development of the thesis is just beautiful.
Stunning read. Makes me rethink everything about AI.
Great article. Lots of things I would like to explain to my friends but would struggle to do so elegantly. What else would you put in the AI ‘circle’ (or set) other than ML?
This sent me on a reading tangent about symbolic AI. Pretty interesting that although ML is undoubtedly a subset of AI, the other major subset is essentially ‘pretend AI’.
Even more interesting that scientists / engineers seemed to think ‘scaling’ symbolic AI would lead to AGI!
https://en.wikipedia.org/wiki/Symbolic_artificial_intelligence
Hmm, symbolic AI is rather "obsolete" (quotation marks needed because it exists in many places and systems, but it's not considered "AI" per se; look up the "AI effect").
Still, people like Gary Marcus champion a neurosymbolic approach toward AGI that tries to get the best of both worlds (symbolism and connectionism), but few people in the industry believe it could work. The reason is that symbolic AI failed in the past, leading to two "AI winters."
However, it's good to know the history of AI, so I encourage you to keep reading! (To answer your first question, many things inside AI are not ML!)
What would you say comes under AI that isn’t ML or symbolic AI?
This goes pretty deep. ML is definitely a type of AI but I have developed ML models that no one could reasonably call AI.
Perhaps there is AI, the field that encompasses ML and symbolic systems, and AI, the goal of a truly thinking machine that has been somewhat perverted by the new goals ‘AGI’ and ‘superintelligence’ (right?).
The AI effect is really interesting reading and shows how much marketing defines the linguistics of this.
I think that’s the really interesting niche your page fills: ‘how our society shapes AI’, on top of the far more documented, ‘how AI shapes our society’.
Great essay. I’d like to call your attention to Daniela Rus’s lab at MIT, where they took inspiration from the nervous system of *C. elegans*, to create what they’re calling “liquid neural nets”. The resulting models are vastly more efficient in terms of memory and energy consumption. Here’s an interview in *Quanta* magazine (this interview led a venture capitalist to get in touch with her, and she’s formed a company, Liquid AI, to commercialize the idea.
Rus comes out of the robotics field, where AGI is not a pressing concern (or distraction).
Liquid AI - I've heard it before, maybe a long time ago haha. Will check the links. Thanks!
here are some articles on the C.elegans->machine learning work:
https://www.quantamagazine.org/researchers-discover-a-more-flexible-approach-to-machine-learning-20230207/
https://www.snexplores.org/article/the-brain-of-a-tiny-worm-inspired-a-new-type-of-ai-liquid-neural-network
https://arxiv.org/abs/2006.04439
https://news.mit.edu/2021/machine-learning-adapts-0128
My biggest frustration in AI is a misalignment of incentives. I am personally interested in the science of AI, which means posing a meaningful scientific question on the nature of intelligence. My own which I adopted from researcher Pei Wang at Temple University is: how can an agent arbitrarily adapt to a new environment under insufficient knowledge and resources (I abbreviate AIKR)? Useful evaluation methodology should follow. See https://cis.temple.edu/~pwang/GTI-book/
In the pursuit of AI, the first major difference in motivation is whether you are doing this for science or engineering. Nothing wrong with engineering, but even before ChatGPT research has mostly shifted towards engineering. Of course now it’s all shallow engineering when money is on the line. Caveat: you need science to really push the boundaries of engineering, no transistors without electromagnetism.
I think there is some good connectionist research on the AIKR in the field of continual learning, although this is basically a coincidence: the scientific question was not the premise of this line of research, although I hope it becomes clearer.
Personally I think you are sort of mixing questions when noting the difference between perceptrons and real neurons. I think this confusion is understandable given the explosion of connectionism. Understanding and modeling neurons is an important scientific question applicable to biology and medicine, but it might not tell us much about intelligence. We could simulate an entire brain yet fail to understand anything about how it works at the operational level.
Yeah, I wanted to add a section clarifying this part, that simulation doesn't entail understanding (we can simulate the entire brain of a fruit fly and yet we know nothing of how its behavior emerges nor can we predict it at all).
Great essay Alberto, thanks for taking the time and energy for educating me on important aspects of how and why we got where we are with 'AI' today. I fear the 'unreasonable effectiveness' of LLMs will continue to drive us all forward in the creation on increasingly powerful 'ghosts' .. it's really up to us to not be fooled and mistake them as fellow humans or God forbid... as God.
Keep up the great work!
Thank you Luca 🙏🏻🙏🏻
In addition to reconsidering the complexity of our artificial neuron models, perhaps it would also benefit us to consider reproducing another quality of biologically evolved neurons - the energy efficiency of their computations.
Global energy production is currently plateauing and, by the 30's, will be going into steep decline. The rest of the 21st century is shaping up to be a brutal game of musical chairs.
With this in mind, maybe the usefulness of AI will turn out to be - not the replication of consciousness - rather, as a bridge between consciousnesses? LLM's as an evolution of language - a more nuanced translator between worldviews. Therapist and relationship counsellor to the human species - a loom to re-weave our social fabric before we alienate ourselves to extinction.
Good one. Gary Markus had this x post
“US tech co’s spending half a trillion each year is not bullish… it’s the single greatest exercise of capital destruction in history”
-https://x.com/GaryMarcus/status/1980333714925363307/?s=09&t=7kD7ygeE4z1Ukd-8CJLBEQ&rw_tt_thread=True
It all sound's like trials before Kepler’s orbital discovery
What if the neuron model chosen by researchers makes perfect sense for how the see the world? These engineers are entirely satisfied with the behavior or their product and see no downsides. It may not be a confidence that their product thinks like they do (highly abstracted with no grounding in reality and a perfect simulation of emotion rather than emotion itself) and also that the neuron model they chose also appeals to their left brain perspective of the world as a simple machine. They ran with that model and they got what they intended to get. FYI the right hemisphere of the brain has a different architecture than the left and maybe we need a dual hemisphere ai with different architectures.
Superb essay. You are on a TEAR in rocking my beliefs. It has shifted my thoughts and left me with many more questions to ponder!
Superb piece that manages the demanding task of being faithful to two mistresses (Neuroscience and AI). One question would be at what point are we likely to realise that additional spend on say compute leads inexorably to diminishing returns? As you strongly imply there will be compelling reasons to keep the party going for some time. Aldo, on a fundamental level, what is the level of research activity on AI that eschew LLM approaches?
A good article. One subject it didn't get into is whether we need any kind of artificial neuron at all to implement human-level AI (AGI). Although there is a sense of frustration indicated in this article, and in the AI world more generally, that we have been trying to figure out AGI for a long time and haven't made sufficient progress, we haven't really explored algorithm space very much at all. Although it's a simplification, we have really tried only two approaches to get to AGI: logic (resulting in the first and second AI winters) and LLMs or deep learning (possibly resulting in a future third AI winter). It wouldn't surprise me if AGI is a much tougher nut to crack than flying. IMHO, it is way past time to try more stuff.