Robots Doing Dishes, Or
My Review Of Max Bennett's A Brief History Of Intelligence: Evolution, AI, And The Five Breakthroughs That Made Our Brains
In A Brief History of Intelligence, Max Bennett asks: why don't we have Rosey the robot yet? In The Jetsons, Rosey the robot could play with and tutor children, as well as complete all manner of household chores. In reality, today we have robots that can do back flips and, separately, AI that can pass difficult professional exams, write entire essays, and beat human champions at chess and go. Despite that, we don't have household robots that can do our dishes. Why?
The answer is a lengthy one, because according to Max Bennett we need to understand human intelligence. And to understand human intelligence we need to understand the intelligence of our evolutionary forebears. By understanding biological intelligence and biological learning we can see how artificial intelligence and machine learning compares. Bennett makes the story of the evolution of learning and intelligence accessible by looking at it in terms of five evolutionary breakthroughs that were essential to modern day human intelligence.
The first breakthrough is the ability to steer. 550 million years ago our most distant animal relatives like coral polyps, anemones, and jellies, primarily have radial symmetry and don't steer toward or away from things. Around the same time, the simple nematode worm appears. It has the same kinds of neurons as radially symmetric ones and these neurons operated in largely the same way as modern human neurons do. That was a shocking fact to me. I didn't know that prior to reading this book.
The nematode had significant bodily differences with the radial animals: it had bilateral symmetry (like Bennett, I will refer to them as bilaterians) with a separate mouth and anus. Bennett makes a good argument that bilateral symmetry is why nematodes had another difference: brains. Nematodes needed not only neurons in order to move, but also brains in order to steer. He writes, "All these sensory inputs voting for steering in different directions had to be integrated together in a single place to make a single decision; you can go in only one direction at a time." Interestingly, Bennett points out that the Roomba, the robotic vacuum cleaner, was explicitly modeled after this bilaterian template!
Another shock to me: emotions also play a part in the earliest bilaterian steering. "And so we begin with the simplest two features of emotions, those that are universal not only across human cultures but also across the animal kingdom, those features of emotions that we inherited from the first brains: valence and arousal." Bennett points out that this doesn't mean nematodes were conscious. He explains that he mostly steers clear of consciousness in his book, because it is a complicated philosophical "quagmire". I really appreciated Bennett's candor and ability to be clear about assumptions and boundaries throughout the book.
The second breakthrough is reinforcement learning. Invertebrates share some neurons and neurotransmitters (e.g., serotonin, dopamine) with humans, but do not have the same brain structures. Another thing I found shocking: unlike invertebrates, the very first vertebrates we know about show new brain structures... some of the very same brain structures we find in humans today! Interestingly, dopamine is an essential neurotransmitter in reinforcement learning (specifically, temporal difference reinforcement learning). Relief, disappointment, and curiosity all emerge thanks to new brain structures such as the basal ganglia. These abilities set the stage for learning to be a valuable activity in and of itself.
As neuroscience progressed in the late 20th century, research in AI helped solve problems in neuroscience and vice versa. Neuro-gammon, a software program from the mid-1990s that had a computer backgammon player, was successfully upgraded using a machine version of temporal difference reinforcement learning which allowed it to challenge even the best human backgammon players. Neuroscientists reading the academic paper about Neuro-gammon came to realize that temporal difference reinforcement learning explained how dopamine is utilized in brains in their studies on animal and human behavior! Here, Bennett cautions us to reflect on human behavior, "Gambling and social [media] feeds work by hacking into our five-hundred-million-year-old preference for surprise, producing a maladaptive edge case that evolution has not had time to account for." I found this to have far more gravity in this book than in other places I've read similar opinions, thanks to the scientific context he provides.
In machine learning there are aspects of learning that we don't yet understand how brains handle. For example, pattern recognition, the problem of invariance, the problem of discrimination, and catastrophic forgetting are all things that the human brain handles, but which machine learning experts had to overcome. For example, machine learning techniques such as convolutional hierarchical layers (inspired by studies on cat brains!), and back propogation. In the case of catastrophic forgetting, simply freezing the AI model after it is trained (i.e., not training it anymore), is one way to fix that problem. While this prevents the AI model from learning more, it also prevents it from overwriting existing patterns it already learned, which is one underlying causes of artificial neural network forgetfulness. We still aren't sure how biological brains do all of these things.
Getting back to the evolution of intelligence, the third breakthrough is simulation, which includes imagination, and learning by imagining. The innovation of simulation was seen in the earliest mammals thanks to a new brain structure: the neocortex. Bennett surmises that, "simulating actions is astronomically more computationally expensive and time-consuming than the reinforcement-learning mechanisms in the cortex-basal-ganglia system." And so it makes sense that the neocortex first appears in the warm-blooded mammals whose metabolisms can sustain the extra effort needed. Another fascinating fact: "The only nonmammals that have shown evidence of the ability to simulate actions and plan are birds. And birds are, conspicuously, the only nonmammal species alive today that independently evolved warm-bloodedness." He also says that long range vision could be a precondition as well, because the animals needed more time to think, strategize, and plan.
The Helmholtz machine, first created in 1995, was an early attempt to create neocortex-like unsupervised learning in neural networks, by having a two-way flow so that inputs could go forward through the network as well as backwards. It is the basis of modern generative models like GPT. "Helmholtz suggested that much of human perception is a process of inference - a process of using a generative model to match an inner simulation of the world to the sensory evidence presented."
Bennett continues, "There is an abundance of evidence that the neocortical microcircuit is implementing a generative model. Evidence for this is seen in filling-in one-at-a-time, can't-unsee visual illusions; evidence is seen in the wiring of the neocortex itself [...], and evidence is seen in the surprising symmetry - the ironclad inseparability - between perception and imagination that is found in both generative models and the neocortex [...] it also explains why humans succumb to hallucinations, why we dream and sleep, and even the inner workings of imagination itself." He then delves into details of some compelling evidence, and shows again that mammals and birds share in these characteristics.
Aside from impressive abilities to recognize things, neocortexes additionally enabled mammals to simulate trial and error vicariously, consider counter-factuals, and create episodic memories. As mentioned before, reinforcement learning emerged in early vertebrates. It enabled fast, but inflexible decisions based on "direct associations between current state and the best actions." This is actually known as model-free reinforcement learning. Mammals retain this ability, but the neocortex allowed our lineage to additionally learn through model-based reinforcement. This is slower, but much more flexible. It's based on "a model of how actions affect the world and uses this to simulate different actions before choosing." Model-based reinforcement learning has proven difficult to implement in machine learning. Bennett cites two reasons for this: "building a model of the world is hard", and "choosing what to simulate is hard". This includes the so-called "search problem". As Bennett puts it, "In most real-world situations, it is impossible to search through all possible options." How do humans perform such a search? We actually don't know.
AlphaZero, the AI famous for beating human chess and go champions, "used search not to logically consider all future possibilities (something that is impossible in most situations) but to simply verify and expand on the hunches that an actor-critic system [model-free temporal difference reinforcement learning] was already producing. [...] this approach, in principle, may have parallels to how mammals navigate the search problem."
"While AlphaZero was a huge leap forward, AI systems are still far from performing planning in environments with a continuous space of actions, incomplete information about the world, and complex rewards [i.e., the real world]." Perhaps more importantly, its search strategy was fixed. It applies the same search strategy to every move it makes. Mammals can employ different strategies depending on the context of the situation. It's unlikely that mammals have a significantly better search algorithm than AlphaZero. Rather, we have more strategies to choose from and can recognize when to use them.
In mammals, the frontal neocortex models the self, while the sensory neocortex models the world. The frontal neocortex model explains one's own behavior. That might make the search problem tractable by allowing mammals to "choose when to simulate things and how to select what to simulate."
The fourth breakthrough is mentalizing and appears in the first primates. Why do humans have such big brains? The social-brain hypothesis is one possibility and this hypothesis comes from the fact that primates are capable of developing a theory of mind. That is, a theory about the minds of others. The gPFC, a new brain structure that first appears in primates, seems to be what allows us to simulate ourselves. Human patients with damage to this part of the brain tend to have difficulty inserting themselves into their own stories and imaginings, and can even find it impossible to recognize themselves in a mirror.
Mirror neurons in the premotor cortex of primates fire when the primates watch *other* primates do something, suggesting a kind of mimicry through simulation happening. Experiments show that the premotor cortex is necessary for learning through imitation, specifically by simulating oneself doing the actions of another. This allows primates (and some birds) to acquire novel skills, not only known skills. Bennett explains that many mammals, octopuses, fish, and reptiles can only learn known skills. In machine learning, learning using a theory of mind (specifically, working from intent first, rather than pure mimicry), is called inverse reinforcement learning.
The fifth breakthrough is language and appears in humans and possibly some of our most recent ancestors. While some animals can learn simplistic naming, like chimps that can use limited forms of sign language, only humans are known to broadly name things and create and use languages with elaborate grammar.
The human mind has more than 80 billion neurons and hundreds of trillions of connections. Unlike computers which shuffle around electrons, neurons interact with dozens of different chemicals. Fascinatingly, according to Bennett, the Singularity, the point in history where intelligence explodes in exponential progress, usually referring to the birth of artificial super intelligence, has already happened. Just not in artificial intelligence, but in biological intelligence! Thanks to language, humans can collect and store knowledge and pass it down for future generations to learn. This is something that no other animal can do and underlies the exponential pace of human technological progress. Hence, why Bennett believes the Singularity has already occurred.
I think at this point in his book, near its end, he gets a little loose with his proclamations. In particular, he thinks that humanity's ability to cooperate via common myths can scale infinitely and also that humans don't have to relearn everything each generation. I believe that his comments here are discounting a lot. For example, what about the difficulty humans have in coordinating our vast numbers on the planet today? Or how about the limitations of how much humans can learn in a single generation? But to his credit he doesn't make wild assertions about the future of humanity and I could be reading too much into his hopeful comments here.
Stepping back to compare human learning to the state of the art machine learning of 2023 (when the book was published), Bennett writes, "The foundation of language learning is not sequence learning but the tethering of symbols to components of a child's already present inner simulation". This is an astute observation of one way that LLMs very significantly differ from humans. Such differences could be the keys to answering why we don't yet have robots that can do our dishes. Bennett's book shines with clarity in its many astute observations. The context he provides throughout is indispensable, even when he's just describing how parents teach their children their first words.
In my opinion, he makes an error when he writes, "Life on Earth is only four billion years old. It will be another seven billion years before our sun dies. And thus life, at least on Earth, has another seven or so billion years to tinker with new biological forms of intelligence." My understanding is that the sun will turn into a red giant long before it dies. When it does that it will expand in size and, even before it engulfs Earth completely, it will boil away the oceans and strip Earth of its atmosphere. Earth's habitability probably has closer to one and a half to two billion years left, meaning we are more than halfway through Earth's period of habitability.
However, that doesn't really change his point. As Bennett says, we don't know what the sixth breakthrough in intelligence will be. Perhaps it will be something biological or perhaps not. Perhaps it will be more like an artificial super intelligence. Hopefully, we'll have two billion years to find out. Preferably with robots that can do my dishes a bit earlier than that.