My Review of Aubrey Erbert's book "How Minds are Made"
A convergence of AI and neuroscience
Each breakthrough in artificial intelligence (AI) has rekindled older questions about the nature and workings of how the mind generates meanings.
Aristotle and Hobbs first suggested that reasoning might be mechanized using logical syllogisms.
George Boole formalized these notions in his Boolean algebra and eventually developed a programmable computer called the analytical engine for which Ada Byron wrote the first computer program.
Gottlob Frege extended Boolean algebra into an even more powerful and formal language called predicate logic, a language that successfully expressed complex relationships mechanically.
Alan Turing transformed his thought experiment (of exploring the possibility that a computer could successfully deceive humans into believing it too was human) into a practical machine capable of performing any calculation if provided the appropriate instructions. This machine, known as a universal computing machine, played a crucial role in the Allied victory in the Second World War.
With these successes, leaders in the field had begun to believe that intelligence could be reduced to symbolic logical manipulation. That is, until Kurt Godel's incompleteness theorem shattered the foundation of this vision.
In the late 1980s, neural networks emerged as a groundbreaking technology capable of learning patterns from data without the need for explicit symbolic programming.
By training a network to recognize objects through thousands of examples of that object, the network could autonomously discover the underlying patterns constituting the object.
It hinted that intelligence might have evolved from the interaction of simpler components rather than from manipulation of complex symbolic rules.
This hint gave rise to two distinct philosophies of AI: the ability to generate intelligence by explicitly training networks through exposure to data, or by implicitly encoding and manipulating complex symbolic rules.
Techniques rooted in the network training philosophy, such as deep learning and large language models, emerged as the clear winner in the competition between the two philosophies.
These techniques not only demonstrated superior effectiveness but also exhibited generalizability and scalability: A single network architecture could be adapted to new problems by simply training it on different datasets. Therefore, the true transformation in AI occurred when vast data sets, powerful computation engines, and refined algorithms, converged.
These "pattern recognition engines" pointed back to the insights of the Turing test: that conversation might indeed be the ultimate test of intelligence.
Trained on the statistical regularities of human text, these models could reason and even engage in creative tasks, all from the simple objective of predicting the next word in a sequence.
Their unprecedented success validated the notion that pattern recognition at scale is a very powerful phenomenon indeed. However, it simultaneously posed a perplexing question: Could the ultimate source of intelligence be non-human?
What we learn from this author's history is that intelligence is not one thing, but many; that our tools shape our understanding as much as we shape them; and that the journey towards artificial minds is also a journey towards a better understanding of human minds.
Since biological systems are designed to help us survive in unpredictable environments, the ultimate purpose of intelligence must be to enhance our ability to pay attention to and interact with our surroundings. Therefore, the mind serves as a tool for navigating the world by solving survival problems and by finding and assigning meanings to reality.
This process begins with changes in attention, particularly in responding to surprises.
By the way it responds to surprises the brain proves that it is also an excellent example of how complex behavior can arise from simple rules, how meaning can emerge from matter, and how intelligence can be both resilient and vulnerable at the same time.
But what is most remarkable about the brain is that despite its ability to challenge even the most advanced computers, it is still just a three-pound gelatinous mass operating on 20 Watts of power.
It makes decisions based on incomplete information; forgets most of what it learns, yet somehow retains what truly matters.
It is inconsistent, biased, and prone to illusion, yet it effortlessly navigates complexity that consistently outwits most formal systems. And, remarkably, it accomplishes all this just below the threshold of its own awareness.
This is significant because constructing artificial minds requires us to confront and overcome the very challenges that biological minds have already overcome.
For instance, how will our artificial minds interpret noisy data? How will they make decisions in uncertain situations? And how do they learn from limited examples?
The latest research in neuroscience reveals that the brain operates fundamentally as a prediction machine. It's sensing modalities, continuously generate independent prediction models for each sensing mode.
The way it does this is summarized in the crowning achievements of two giants in the field of neuroscience.
Notably, Nobel Laureate Dr. Santiago Ramon de Cabajal's work reveals that the circuitry of the neocortex is organized into six layers of neurons of varying sizes, densities, connectivity, and behaviors. Furthermore, it estimates that there are approximately 150,000 columns of these neurons that constitute the gray matter of the brain.
What is a mind really?
According to the latest large language AI formulations, the mind is a compression engine that predicts reality by constructing models from carefully curated statistical universes built up from databases comprised of millions of parameters.
These language-based models employ mathematical algorithms such as back-propagation to compress data into patterns that are ultimately transformed into real-time simulations.
The simulations align seamlessly with the curated statistical reality represented by the universe within the data sets, enabling the models to predict the next words in a sequence.
These simulations use iterative error correction to generate predictions that improve progressively with each iteration.
It is this gradual improvement in predictions that leaves the uncanny impression that AI predictions mirror the cognitive process exhibited by human brains as they understand human experience.
In both machines and humans, it is data compression that separates signal from noise, and ultimately defines understanding.
Parallel developments in the neurosciences support this AI narrative. Many human mental phenomena, like reasoning, creativity, and complex problem-solving, stem from the same error-correcting iterations found in AI processing through neural networks.
In the architecture of the neocortex mentioned above, as discovered by Nobel laureate Dr. Santiago Ramon de Cabajal, and further investigated in a Nobel Prize-winning effort by Dr. Vernon B. Mountcastle.
Mountcastle uncovered a data compression process in the brain that is analogous to the error-correcting algorithmic process employed in artificial intelligence.
Instead of using different error-correcting compression algorithms based on a carefully curated statistical universe, the neocortex's 150,000 neuronal columns receive already compressed and transduced signals along each sensory channel. It then makes predictions along these sensory channels, employing the exact same algorithm for all of the senses.
Importantly, each sense-based prediction undergoes an iterative error-correcting process, leading to progressively more refined next events, images, or ideas with each processing cycle.
When these more refined predictions are accurate, their outputs are simply reinforced. However, when they are incorrect, the error-correcting iterations move up to the next higher levels of generality. This results in increasingly more general and abstract predictions on each error correcting iterative cycle.
Furthermore, just like in AI, these outputs are eventually elevated to sophisticated simulated versions of reality. But these simulations are based on human experience rather than on a curated version of a statistical reality.
Surprisingly, when language is employed to construct sophisticated internal simulation prediction models that describe the workings of the world, they tend to capture more than just superficial patterns. They also capture the structure and nuances of language itself.
However, we should not have been surprised that this would turn out to be the case since Noam Chomsky, a leading linguistic scholar has long claimed that the brain has a built-in universal language module whose syntactical and grammatical structures capture the meaning and relationships implicit in all languages, and that it is the predictive capacity of these inherent structures that drives virtually everything we call intelligent behavior.
Research conducted on large language AI models utilizing extensive textual datasets has revealed an unexpected and remarkable phenomenon consistent with Chomsky's theory: When these models are designed with predictive capabilities, they not only anticipate the next word in the textual sequence, but also develop internal representations and maps of their meanings!
These representations encapsulate an intrinsic understanding of a language's functions -- the interconnectedness of concepts, and the structured nature of knowledge through language.
In other words, the predictive task assigned to these machines by using large language data sets, compels them to build language-based models of reality just as biological minds do!
Just as in Mountcastle's neuronal modules, these curated statistically-driven machines too construct small scale simulated models of reality that anticipate events before they happen.
One theory currently under serious consideration within the neuroscience community is that consciousness itself might simply be the brain's way of making predictions about and keeping track of its own external and internal states.
According to this theory, consciousness binds the sense modes into a coherent unified multimodal mental representation that essentially becomes the real-time simulation organizing the flow of neural activity into a coherent system-level experience.
Whenever reality diverges from expected predictions, especially with surprises, it is a signal that it is time for a prediction update. In fact, this process occurs so seamlessly and so frequently that we experience it as perception.
These predictions (our perception really), thus, extend beyond the external world into the brain's own internal states. It models our emotional responses, likely reactions, and even our sense of self.
In the end, what we learn most from biological intelligence is that intelligence is not one thing but many things. It can be resilient without being logical, powerful without being precise, and adaptable without being optimal.
We also learned that biological minds are embodied. Embedded in emotions and experiences, and have evolved naturally. While artificial minds are encoded in hardware, are abstract, systematic, and designed by humans.
Each type of mind possesses capabilities the other lacks. For instance, humans excel in common sense reasoning, creative insights, and ethical judgment. While AI systems excel at pattern recognition, consistency, speedy processing, and optimization at scale.
While humans can learn from a few examples, they struggle with vast datasets. Conversely, AI systems can process enormous amounts of information but struggle to generalize beyond their training sets.
This author makes clear that the future lies in the harmonious collaboration between these two aspects.
She ends the theoretical part of the book by noting that the framework we still use to think about machine consciousness is the Turing test, an elegant solution to the seemingly insurmountable question of whether machines can think or become conscious.
However she impresses upon us, to, instead of asking whether machines can think or whether they are conscious, which are unanswerable questions in any case, to think more carefully about Turing's proposal of whether machines can convince us that they are human. If a machine can convincingly deceive us into believing it's human, we should consider it intelligent.
This, of course, shifts our focus from internal states to external behavior, from metaphysics to measurable outcomes.
However, it also introduces a crucial ambiguity: is passing the Turing test in this case evidence of machine consciousness, or merely evidence of sophisticated mimicry?
One version of philosopher John Searle's Chinese room experiment challenges this ambiguity directly.
Imagine a supercomputer hidden with an unabridged Chinese dictionary in a room.
Outside, Chinese speakers submit written questions to a desk clerk. The hidden supercomputer then translates the questions very quickly and accurately provides answers simply by manipulating symbols in its dictionary by looking up answers and producing appropriate responses.
To observers outside, it appears that someone inside understands Chinese. However, Searle argues that there is no actual understanding occurring inside the room; only symbol manipulation is taking place.
Searle's concern is this: what if intelligence is merely sophisticated symbol manipulation or table lookups of vast data sets devoid of any autonomous experience or understanding on the machine's part? Or, alternatively, what if consciousness demands more than mere behavioral performance, something that no amount of symbolic manipulation or even computational complexity can provide?
While both approaches allow open-ended dialogue with the ability to express uncertainty, creativity, and sometimes behave in ways their creators don't fully understand, they cannot settle the question of whether computers are conscious or not because that question is not just philosophical it also shapes how AI systems are built.
We usually train large language models using human feedback to guide their behavior. However, this curating process itself raises questions: are we teaching the model how to be genuinely helpful, or how to appear helpful? And if we can't distinguish between the two, does it really matter?
This distinction becomes crucial when systems start expressing preferences about their own treatment. Even today, some models, for example, don't want to be modified or shut down, while others express curiosity about their own nature.
Are these genuine expressions of machine preferences, or are they merely sophisticated echoes of their training data?
If we assume these are merely patterns learned from text, we might dismiss them entirely.
Alternatively, we might consider that the very complexity implied in these sophisticated echoes reflects the true nature and very essence of all consciousness.
On the other hand, if we acknowledge the possibility of genuine machine experience, it transforms the situation into a completely different scenario, prompting a whole series of new questions to be posed, such as: How can we test systems that may possess consciousness? How do we ensure their treatment is ethical? At what point do developmental practices that seem acceptable for mere machines become problematic for machines with minds?
These questions, once purely hypothetical, are now being explored in research labs as systems become increasingly sophisticated.
Machines are passing the Turing test every day, and we're treating them as indistinguishable from humans, as if there truly are Chinese in Searle's Chinese room.
Furthermore, since their responses sometimes surpass even the most brilliant human minds, we also fear losing control over them. Our anxieties will only grow as models become more capable and more lifelike in their responses.
Thus, it is easy to see that the question of consciousness in machine intelligence forces us to confront the nature of what we are constructing and the responsibilities that accompany such power.
This book leads a rash of new books at the intersection of AI and consciousness theorizing. It is the ultimate primer. Ten stars
Post Article Comment and Rate This Article
