How Not to Create a Mind

I had heard the name Ray Kurzweil before, and when I was researching material to read on Artificial Intelligence his recent book How to Create a Mind: The Secret of Human Thought Revealed came up several times. Naturally, the title intrigued me, and I dug deeper into who Ray Kurzweil was; I was impressed with much of his biography. Ray is a computer scientist/inventor who has founded several companies and led the early advances in Artificial Intelligence. He is widely regarded as a “genius”, holding several honorary degrees, patents, and other honors and medals, including being inducted into the “National Inventors Hall of Fame”. His background in AI, combined with his entrepreneurial tint, resonated deeply with me, and I immediately purchased a copy of his book, beyond excited to see what “secrets'' it would reveal. However, it quickly becomes clear that Kurzweil offers very little insight into how we can build Artificial General Intelligence; the book contains a hodgepodge of supposedly relevant information without a coherent narrative, or even a strong, innovative theory. The book unequivocally fails on its promise. Kurzweil’s supposed contribution in this book is to present his Pattern Recognition Theory of Mind (PRTM) as the way in which humans think and represent information. To start, knowledge that the brain acts as a pattern recognizer is well established. Every undergraduate in their first year learns this in PSYCH 101. Kurzweil proposes that our neocortex, the outermost layer of our brain, and the most recent in our evolutionary development, is the area that is responsible for this pattern recognition. He claims that the neocortex is made up of repeating structures of “pattern” recognizers. To support this, he offers two recent scientific studies and makes no effort to show the readers any other pieces of conflicting evidence. In fact, there is still much we do not know about how the neocortex is structured. Based on his PRTM, Kurzweil estimates that the brain has 300 million pattern recognizers. Each one recognizes one pattern, and will communicate with the other pattern recognizers when it gathers enough evidence to support that the pattern exists. Nowhere does he give a proper definition of what he considers a “pattern”, and it is up to the reader to infer that based on examples. His connection between pattern recognition and thought is wildly incomplete; how does one go from mere recognition of patterns to full-blown thoughts in a conscious mind? He goes on to make wild conjectures about how the brain could be connected to the cloud, giving us more “pattern recognizers” and thus increasing our processing power and memory. This is a tempting idea, but Kurzweil offers zero advice on how this would actually be implemented. The reader is expected to simply agree with Kurzweil that his theory must be correct and that all potential corollaries of the theory hold true. I found his connections between memories and pattern recognition to be even less flushed out. He claims that our brain, which he acknowledges is a highly parallel processor, stores memories in ordered lists. He thinks that memories themselves are a series of patterns, but how exactly that is the case is unclear. If patterns are to be so generic as to merely mean a “representation of something” then of course memories would be patterns; this would not be new. But the implication in Kurzweil tone and style is that he has stumbled and reached a novel theory, that his understanding of patterns and memories is somehow profound. It is hard-pressed for the reader to know why. The parts of the book that I did find interesting are Ray’s forays into computer science, and his explanations of how Watson, von Neumann machines, and Turing machines work. He tries to show how hierarchical pattern recognizers, embodied in computer software, could mimic the human brain in thought, and asks us to “take a leap of faith” that a computer software that acts exactly like a human does would necessarily be conscious. However, he gives very little support for those reasonings. Firstly, simply because we can build hierarchical pattern recognizers in machines does not mean we have mimicked human thought (or consciousness). We have merely SIMULATED one ASPECT of human computational power; it is NOT duplication. Secondly, as we had already discussed, even if it were a duplication, Kurzweil pattern recognition theory is weak at best. Just because the mathematical structures work so well for computers,does not mean that this is how the brain works, or how a conscious mind could be created. Kurzweil’s attempts to talk philosophy of mind and consciousness also embarrassingly miss the mark. He thinks that “understanding” could be created through higher and higher hierarchical pattern recognizers, coding pattern recognizers that include semantic meaning. However, simply having recognizers for semantic meaning is far from creating a generally intelligent//conscious mind; we already have many NLP algorithms that can pick up on semantic meaning, with layers to represent them (I am working on some side projects of my own to do this). He is a reductionist, and thinks that having a layer, or representation of understanding is enough for conscious minds. Fair enough, but this is such a haphazard attempt at reductionism as to barely be worth refuting. The jump from hierarchical understanding of semantics to a thinking, understanding thing, is massive and while I think it can in principle be bridged, Kurzweil does not show us how. Thus, while the technological portions of Kurzweil’s book are fascinating, he is more focused with describing how they work rather than showing how these concepts could be realistically linked to biological brains and human thought. His book appears to be about describing the PROCESS of thought, but not how thought manifests itself and shows itself to conscious beings. Kurzweil does not describe the user interface, only the code of the program. His chapter “Thought experiments on the mind” was perhaps the most difficult to read; he introduces famous debates on Free Will, identity, and split brains but offers nothing new beyond recounting what those debates are and drawing weak connections to his PRTM. My frustrations with the book aside, the time spent reading it was not a complete waste. There are a few nuggets of interesting information and ideas that I was able to extrapolate. One is the idea that we should think about understanding the brain at the right level of computation. Kurzweil is right in this; if we sought to simulate the brain by enacting every single neuron’s activity at all times, this would be a fruitless endeavor. Instead, we should try to go up a few levels of understanding, closer to the computation//algorithmic level of what the brain is doing and simulate that. Of course, Kurzweil’s PRTM is far too high level, and is not sufficiently close to what we know about neuroscience today to serve as our candidate explanatory model. Second, Kurzweil makes some good points about what the CONTENTS of our conscious experiences are. He notes that we are often not aware of all the contents of our experience, and experiments have shown again and again that we confabulate our experiences to make better sense of them in our head (so that they can fit a more cohesive narrative). His remarks on how the entirety of conscious experience is only what is reportable are well-taken. Lastly, while Kurzweil’s lack of attention to providing the reader with how exactly our minds will come to be augmented by cloud computing, this general idea is an interesting one. There is much more research that needs to be done (Neuralink!) but I am excited to see where it goes, and this perhaps is the only real new insight I gleaned from reading this book.