In this latest offering, the renowned futurist revisits his groundbreaking concept of technological singularity. However, the eagerly anticipated updaIn this latest offering, the renowned futurist revisits his groundbreaking concept of technological singularity. However, the eagerly anticipated update to the nearly 20-year-old seminal work falls short of expectations over and above the revised date when he expects Singularity to arrive. Instead, the book feels hastily assembled. Absolutely, the new date – before the end of this decade – will likely make this book a must-read piece of work for many, but it is likely that most, like this reviewer, will walk away not much smarter.
The book's primary shortcoming is its lack of fresh insights. Rather than delving deeper into the evolving landscape of technology and its implications for the singularity, it rehashes familiar territory. The author misses a golden opportunity to provide better justifications for why he expects machines to be better than humans in almost all aspects by 2029 and not 2045. More importantly, the book fails to discuss the implications of machines working on themselves.
At the least, the update book should have re-examined the core concepts of singularity in light of the vast amount of new information available since the original publication. A glaring omission is the lack of discussion on recent technological breakthroughs. The book overlooks innovations in mobile telephony and social media that were not expected in the first work, which is perhaps ok, but also highly topic-relevant developments in deep tech. Notably absent is any meaningful exploration of neural networks, including RNNs, CNNs, and the game-changing advent of transformer and post-transformer technologies. These advancements have profound implications for machine intelligence, intentionality, and purpose-driven AI – topics that are supposed to be what the book is all about. Instead, the book veers into well-trodden territory, offering a broad overview of technological progress over centuries and projecting exponential growth into the future—a topic extensively covered in numerous other books, TED Talks, and industry reports.
A significant portion of the book is devoted to societal progress, summarizing work better articulated by other authors like Steven Pinker. While interesting, the author's optimism about technology's impact on employment and his speculations about future innovations across various fields don't offer much novelty either.
On the positive side, the author's unwavering optimism and recounting of technological advancements do provide some valuable insights. His ability to picture potential future developments across various sectors is commendable, even if not groundbreaking.
Overall, the book may become the book of the season for most readers, but it serves more as a general recap of well-covered subjects than a pioneering work like its predecessor.
As one of AI's pioneering founding mothers, Dr. Fei-Fei Li is destined for the history books. While shaping the AI's true history-defining years of thAs one of AI's pioneering founding mothers, Dr. Fei-Fei Li is destined for the history books. While shaping the AI's true history-defining years of the first quarter of the twenty-first century, she had a unique vantage point to narrate the evolution that has now turned into a revolution. The chronicle of the neural network renaissance sketched in the book should prove a trove for future scholars tracing AI's origins. Alas, this book's desultory storytelling does not do justice to the topic. It will go down as not only incomplete but also inadequate in the goodness of time.
The book is best when it captures the emergence of Convolutional Neural Networks (CNNs) and then Recurrent Neural Networks (RNNs) awakened from the field's winter slumber, breaking past roadblocks in ImageNet and WordNet. The author superbly shows the critical role played by her teams and their heroic labors in generating vast human-curated datasets. For some reason, the author shies away from more detailed technical and technological factors, which would have made the book a far greater source of historical importance.
The author's weaving of this with her personal narrative does not work. The book overreaches in painting personal setbacks as the stuff of Greek tragedy. As much as the author tries, her story is not of an extreme underdog. This person is blessed with great intellect and has always been a part of the world's best institutions. She was rarely hindered because of her gender or ethnicity. Yes, she has her own occasional setbacks and personal tragedies, but not of the kind that make compelling narratives. Like her Congress appearance, the endeavor to make everything a rough climb has almost anecdote ending with a damp squib. They serve occasional purposes in providing a base for some of the author's ethical AI pursuits, but once again, not of a kind that will set either the goals or the work apart from those of many others.
The biggest weakness is where the book stops - just at the time of the arrival of the LLMs, which, as we know it today, dwarfs all that came before. The book is akin to an early 20th-century treatise on electromagnetism that stops at 1905 and Einstein's emergence in physics. The book excels in showcasing neural networks' sporadic successes and pitfalls before taking absolute center stage, but it still appears like the tales of early Korolev rockets before Apollo's glories.
Before the review, it is important to mention that this reviewer has rarely agreed with any books more than this one on almost any topic. I have writtBefore the review, it is important to mention that this reviewer has rarely agreed with any books more than this one on almost any topic. I have written on at least a dozen topics covered in the book in various past reviews and my LinkedIn posts with almost identical conclusions on AI's impact. Now on to the long review:
Mustafa Suleyman's The Coming Wave offers a compelling narrative of the promises and perils of artificial intelligence, emphasizing the urgency of collective action to mitigate its risks before they spiral out of control. The book excels in illustrating the far-reaching impacts of AI, although these sections are short and sporadic. It falls short in its central theme of potential solutions.
As the founder of DeepMind, Suleyman provides an insider's, authoritative perspective on the recent advances. While AI optimism pervades the book, with the topic being its risks, these sections' message on transformative power is often too hurried.
That said, the book convincingly argues that we have moved beyond the rudimentary applications of AI, such as chatbots and photo editing tools, to a new era where machines can think on their own. Techniques like deep learning and transformers have enabled AI to tackle complex real-world tasks like protein folding, autonomous driving, and understanding of human languages that were far beyond their capabilities until now. The author effectively argues these are not incremental advances but a paradigm shift – AI can now learn and reason independently in ways that were unimaginable even a few years or even quarters ago. Many who feel AI has been around for years completely miss this point: various giant thresholds have been crossed, and we all need to think about AI anew to appreciate its impact from hereon truly.
The main positive effect of AI is how it has begun to turbocharge innovation across sectors globally. It can deal with the complexity of orders higher than ours and exponentially rising. In this reviewer's language, we have been tackling all the life's NP-hard problems at a particular level defined by the limitations of our 100W-powered biological neural networks, aka brains. Our solutions have been the best so far, as nothing in our toolkit could process the most elementary of our level complexity issues, let alone anything higher. This last fact is no longer true.
In other words, machines of higher capabilities will have repeated relooks at the complicated problems of all life's domains, with the promises of solutions that will revolutionize fields as far away as synthetic biology and robotics to drug discovery and new chemicals and everything in between including quantum computing, battery alternatives, superconducting candidates, etc. Whether linked to machine vision or solving micro-scale climate tech issues, supply chains, policy regulations, and finance - none of our domains will likely remain untouched. More importantly, the changes will be hyper-paced, with the risks of better technologies around the corner forever even before any new solution settles, all accentuated by these technologies building on each other, with quantum computing plus AI an obvious hyperscaling candidate.
This ability to acquire and deploy knowledge at a more complex level makes AI a universally applicable technology. Suleyman illustrates this through examples like AlphaFold cracking protein folding, then showing how the same system can learn physics or math.
On the one side, The Coming Wave refreshingly looks beyond common Silicon Valley chat/edit examples of AI's emerging impact. On the other, the book's focus is squarely on risks. The author balances every positive prognostication with a historian's lens, cautioning against blind techno-optimism. He convincingly argues against optimists who ignore logical risks through the simplistic interpretation of history based on a small number of data points, most famously those who love to site a Malthus or a Ludd.
The author's starting point is simple: AI should not be contained, given the positives, but it cannot be contained given the distance we have already covered in our world of competitive nations, corporates, and ego-driven people. Paradoxically, his eventual suggested solutions are too idealistic and rhetorical to have any chance of succeeding in a world where no two central authorities have the same moral, ethical, political, or legal framework, not just across nations but even within the same nation between rival parties.
To this reviewer, the author should have focussed on the best defensive strategies given his technological skillsets and understanding. Good, conscientious folks should continuously work on mitigating risks by building rival technologies - like anti-virus or anti-missile technologies - that perpetually monitor early signs of malfeasance. Yes, policymakers, too, must come together to do their best on global guardrails, but it is unlikely high-level global agreements can prevent much of anything the author warns about through the book.
Other societal risks, like technological unemployment, are more important, but they do not get the same treatment as risks of malintent or machine errors. Most discussions on these topics, including issues like bias, transparency, and privacy invasions, have little freshness.
In fairness, few books could offer definitive solutions to challenges this enormous and complex. If anything, this ambiguity leaves readers recognizing their own role and agency in shaping the AI future. The book succeeds most in giving readers a conceptual foundation to wrestle with the coming wave.
The book dwells heavily on historical examples to frame AI. While partly instructive for context, extended discussions about innovations like electricity feel dated and distracting compared to the remarkable technological forces described elsewhere. More perspective from the bleeding edge of research could have reinforced the book's vital message around AI's transformational potential.
The discussions on the complexity above are my own, although they mirror those of the author's. Before leaving the review, let me add some more personal views on the same topic, although this has no connections with anything in the book. The book does not recognize that we have likely built a completely different form of intelligence using the latest neural network methods, one that is likely far superior to human intelligence and constantly improving. This new intelligence may or may not beat The Turing Test. We may constantly encounter examples where humans do things better or differently, but this will not contradict our machines' new ability to deal with higher forms of complexity.
It is the machines' ability to deal with a higher level of complexity that is revolutionary and transformative. Ever since Turing, the arguments that machines would need to mimic human behavior to be considered intelligent have been fundamentally flawed and, at times, regressive. We don't need to understand a dolphin's language to know that we are more intelligent, and similarly, machines don't need to mimic us to surpass us in intelligence.
Lastly, through unsupervised learning, machines create their own languages, classifications, tags, etc., to analyze structures like genes, proteins, vision, and everything else. This allows them to approach various scientific queries in ways our languages, including coding languages, were incapable of. For example, deciphering a gene would never be possible using our language dictionaries and any human-driven categorization. The same happens in LLMs through machines' indecipherable ways of dealing with tokens and converting them into usable symbols. This is another new methodical change with far-reaching implications everywhere....more
The Self-Assembling Brain is a fascinating examination of the intersection between neurobiology and artificial intelligence. As the title suggests, thThe Self-Assembling Brain is a fascinating examination of the intersection between neurobiology and artificial intelligence. As the title suggests, the author Jonas Hielsinger posits that the brain - let's call it BNN or biological neural network for this review - is a self-assembling system with simple low-level rules resulting in incredibly complex high-level behaviors and cognition. Through dense yet lucid descriptions of cutting-edge research in neuroscience, the author makes the case that understanding how the brain wires itself may hold the key to advancing AI or artificial neural networks, henceforth called ANNs - again for this review.
The core argument underpinning the book is that neurobiology and AI are deeply intertwined fields with much to learn from each other. It emphasizes the need for collaboration between experts across disciplines to unlock the secrets of ANNs and BNNs. In some ways, the author's views are too biased toward the potential payoff of connections between the fields. The book could have benefited from more focus on the enormous divergence that has grown between these fields by now, but it still does not take away anything from the enormous value it provides, regardless.
The book truly shines while discussing the details of neuroscience. A particular highlight is the in-depth discussions of how simple local learning rules, evolved over millions of years, lead to the complex phenomena we associate with cognition and consciousness. Take language acquisition as an example – babies are not explicitly programmed with grammatical rules but rather absorb the statistical regularities in the speech patterns around them. The brain, a BNN, is wired to detect and internalize these regularities through brute repetition, unlike how we train ANNs these days.
The book illustrates this and other similar concepts through clever hypothetical dialogues between experts at the start of each chapter. In one exchange, an AI researcher presses a neuroscientist on how children acquire language without direct instruction. The neuroscientist explains how the rapid formation and pruning of neural connections allow the BNN to build statistical models reflecting the environment. While fictional, these dialogues neatly encapsulate the core themes around self-assembly and help make the later technical sections more intuitive.
An early section analyzes systems like our BNNs that are fundamentally unpredictable despite relying on simple deterministic rules. And, then, there is the reverse. Networks of neurons in lower-level areas operate largely randomly at an individual level yet produce reliable signals when aggregated. Out of disorder emerges order. The book covers the opposite phenomenons exceptionally to describe various aspects of both neural networks' complexity.
The book argues that grappling with these chaotic systems holds lessons for AI researchers seeking to build adaptable, resilient models. The brain achieves robustness despite – or perhaps because of – underlying chaos and randomness percolating through its networks.
While the author makes a strong case for collaboration between neuroscience and AI, the rapid progress of artificial intelligence over the past decade suggests the arrow of learning between the two fields has reversed in crucial ways. This reviewer feels that back when ANNs were in their infancy, AI researchers had much to gain from understanding the workings of organic BNNs. Insights into biological neural architecture and plasticity accelerated early ANN development. However, ANNs today operate unconstrained by the limitations of their organic counterparts - they do not have to be energy efficient or constructible from genetic code. They are not survival maximizers without a goal. The environments and design parameters for ANNs are now so distinct that neuroscience, for all its intricacies, likely has more to learn from AI than vice versa moving forward. While exceptions exist, the utility of modeling AI systems on detailed neurobiology has also diminished because of the incompleteness of our understanding of low-level brain function.
In summary, while conceptual inspiration clearly flowed from neuroscience to AI originally, ANNs have evolved so dramatically in recent years that they operate under very different principles and design constraints compared to BNNs. While fascinating, the complex mechanics of actual brain processes seem unlikely to offer meaningful shortcuts for today's leading AI techniques.
Such disagreements aside, here is a book where one learns in every para. The details are exhaustive but also fascinating when one begins to think how evolution has produced a gadget of such intricacy. The book not only succeeds at conveying the awe-inspiring complexity and magic of the BNNs but also throws light on how we will struggle to truly understand and master ANNs despite being their creators. ...more
The AI Revolution is an engaging and thought-provoking read, offering insightful examples that highlight the potential of artificial intelligence. TheThe AI Revolution is an engaging and thought-provoking read, offering insightful examples that highlight the potential of artificial intelligence. The authors eloquently capture the essence of their argument in the middle of the book:
"Medicine traditionally refers to a sacred relationship between a doctor and a patient — a twosome, a dyad. "And I'm proposing that now we move to a triad," he said, with an AI entity like GPT-4 as the third leg of that triangle.
Today's LLMs are likely to appear elementary in a few years. As impressive as their feats are, as shown in the book, they still have much to demonstrate in order to surpass the expertise of our finest medical professionals indefinitely. Even if they claim to outperform the average practitioner, it is natural for many of us to harbor reservations and doubt their abilities, regardless of the irrationality behind such sentiments. Nevertheless, the book masterfully showcases the incredible potential of integrating GPT as the third agent in the doctor-patient dynamic. From aiding in diagnosis, documentation, and explanations to serving as an error handler, facilitating patient-doctor communication, optimizing planning, and enhancing overall efficiency—the possibilities are vast. Furthermore, the book hints at the future prospects of LLMs as long-term record-keepers and even contributors to drug discovery, further emphasizing their potential value.
It is pretty likely that healthcare and pharmaceuticals emerge as generative AI's most significant application sectors over time....more
"What is ChatGPT doing" has the perfect combination of well-explained, highly technical details with clear, concise, human writing (!) to make some pr"What is ChatGPT doing" has the perfect combination of well-explained, highly technical details with clear, concise, human writing (!) to make some profound points about not just the product that is at everyone's tongue tips but also a scientific discovery that has the potential to change everything in humanity's future.
Superficially, the reader is taken on an enlightening journey through the intricate realm of generative artificial intelligence (AI) and ChatGPT's technical architecture with the emphasis on things that make it completely different from everything that carried the tag of "artificial intelligence" before. The least one learns is an appreciation for why this "AI" is different and why it needs learning. There will always be those who will try to confuse previous-generation neural network technologies with generative AI. Still, these people could be making even worse mistakes than those made by many celebrated handphone makers of 2007 who continued to call their devices also as "Smartphones" without ever stopping to understand what had changed post the arrival of the iPhone.
However, there is more it what has changed than just the arrival of a new product or a large language model. Philosophically, the LLMs are a new lens through which we are able to dissect the mechanisms of human cognition.
The enigmatic structure of the human brain, its nebulous configurations of neurons and synapses, has long defied capture in simple equations. Our brains have been black boxes with constellations of thoughts and perceptions, ideas and inspirations that form the stars of our individual universes. We have been able to do little more than brandish them a chaotic cosmos where the most we could attempt to understand were in small parts. Yet, as Wolfram illustrates, the advent of generative AI has begun to chart a course through this complexity, offering profound insights into how we perceive everything from language to visuals and more.
Despite the differences in the inner workings of our brains and the current constructs of generative AIs, there is a striking similarity in the results both produce. This resemblance is a testament to the role of near-by or near-term parameters ("attention") in shaping our subsequent thoughts and ideas, just as generative AI uses immediate context to generate responses. It's like looking at a painting, where each brushstroke contributes to the overall picture while also extending the meaning of what it surrounds.
However, the parallels between the human brain and AI are not without much-discussed terrifying implications. They are not the topic of the book, but the explanations only cement the thought that for ever-improving machines, passing our level of intelligence will be just a small milestone on its light-speed march.
Wolfram's book excels in describing the concept of computational irreducibility and the constraints of generative AI. There are limitations to what an AI model can infer or predict about a system's future states without simulating the system step by step. In other words, there are realms within which the process of querying will have to take alternate forms, including the good old lookup-type forms. This harks back to traditional methods like reaching for a dictionary when one comes across an unfamiliar word, a process not yet completely replicated by LLMs that rely solely on training-based generative methods. This is sure to change soon.
The heart of the book lies in its elucidation of how generative AI creates its own classifications through tokens and navigates the complexity of information through attention mechanisms. This process mirrors human cognition in many ways, particularly because it makes LLMs their own kind of black boxes that work but without us understanding - in human terms, at human levels - how.
Wolfram's ability to explain equations makes the book far better than hundreds that spring up on a daily basis. A must-read guide for almost everyone....more
The Alignment Problem is the perfect book to read to understand the machinery behind generative artificial intelligence (AI) systems that are taking oThe Alignment Problem is the perfect book to read to understand the machinery behind generative artificial intelligence (AI) systems that are taking over the world. This insightful and thought-provoking book comprehensively explains the underlying mechanisms of AIs. Directly and indirectly, it also explores one of the most pressing issues of our time: how to ensure that artificial intelligence is used for good rather than for harm.
If the previous era programming were based on conditionals and loops, variables and functions, the machine learning algorithms have their equivalents in reinforcement learning, shaping, and IRL. This book is a fantastic introduction to the lexicon of this emerging field.
The author delves into the challenge of aligning AI with human values, arguing that AI systems have become exceedingly powerful and opaque. While the call for greater accountability and alignment is compelling, the author often appears to state an aspirational goal—akin to the desire for no child to go hungry—rather than a concrete, actionable solution provider.
The reviewer wishes to share some personal takeaways from the book, focusing on the two developmental phases of machine learning systems: The initial phase involves constructing an AI system to accomplish broad goals through learning, imitation, and problem decomposition. This phase is reminiscent of the AlphaGo machine that mastered the game of Go by studying human gameplay history. In the subsequent phase, the AI system evolves to set its own objectives, generate its data, and create learning processes with diminishing inputs from its human masters. This is the scary stage where machines are on autopilot learning, including devising better learning methods to do things we barely comprehend outside the end results. An example of this is AlphaGo Zero. A future example of this could be the use cases of AutoGPT.
Most of us who love to quote Asimov's Robotics rules or Turing tests discuss AIs in the first phase above. Discussions on topics like "Alignment" are also more relevant when humans have some control over end goals, interim steps, process creations, and inputs. We are incapable of imagining what machines could do once they internalize more and more away from our provided structures. Positively, machines could create scientific or medicinal discoveries and solve some of our biggest problems. The list of negatives is far more inscrutable and hence frightening for most of us.
Although the desire for control is understandable, implementing safeguards may be challenging due to competition among nations and corporations. For instance, one country's attempts to restrict AI applications in a certain sector could be undermined by providers in another jurisdiction. Additionally, machine learning algorithms are not difficult to develop and are accessible to many. As a result, small teams can rapidly create innovative products that are quickly imitated, rendering traditional regulatory approaches ineffective.
Another thing one learns from the book is how most AI/ML tools - aka libraries, stacks, programming methods - are available to everyone, everywhere. Without a doubt, a few will emerge as commercial winners with their AI implementations, but this is going to be a crowded field in every aspect of AI use cases, making control extremely difficult.
In conclusion, The Alignment Problem is a timely and invaluable resource for readers looking to understand the complexities of AI alignment and its implications for society. This book is an important contribution to the ongoing conversation about AI's potential to shape our future, both positively and negatively, and is a must-read for anyone interested in the subject....more
Superficially, "A Brief History of Artificial Intelligence," is an ambitious attempt to provide a comprehensive account of the development of artificiSuperficially, "A Brief History of Artificial Intelligence," is an ambitious attempt to provide a comprehensive account of the development of artificial intelligence (AI). However, the different and evolving meaning of the central term, artificial intelligence, comes in the way. The book simply fails to address the rapidly evolving state of AI as it is known today. The way the term AI was meant at various times, the book turns into a computer science history.
The book was published in 2021. For the reader, however, a complete absence of generative AI, ChatGPT, Bard, or other similar terms makes the history itself too historic!
Its treatment of AI's potential uses and abuses is similarly weak. Discussions are cursory, often recycling arguments and examples covered well in popular journals. This lack of depth is especially apparent when the author addresses weighty topics like AI consciousness or its impact on jobs.
The author's biggest failure is in its predictions about the future of AI. Wooldridge's vision of what AI will be capable of in the coming decades seems painfully outdated compared to the developments we've seen in just the past few weeks. The rapid pace of AI advancement has outstripped the author's expectations, rendering many of his predictions obsolete so soon after its release.
Given that the topic is soon to have many books with the central focus on the path-breaking latest innovations, this book has lost its relevance even more....more