第1章
The AI Illusion: Why Today's Machines Are Smarter Than They Seem
When Elon Musk proclaimed artificial intelligence "the biggest existential threat" to humanity, he wasn't alone. Stephen Hawking warned AI could "spell the end of the human race," while tech titans like Bill Gates expressed "serious concerns." Yet ironically, as these warnings circulated, Facebook's much-hyped M chatbot was quietly discontinued after failing to live up to expectations. Tesla's "self-driving" cars kept crashing into parked emergency vehicles. IBM's Watson, once celebrated for beating Jeopardy champions, saw its healthcare initiatives shelved after making "unsafe and incorrect" medical recommendations.
"Rebooting AI" became an instant classic in Silicon Valley upon its 2019 release, with tech leaders from Melinda Gates to Garry Kasparov praising its clear-eyed assessment of AI's true capabilities. As AI enters mainstream consciousness through tools like ChatGPT, this book's message grows increasingly relevant: today's AI isn't nearly as intelligent as headlines suggest, and building truly reliable systems requires a fundamental rethinking of our approach.
第2章
The Great AI Disconnect
Artificial intelligence has consistently overpromised and underdelivered since its inception. The field's pioneers were notoriously optimistic - Marvin Minsky predicted in 1967 that "within a generation, the problem of creating 'artificial intelligence' will be substantially solved." Herbert Simon declared in 1957 that machines would be chess champions, discover mathematical theorems, and compose "aesthetically satisfying" music within a decade. While some predictions eventually came true, they took far longer than anticipated.
This pattern of overconfidence continues today. Waymo promised driverless cars without safety drivers by 2018 - a deadline that silently passed. IBM Watson's healthcare initiatives were quietly shelved after doctors found its recommendations "unsafe and incorrect." Facebook's M chatbot was canceled just three years after launch, having failed to live up to its promise of handling complex tasks without human intervention.
Yet despite these failures, the rhetoric around AI remains almost messianic. Google's Eric Schmidt claimed AI would solve climate change, poverty, war, and cancer. Simultaneously, others like Nick Bostrom, Henry Kissinger, and Stephen Hawking warn of existential threats from superintelligent machines. This disconnect from reality becomes stark when you consider that current robots struggle with basic tasks like turning doorknobs while Teslas keep rear-ending parked emergency vehicles.
Media coverage exacerbates these misconceptions. Headlines claiming "Robots Can Now Read Better than Humans" grossly exaggerated minor progress on a narrow reading test that only required finding answers explicitly stated in text. Similarly, Facebook's simple question-answering system for a four-line Lord of the Rings synopsis was portrayed as breakthrough technology.
The reality is that despite impressive advances in game playing, speech recognition, and facial identification, today's AI systems remain fundamentally limited. They excel at narrow tasks in controlled environments but fail spectacularly when confronted with novelty or complexity. We wanted Rosie the Robot but got Roomba - a device that can vacuum floors but does little else.
第3章
The AI Chasm: Why Current Systems Fall Short
The gap between AI ambition and reality stems from three fundamental challenges. First is the "gullibility gap" - humans easily attribute intelligence to machines through what psychologists call the "fundamental overattribution error." We saw this with Eliza, a 1960s chatbot that merely matched keywords but convinced users it understood them through simple tricks that appeared as sympathetic feedback. Users would chat for hours, forgetting they were talking to a machine despite its paper-thin understanding.
This overattribution error can be deadly. In 2016, a Tesla owner trusted Autopilot enough to allegedly watch Harry Potter while driving, before fatally crashing into a white tractor-trailer the system failed to detect. The driver believed the system was more capable than it actually was - a mistake encouraged by Tesla's marketing.
The second challenge is the "illusory progress gap" - mistaking progress on easy problems for progress on hard ones. IBM's Watson winning Jeopardy! impressed viewers, but the game's format of question-answering based on factual information played to the system's strengths. Games like chess and Go are perfect-information systems with complete data, unlike real-world scenarios with uncertainty and incomplete information.
The third challenge is the "robustness gap" - assuming solutions that work sometimes will eventually work all the time with more data. This affects driverless cars that perform well in ideal conditions but struggle with unexpected situations. Current AI focuses on problems with high error tolerance like ad recommendations, not critical applications requiring near-perfect reliability.
What's missing is broad (or "general") intelligence - the ability to adapt flexibly to an open-ended world. While narrow AI excels in closed systems like Go with fixed rules, real life has no fixed rules and unlimited possibilities. We can't practice every situation in advance or foresee what information we'll need.
第4章
Deep Learning: Breakthrough and Limitations
The current AI enthusiasm stems largely from deep learning - a technique that has transformed the field by enabling machines to learn directly from data rather than following explicitly programmed rules. This approach dates back to the 1950s when Frank Rosenblatt built early neural networks aimed at recognizing objects without requiring programmers to anticipate every contingency.
These early systems were underpowered by today's standards, with limited network depth and resolution of only 400 pixels (1/30,000th of an iPhone X). Neural networks struggled until the big data revolution of the early 2010s, when researchers like Hinton, Bengio, LeCun, and Schmidhuber finally had their day. The breakthrough came not from mathematical advances but from repurposing GPUs (graphics processing units) originally developed for video games.
In 2012, Geoffrey Hinton's team achieved 84% accuracy on ImageNet recognition, dramatically better than previous techniques, and by 2017 accuracy reached 98%. This success triggered an explosion of investment and applications. Deep learning now powers speech recognition, translation, image labeling, and recommendation systems used by billions of people daily.
Deep learning is based on two fundamental ideas. First is hierarchical pattern recognition, inspired by Hubel and Wiesel's Nobel Prize-winning neuroscience research showing that visual neurons respond to stimuli of varying complexity in a hierarchical fashion. The second key idea is learning through weight adjustment - neural networks consist of nodes connected by weights that are gradually optimized through exposure to examples.
Despite its impressive achievements, deep learning faces fundamental limitations. It's greedy, requiring enormous data compared to humans who learn from few examples. It's also opaque - neural networks consist of vast arrays of numbers defying intuitive understanding, making it impossible to explain specific decisions or predict when systems will fail catastrophically.
Most troublingly, deep learning is remarkably brittle. Systems can hallucinate nonexistent objects or miss obvious ones. Researchers have demonstrated dozens of ways to fool these networks: 3D-printed turtles misidentified as rifles, baseballs with foam seen as espresso cups, and stop signs with subtle alterations interpreted as speed limit signs. Adding tiny patches of random noise can trick networks into seeing "tiger cats" instead of piggy banks.
This creates a "long tail" problem: deep learning handles common cases but fails with unusual configurations or relationships between objects. Despite commercial success in narrow applications, deep learning remains an idiot savant with miraculous perceptual abilities but little comprehension.
第5章
The Reading Comprehension Challenge
Despite incremental improvements in AI systems like Google Search and virtual assistants, we're nowhere near achieving genuine machine reading comprehension. Virtual assistants excel at specific tasks but fail at basic reasoning. They might correctly identify that Donald Trump is a person or an Audi is a vehicle, but struggle with questions like "Can an Audi use gas?" or misinterpret requests like "Find a fast food restaurant that is not McDonald's" by returning only McDonald's locations.
Even WolframAlpha, despite its vast knowledge databases, makes fundamental errors like calculating the distance from San Diego to Mexico's geographic center rather than to its border, or failing to answer whether Ella Fitzgerald was alive in 1960.
Watson's famous Jeopardy! victory is less impressive than it appears - almost 95% of Jeopardy! answers are Wikipedia page titles, making the game largely an information retrieval challenge rather than one requiring genuine understanding. No current AI system can integrate information across multiple sentences or sources the way humans naturally do when reading.
Deep learning approaches to language have improved machine translation by learning word correspondences from parallel texts, but they still fail when deeper understanding is required. Google Translate might render "Je mange un avocat pour le dejeuner" as "I eat a lawyer for lunch" rather than the correct "I eat an avocado for lunch," choosing the statistically more common meaning of "avocat" while sacrificing sense.
The inability to understand unstructured text remains a significant bottleneck in commercial AI applications. We cannot yet automate reading of legal contracts, scientific articles, or medical records because AI lacks compositionality - the ability to construct meaning from parts. Deep learning doesn't recognize that sentences contain phrases referring to objects and their relationships, nor does it understand how concepts relate to each other in structured ways.
第6章
The Robotics Reality Check
Unlike science fiction robots, real-world robots are far more limited in their capabilities and applications. While popular media portrays robots as sophisticated beings like R2-D2, C-3PO, or the menacing Terminator, the reality is far more modest. The world's bestselling robot isn't a humanoid companion but the humble Roomba vacuum cleaner, which has sold over 30 million units worldwide. Despite significant investments by companies working on specialized robots for tasks like construction site excavation, warehouse automation, or agricultural applications like apple-picking and crop monitoring, we're still decades away from having anything resembling Rosie from The Jetsons handle our diverse domestic chores.
Roomba's success stems from its focused functionality - it requires minimal intelligence for its narrow task because vacuuming follows relatively simple rules and patterns. The real challenge lies in creating robots that can perform the broad range of physical tasks humans handle effortlessly throughout their day, from opening jars with varying lid types to wrapping oddly shaped presents or folding laundry with different fabric types. While progress exists in specific areas - like Manuela Veloso's CoBots that navigate Carnegie Mellon's university halls delivering packages, or autonomous drones that can track runners and capture aerial footage - most impressive robotics demonstrations remain confined to controlled laboratory environments rather than real-world consumer applications.
Companies like Boston Dynamics have created remarkable machines that capture public imagination - their humanoid Atlas robot can perform backflips and parkour movements, while their dog-like SpotMini can open doors and navigate complex terrain. However, these are carefully orchestrated demonstrations in controlled environments rather than practical household solutions. The primary limitation isn't hardware capabilities but software intelligence. Any truly intelligent robot needs to compute five fundamental things simultaneously: precise self-localization, environmental perception, immediate action selection, motion planning implementation, and long-term strategy optimization.
Even industry leaders face significant challenges. Tesla's manufacturing struggles, which Elon Musk attributed to "excessive automation," highlight how robots struggle to adapt to dynamically changing environments and unexpected variations in tasks. Real-world robots must handle countless unpredictable scenarios: from encountering unusual obstacles like cockroaches in wine glasses to making complex value judgments about whether to risk creating a commotion to pick up a fallen cracker versus an expensive diamond earring.
While robotics has made excellent progress in areas like self-localization (knowing where they are in space) and reasonable advances in motor control (executing physical movements), the field continues to struggle with three essential capabilities: accurately assessing complex situations, predicting probable future scenarios, and making dynamic decisions as circumstances rapidly change. The challenge lies in the virtually infinite space of possibilities in real-world environments - from varying lighting conditions to different surface textures and object arrangements - making it impossible to rely solely on data-driven approaches or pre-programmed responses. Recent advances in artificial intelligence and machine learning have helped, but we're still far from achieving the adaptability and general intelligence that characterize human physical interaction with the world.
第7章
Lessons from Human Cognition
Marvin Minsky once observed that intelligence stems from diversity rather than any single perfect principle. This insight contradicts the current AI trend of seeking a "master algorithm" that can solve all problems. The brain's immense complexity - with its 86 billion neurons, trillions of synapses, hundreds of proteins per synapse, and 150+ distinct brain areas - suggests that intelligence cannot be reduced to a single principle.
From cognitive science, we can identify several key principles that AI systems should incorporate:
1. Internal representations matter. While neural networks technically have representations (vectors), they lack richer structures like propositions that describe relations between entities. Classical AI systems explicitly represent facts like "PART-OF (BERLIN, GERMANY)" and "VISITED (KENNEDY, BERLIN, JUNE 1963)," enabling reliable inference.
2. Abstraction is essential. Humans naturally work with abstractions, understanding concepts like "sisterhood" that apply across many specific instances. We use abstract reasoning to make inferences - if we learn that two people share parents, we can deduce they're siblings and likely know each other.
3. Cognitive systems need structure. Neuroscience reveals the brain as highly structured, with different areas serving distinct functions. This contrasts with current AI trends toward end-to-end models using single homogeneous mechanisms with minimal internal structure.
4. Multiple tools are required even for simple tasks. Even basic cognitive functions like verb past tense formation use both rule-based systems (walk-walked) and associative networks (sing-sang). True AI will need both deep learning mechanisms for pattern recognition and classical AI machinery for reasoning and abstraction.
5. Language and thought are compositional. We construct complex sentences from smaller components, with meaning depending on precise structure. Deep learning systems often miss subtleties, like distinguishing between "Loved it until I hated it" and "Hated it until I loved it."
6. Understanding requires both bottom-up and top-down processing. Our perception integrates both what our senses detect and our prior understanding of the world. When shown ambiguous images, people's interpretations vary dramatically based on the context they're given.
7. Concepts exist within theories. We understand concepts not just by their physical properties but through the theories in which they're embedded. A quarter-sized pizza seems perfectly acceptable, but a pizza-sized quarter would never be accepted as currency.
8. Causality is fundamental. While physics could theoretically explain all causal relationships through detailed simulation, the real world contains too many variables to track. Instead, we rely on approximations - knowing aspirin relieves pain without understanding its biochemistry.
9. We track individual entities. Humans naturally track individual people and things through time - your spouse's career and preferences, your daughter's fears and loves, your car's maintenance history. Deep learning systems excel at category-level generalizations but struggle with tracking specific individuals.
10. Complex cognition isn't a blank slate. While machine learning researchers often pursue approaches that learn everything from scratch, biology shows that nature and nurture work together. Humans are likely born understanding that the world consists of enduring objects, with innate senses of geometry, quantity, and intuitive psychology.
第8章
Building Common Sense into Machines
Common sense - the basic knowledge we expect ordinary people to possess - has proven remarkably difficult to build into machines. John McCarthy first identified this problem in 1959, but progress has been remarkably slow. Neither classical AI nor deep learning has made significant headway.
Multiple approaches have been attempted with limited success. NELL tries to learn everyday knowledge by crawling the web, but produces spotty results. Crowdsourcing projects like ConceptNet collect commonsense facts, but often lack precision computers require. The largest such effort, CYC, contains millions of carefully encoded facts but remains incomplete despite three decades of work.
To make progress, we need both an inventory of knowledge a general intelligence should have and clear ways to represent that knowledge in machines. Taxonomies (like "dogs are mammals") are relatively straightforward to encode, but common sense extends far beyond taxonomies. Semantic networks attempt to represent broader relationships but suffer from ambiguity - machines struggle with what isn't made explicit.
Beyond encoding individual facts, AI systems need coherent knowledge frameworks that connect related concepts. Rather than learning that writers write books, painters paint pictures, and composers compose music as separate facts, systems should recognize these as instances of a more general "individual creates work" relation.
If limited to just three knowledge frameworks, we would focus on what Kant identified as fundamental: time, space, and causality. Classical AI researchers have made progress on temporal logic systems that can represent how situations develop over time, allowing machines to plan sequences of actions. However, significant challenges remain in mapping natural language to these internal representations.
Understanding space is equally challenging. While we understand Euclidean space and can perform geometric calculations for realistic computer graphics, machines still can't grasp how an object's shape relates to its function. Understanding that cucumbers stay in a mesh bag while peas fall through requires reasoning about the relationship between object sizes and mesh spacing.
Causality encompasses all knowledge about how the world changes over time, from Newton's gravity to the specific fact that pressing a TV remote's power button turns the television on. The hallmark of human causal thinking is versatility - understanding the remote-TV relationship lets us make predictions, plans, and explanations.
To achieve deep understanding, AI needs more than just improved knowledge representation and reasoning - it requires a fundamental rethinking of learning itself. The solution isn't purely hand-coding knowledge nor pure machine learning, but a carefully structured hybrid. We need systems with strong innate foundations that can learn at conceptual and causal levels - learning theories, not just isolated facts.
第9章
The Path to Trustworthy AI
Trustworthy AI requires solid engineering practices that are currently lacking in the field. Unlike other safety-critical industries, AI development often prioritizes short-term functionality over robustness, neglecting crucial engineering principles like building in safety margins, designing for failure with redundant systems, incorporating fail-safes, and ensuring maintainability.
Program verification offers promising approaches for ensuring AI safety. This technique, which formally verifies program correctness using mathematical logic, has already proven valuable in critical systems - Microsoft's Static Driver Verifier significantly reduced system crashes by ensuring device drivers follow strict rules, while verification techniques have been used to guarantee the Airbus control program won't crash and to verify aircraft collision avoidance systems.
Deep learning poses unique challenges compared to traditional software engineering. While conventional programming allows developers to understand program logic and systematically debug errors, deep learning systems operate more like pharmacology - we understand broadly why they work but lack detailed insight into their internal operations. With hundreds of millions of parameters, these systems provide no clear explanations for their successes or failures, making debugging largely trial-and-error.
For AI to be truly trustworthy, it must also incorporate ethical values. While Asimov's Three Laws of Robotics provide basic guidance for straightforward ethical decisions, real-world applications raise complex questions about what constitutes harm, what responsibilities AI systems have to prevent indirect harm, and how to balance competing priorities.
Nick Bostrom's famous paper-clip maximizer thought experiment - where a superintelligent AI converts all matter in the universe into paper clips - represents a fundamental flaw in thinking about AI risk. The scenario assumes a machine smart enough to master interstellar travel yet lacking the common sense to recognize its quest is both pointless and morally wrong. The real danger isn't superintelligent machines seeking to obliterate humanity, but our reliance on systems without common sense - "idiots savants with power" like autonomous weapons without moral constraints or AI-driven newsfeeds that prioritize engagement over societal impact.
第10章
The Future of Intelligence
Trustworthy AI grounded in reasoning, commonsense values, and sound engineering will eventually transform our world, whether in decades or centuries. While we've seen significant advances in machine learning applied to big data - improving speech recognition, translation, and image labeling - these will ultimately prove mere appetizers compared to what follows.
The true turning point won't be the 2012 rebirth of deep learning but the eventual solution to common sense and reasoning challenges that yield deeper understanding. This breakthrough will enable machines to learn like children - easily, powerfully, requiring minimal examples to create valid models of new concepts. Such AI will truly comprehend novels, films, and news stories, while robots with deep understanding will safely navigate the world, manipulate objects, and interact naturally with people.
Once computers genuinely understand the world, search will vastly improve, answering complex questions like "Who was the oldest Supreme Court justice in 1980?" Digital assistants will become democratized, handling complex tasks like planning corporate retreats. Computers will become dramatically easier to use, with natural language interfaces eliminating tedious tasks. Virtual reality will advance to Star Trek holodeck levels, and domestic robots will finally become practical for cooking, cleaning, and household maintenance.
Eventually, AI will extend beyond common sense to expert-level understanding in medicine, law, teaching, and programming. This will empower individuals to accomplish far more, while automating dangerous work through rescue robots and robotic firefighters. Artists and hobbyists will gain AI-powered assistants that dramatically enhance creative possibilities. Scientific discovery will accelerate as machines combine computational power with human-like intuition, potentially solving challenges in neuroscience, mental health, agriculture, and clean energy.
While this future brings many benefits, employment will likely become scarcer, intensifying debates about basic income and income redistribution. Society's fabric will change fundamentally, with decreased drudgery but fewer jobs. The cognitive revolution in AI will spark societal changes comparable to the Industrial Revolution - mostly positive if approached wisely, but requiring careful navigation of the challenges ahead.
