如何创造可信的AI ( etc.) (Z-Library)

# How to Create Trustworthy AI — Reading Guide ## 【One-Line Pitch】 A sharp, contrarian critique of deep learning's limits and a roadmap for building AI that can actually be trusted — essential reading for AI practitioners, product leaders, and anyone making decisions about where AI should go next. ## 【Book Arc】 - **Opening (~0%–25%)**: Establishes the authors' credibility and the book's core mission — separating AI hype from reality. Gary Marcus's background in cognitive science, his public debates with deep learning advocates, and endorsements from figures like Noam Chomsky and Steven Pinker frame the book as a corrective to AI over-optimism. - **Early (~25%–50%)**: Introduces the central concept of the "AI gap" — the disconnect between what AI is reported to do and what it can actually do. The authors begin dissecting deep learning's fundamental limitations, setting up the detailed critique that follows. - **Middle (~50%–75%)**: The book's first major section (Chapters 1–5) systematically catalogs nine categories of deep learning's weaknesses, including lack of robustness. Each limitation is illustrated with concrete examples from natural language processing, computer vision, and robotics, showing why current approaches fail in real-world conditions. - **Late (~75%–100%)**: The second major section (Chapters 6–8) pivots from critique to construction. The authors propose a path forward based on insights from cognitive science — 11 key principles for building AI with richer internal structure, common sense reasoning, and causal understanding. - **Ending (~100%)**: Concludes with the practical capabilities needed for trustworthy AI: reliable engineering practices, safety protocols for mission-critical applications, and ethical frameworks for autonomous systems. ## 【Key Takeaways】 - **The "AI gap" is real and costly** (Early): Media reports and academic claims about AI capabilities far outstrip what systems can actually do in the real world. The authors identify three "traps" — the gullibility trap (anthropomorphizing machines), the illusory progress trap (assuming narrow success generalizes), and the robustness trap (systems failing on slight variations). Understanding these traps is essential before investing in AI applications. - **Deep learning has fundamental, not incidental, limitations** (Middle): The book catalogs nine specific weaknesses, from lack of robustness to inability to handle novel situations. These aren't engineering problems to be tuned away — they're structural issues rooted in how deep learning models learn from data without understanding underlying principles. - **Common sense is the missing ingredient** (Late): Human intelligence relies on vast, implicit knowledge about how the world works — time, space, causality, physical objects. Current AI systems lack this entirely, which is why they fail at tasks humans find trivial. Building common sense into AI is presented as the central challenge for the field. - **Internal structure matters more than raw data** (Late): Drawing on Marcus's cognitive science background, the book argues that AI needs richer internal representations — not just statistical patterns but structured knowledge that can support reasoning, causal inference, and flexible problem-solving. - **The path forward requires cognitive science, not just computer science** (Late): The authors extract 11 principles from how human minds develop and function, arguing these should guide next-generation AI architecture. This cross-disciplinary approach is the book's most distinctive contribution. - **Trustworthy AI needs engineering discipline** (Ending): Just as software engineering matured over decades, AI needs similar rigor — reliable development tools, testing methodologies, and safety protocols. Mission-critical applications like autonomous vehicles demand especially high standards. - **Ethics must be built in, not bolted on** (Ending): AI systems, especially robots, need explicit ethical frameworks from their creators. The book references Asimov's laws as a starting point, though the authors acknowledge these are insufficient without deeper understanding baked into the systems themselves. ## 【Reading Tips】 - **Skim the endorsements and foreword** (0%–25%): The praise from figures like Chomsky and Pearl is useful context, but the foreword by Lu Qi provides the clearest structural overview — read it to understand the book's two-part architecture before diving in. - **Deep-read the critique section** (Middle): Chapters 1–5 contain the book's most valuable content — the nine limitations of deep learning. Take time with the concrete examples; they make abstract criticisms tangible and are the book's strongest evidence. - **Focus on the 11 principles** (Late): The second half's proposal for building trustworthy AI is dense but rewarding. Don't rush through the cognitive science foundations — they're essential for understanding why the authors propose specific architectural changes. - **Watch for the debate context**: Marcus is openly critical of deep learning pioneers like Yann LeCun. Understanding this adversarial context helps you evaluate his arguments fairly rather than accepting them at face value. - **Consider reading critically**: The book is persuasive but one-sided. Pair it with responses from deep learning advocates to get a balanced view of the debate. ## 【Coverage Limits】 This guide is based on excerpts that include the book's front matter, the foreword by Lu Qi, and structural summaries of both major sections. The detailed technical arguments within Chapters 1–8 are summarized at a high level; specific examples and case studies from the full text are not covered in depth here. ##