[Book Notes] David Deutsch: Good Explanations and the Beginning of Infinite Progress

20 minute read

Published:

This post supports English / 中文 switching via the site language toggle in the top navigation.

Introduction: Why Did Progress Begin?

David Deutsch’s The Beginning of Infinity: Explanations That Transform the World begins from a historical asymmetry. Human beings have existed for a very long time, yet rapid, sustained progress appeared only recently, around the scientific revolution and the Enlightenment. Why did societies that had changed slowly for millennia begin to transform science, technology, politics, morality, and everyday life within a few centuries?

Deutsch’s answer is a philosophy of knowledge. Progress begins when a culture develops a tradition of seeking good explanations, exposing them to criticism, and replacing them when better explanations are created. Observations and experiments matter enormously, though they cannot manufacture theories for us. Knowledge grows through creative conjecture and error correction.

The book is unusually ambitious. Its eighteen chapters move through epistemology, quantum physics, computation, mathematics, biological evolution, artificial intelligence, aesthetics, culture, political institutions, and environmental thought. These subjects are connected by one claim: explanatory knowledge is a real physical force, and its possible reach has no known upper bound beyond the laws of nature.

The “infinity” in the title therefore does not mean that humanity will automatically survive forever or eventually know everything. It means that there need be no final problem, final explanation, or final stage of improvement. We are at a beginning whenever the conditions for open-ended knowledge creation come into existence.


1. Good Explanations Are Hard to Vary

The central unit of Deutsch’s philosophy is the good explanation. A good explanation accounts for what happens in a way whose details are constrained by the problem. Alter an important part casually and the explanation stops working. Deutsch calls this quality hard to vary.

Ancient stories about the seasons illustrate the opposite. If winter is caused by a god’s sadness, almost any change in the story can be made to fit the same observations: a different god, a different emotion, a different sequence of events. The flexibility looks like explanatory power, but it is precisely the weakness. Nothing in the phenomenon selects those details.

The explanation based on Earth’s axial tilt is different. The tilt, orbit, angle of sunlight, and opposite seasons in the two hemispheres constrain one another. Changing one component creates consequences elsewhere and often contradicts observation. Its parts are connected by an account of reality rather than assembled to protect a conclusion.

This criterion reaches beyond laboratory science. An organizational explanation such as “the team lacks ownership” is easy to repeat and easy to vary. A stronger account identifies which decision rights, incentives, information flows, and feedback delays produced a particular failure. It may still be wrong, but it gives criticism somewhere to land.

Good explanations are not synonymous with complicated explanations, mathematical models, accurate predictions, or expert consensus. A formula can predict a pattern without explaining its mechanism. A simple idea can have immense reach. The decisive test is whether the explanation solves the problem without arbitrary adjustment and survives serious attempts to improve it.


2. Knowledge Grows Through Conjecture and Criticism

Deutsch develops Karl Popper’s fallibilist epistemology. We do not derive general theories from repeated observations. Experience supplies problems, constraints, and tests; people supply conjectures. A theory of stars cannot be read directly from points of light, just as a theory of flight could not be extracted from centuries of watching objects fall.

The knowledge-creation cycle has no guaranteed first step and no final certificate:

  1. We encounter a problem within existing knowledge.
  2. We propose explanations that might solve it.
  3. We criticize those proposals through argument, experiment, and comparison.
  4. We eliminate detected errors and inherit new problems.

Evidence is indispensable because it helps choose among explanations. It does not turn a surviving theory into certain or finally justified truth. Every explanation remains open to criticism, including our deepest physical theories and the philosophical rules we use to judge them.

This is why Deutsch rejects both authority and radical skepticism. Authority fails because no person, tradition, dataset, or method is infallible. Skepticism fails when it treats the possibility of error as a reason to deny improvement. We can be wrong and still know that one explanation solves problems that another leaves untouched.

Fallibility is therefore productive. A scientific community becomes rational through organized correction, not through the purity of individual scientists. Peer review, replication, open data, adversarial discussion, and peaceful institutional change are valuable when they make errors easier to discover and remove. Their legitimacy comes from correctability rather than prestige.


3. The Reach of Explanations

Human access to evidence is local and narrow. Our best explanations reach far beyond it. From light entering a telescope, we infer the composition and history of stars no human will visit. From traces in rocks and genes, we reconstruct events that occurred before any observer existed. Theory connects a small interaction here to structures separated from us by enormous distances and times.

Deutsch uses this reach to challenge empiricism’s image of the mind as a passive receiver. Instruments do not simply enlarge the senses. Their readings become meaningful inside explanatory theories. A radio telescope, particle detector, or microscope lets us perceive reality because layers of theory tell us what its signals represent and how errors can be corrected.

Explanations also travel between domains. Computation links physical machines to abstract algorithms. Evolution explains biological adaptations and helps illuminate how non-genetic ideas persist. Epistemology connects scientific discovery to political institutions because both depend on finding and correcting error.

This does not imply that one theory reduces every level of reality to particle physics. Deutsch is a realist about abstractions. Numbers, programs, institutions, and knowledge have physical embodiments, while explanations at those higher levels can be indispensable. Knowing every atomic position in a computer would not by itself explain which algorithm it is running or why the program solves a problem.

The reach of an explanation is a reason for wonder and also a demand for discipline. A wide-ranging theory earns its reach by solving linked problems under criticism. Merely attaching the same fashionable vocabulary to many subjects produces breadth without explanatory depth.


4. The Jump to Universality

One of the book’s deepest recurring patterns is the jump to universality. Many systems improve gradually until a particular architecture crosses a threshold and becomes capable, in principle, of an entire class of transformations.

A universal computer is the clearest example. Given enough time, memory, and the right program, it can perform any computation that any other physical computer can perform. Its universality does not make every computation easy, fast, or currently known. It removes the need to construct a fundamentally new machine for each new computable task.

Deutsch argues that humans are universal explainers. A person is not confined to the narrow set of knowledge encoded genetically for survival in one ancestral environment. Human beings can create explanations about black holes, prime numbers, extinct organisms, constitutions, and possible technologies that have never existed. We remain finite, error-prone, and dependent on culture, yet the scope of what we may understand is not tied to a fixed cognitive niche.

Language and writing amplify this universality. They allow explanatory knowledge to survive individual minds, encounter criticism across generations, and recombine with distant ideas. A dynamic culture becomes a distributed error-correcting system whose future repertoire cannot be listed in advance.

Universality should not be confused with omnipotence. Physical law still distinguishes possible transformations from impossible ones; computational complexity, energy, time, and missing knowledge remain genuine constraints. The claim is subtler: there may be no fixed catalogue of subjects that human-style explanation is inherently unable to reach.


5. Knowledge Changes What Counts as a Resource

Resources are often discussed as if usefulness were an intrinsic property stored inside matter. Deutsch emphasizes that a resource is matter plus the knowledge of how to transform it. Uranium was not an energy resource for a society without nuclear physics. Sand became central to computation only through theories and techniques that made semiconductors possible.

This view changes the relationship between people and their environment. Human survival has never depended on preserving one naturally given niche unchanged. Clothing, agriculture, sanitation, medicine, buildings, and energy systems let people inhabit environments for which our bodies were not biologically adapted. Knowledge creates new options and also new responsibilities.

Limits remain real. The laws of physics cannot be negotiated, and local materials, energy, time, ecosystems, and institutional capacity can all become binding constraints. Deutsch’s point is that we rarely know in advance which apparent limit is fundamental and which is a symptom of missing knowledge. A shortage can sometimes be solved through substitution, efficiency, recycling, discovery, or a transformation no one has yet proposed.

This is also why people occupy an unusual place in his worldview. Knowledge can produce physical effects vastly disproportionate to the mass carrying it. A small encoded idea may cross the world, reorganize factories, redirect energy, cure a disease, or destroy a city. Humans matter cosmically because they can create explanations that transform which physical events become possible.

That claim carries an ethical edge. Greater reach magnifies both correction and error. The same universality that expands solutions can expand harm, so institutions capable of criticism must grow alongside technological power.


6. Optimism Is a Theory of Problems

Deutsch’s optimism is often mistaken for a forecast that everything will work out. His actual position is methodological. Problems are unavoidable because every solution changes the situation and reveals further problems. Progress depends on treating those problems as soluble through new knowledge unless a good explanation shows otherwise.

He compresses the attitude into two short propositions:

“Problems are inevitable. Problems are soluble.”

The first sentence blocks utopianism. No policy, technology, leader, or social arrangement ends history. A solution can create side effects, distribute costs unfairly, or expose a deeper difficulty. The second sentence blocks fatalism. Present failure does not establish impossibility, and an absence of known solutions is not an explanation that none can exist.

Optimism therefore requires energetic error correction. It asks what knowledge is missing, which assumptions conceal alternatives, how experiments can discriminate between proposals, and how institutions can recover when a proposal fails. Pessimism becomes intellectually dangerous when it quietly converts “we do not know how” into “it cannot be done.”

This philosophy also rejects prediction as the foundation of long-range planning. Future knowledge cannot be known in advance; if we already possessed its contents, it would be present knowledge. We can prepare by building wealth, scientific capacity, redundancy, open communication, and institutions that permit rapid criticism. Resilience comes from the capacity to create responses that no plan could specify beforehand.


7. Static and Dynamic Societies

For most of human history, Deutsch argues, societies were static: they changed slowly because cultural knowledge was transmitted in forms designed to suppress variation. Ritual, taboo, authority, and punishment kept behavior sufficiently faithful for the society to reproduce itself. Such cultures could contain sophisticated knowledge while making deliberate improvement exceptionally difficult.

A dynamic society sustains traditions of criticism. Its customs and institutions can be discussed, challenged, and deliberately changed without destroying the entire social order. The Enlightenment mattered because it strengthened this meta-tradition: no source of ideas is above criticism, and improvement can continue without a final authority.

Democracy fits this account through its error-correcting function. The central achievement is not a reliable method for selecting uniquely wise rulers. It is the ability to remove governments and revise policy peacefully. A political institution should be judged partly by how it behaves after error becomes visible.

The distinction also applies inside laboratories and companies. An organization may celebrate innovation while punishing the person who reports a failed assumption. It then remains culturally static beneath modern tools. A genuinely dynamic organization preserves channels through which inconvenient evidence can change plans, incentives, and leadership decisions.

Tolerance alone is insufficient. Criticism must be connected to explanation, and alternatives must be allowed to improve through contact with reality. Dynamic institutions need memory as well as openness: error correction works when useful knowledge survives while defective parts can be replaced.


8. Why Physics, Beauty, and Culture Belong in One Book

The book’s range can initially feel disorienting. Chapters on infinity in mathematics, Everettian quantum theory, free choice, floral beauty, cultural evolution, and sustainability seem to belong to different books. Deutsch treats them as tests of one worldview.

Quantum theory provides his most controversial physical case. He defends the Everett or multiverse interpretation and argues that quantum phenomena should be explained as objective features of reality, not reduced to rules for predicting observations. Mathematics supplies examples of finite procedures whose explanatory reach opens onto infinite structures. Aesthetics raises the possibility that beauty contains objective problems and improvements rather than being exhausted by personal preference.

Culture extends the evolutionary story. Genes contain knowledge created by biological variation and selection, but they do not understand what they encode. Human creativity introduces explanatory knowledge: ideas can be criticized for their content, intentionally redesigned, and transmitted in ways that change the selection environment itself.

The unifying move is realism. Deutsch resists declaring difficult domains unreal, subjective, or unknowable merely because current explanations are weak. Black holes, mathematical abstractions, moral improvement, and beauty pose different problems, yet each invites better explanations rather than a retreat to authority or relativism.

Readers need not accept every component to benefit from the architecture. The multiverse interpretation, objective aesthetics, and some claims about culture remain disputable. Fallibilism applies to Deutsch’s system too. Its strength lies in making disagreements productive: which problem does an alternative solve better, and which details can survive criticism without arbitrary repair?


9. A Contemporary Interpretation for Research and AI

Deutsch wrote before modern foundation models and today’s agentic AI systems. The following applications extend his framework; they are not claims that he directly made about current technology.

Research: From Scores to Explanatory Progress

Benchmarks, ablations, leaderboards, and statistical tests are valuable instruments for criticism. They become epistemically thin when the score substitutes for the problem. A model may improve an aggregate metric through dataset artifacts, extra scale, or a narrow evaluation advantage while leaving the mechanism obscure and the underlying failure unsolved.

Deutsch’s framework suggests a stronger research question: What new explanation has the result made possible? A useful contribution identifies why a method works, where it fails, which assumptions matter, and what observation would force revision. Even an empirical system that resists full mechanistic analysis can generate explanatory progress when experiments sharply distinguish competing accounts.

AI: Prediction, Explanation, and Creativity

Modern language models are extraordinary predictors and practical tools. Benchmark success, fluent output, and broad task coverage do not by themselves settle whether a system creates explanatory knowledge in Deutsch’s sense. That is an open scientific and philosophical problem, not a label that can be inferred from one behavior.

His approach directs attention toward processes: Can the system formulate a problem that was not supplied? Can it propose explanations whose details are constrained by reality, expose them to criticism, recognize a failed framing, and create a better one? Can it transfer criticism across domains without merely protecting its previous answer?

This also changes human–AI workflow design. An AI assistant is most valuable when it shortens the conjecture–criticism loop while keeping claims inspectable. It can generate alternatives, search for counterexamples, run experiments, and reveal inconsistencies. The human contribution remains problem selection, judgment, responsibility, and the construction of institutions in which errors can be surfaced safely.

Robotics: Open Worlds Need Error-Correcting Agents

Robotics makes the gap between fixed success and open-ended knowledge visible. A policy can master a benchmark distribution and fail after a small change in objects, goals, or social context. Scaling experience may reduce the gap, while deployment continues to produce situations absent from training.

An error-correcting robot would need more than robustness. It would identify surprise as a problem, generate candidate causes, seek informative evidence, ask for help, and revise its representation. This is still a research program rather than a solved architecture. Deutsch’s philosophy provides a useful criterion: intelligence should be evaluated by its capacity to create and correct explanations, not only by the number of tasks already inside its repertoire.


10. How to Read the Optimism Critically

The book is powerful partly because it pushes its claims farther than most readers will immediately accept. That ambition creates several productive tensions.

First, “hard to vary” is a standard for judging explanations, not a mechanical algorithm for discovering the correct one. People can disagree about which details are arbitrary, and scientific communities can preserve elegant errors for long periods. Social conditions, incentives, and power shape which criticism is heard.

Second, classifying evils as soluble problems does not make solutions technically or politically easy. Some conflicts involve incompatible interests, delayed harm, coordination failure, and institutions that reward the status quo. Knowledge is necessary, while implementation may require trust, authority, resources, and moral courage.

Third, unbounded progress is a claim about possibility under physical law, not a guarantee about history. Civilizations can destroy knowledge, close criticism, or fail to act in time. Optimism creates obligations because preventable catastrophe cannot be outsourced to destiny.

Finally, the book’s confidence in objective progress asks the reader to explain standards of moral and aesthetic improvement with the same care demanded in physics. Deutsch opens those domains to rational criticism; he does not supply a finished theory of each one.

These tensions strengthen the book when they are treated as new problem situations. A philosophy of fallibility should generate questions it cannot answer in advance.


Conclusion: Infinity as an Error-Correcting Tradition

The Beginning of Infinity is ultimately a book about beginnings. A good explanation begins a larger field of questions. A universal machine begins a repertoire no designer needs to enumerate. A dynamic society begins a history whose institutions can be deliberately improved. Optimism begins when a difficulty is treated as a problem in knowledge rather than a sentence imposed by fate.

Closed orientationOpen-ended orientation
Protect an explanation from revisionMake errors easier to expose
Derive knowledge from authority or raw dataCreate conjectures and test them critically
Treat present resources as a fixed inventoryAsk what new knowledge could make possible
Predict a problem-free final stateExpect new problems and improve the means to solve them
Optimize the accepted benchmarkUse results to deepen explanation
Preserve institutions by suppressing variationPreserve society through peaceful correction

Read beside James P. Carse’s Finite and Infinite Games, the two books offer complementary meanings of infinity. Carse asks how we keep participation and possibility alive; Deutsch asks how explanatory knowledge can continue to grow. One emphasizes the continuation of play, the other the correction of error. Together they suggest a demanding principle for research, organizations, and technology: keep the future open, and keep improving our ability to understand it.

The book’s most useful question is therefore not whether progress will happen automatically. It is: Are we building cultures in which a serious error can become the beginning of a better explanation?


Further Reading

本文支持通过网站顶部语言切换按钮在 English / 中文 间切换。

引言:进步为什么会开始?

戴维·多伊奇的《无穷的开始:世界进步的本源》(The Beginning of Infinity: Explanations That Transform the World)从一个历史上的不对称现象出发。人类已经存在了漫长时间,但快速且持续的进步直到科学革命与启蒙运动前后才出现。那些数千年间变化缓慢的社会,为什么能在几个世纪内连续改造科学、技术、政治、道德和日常生活?

多伊奇给出的答案是一套知识哲学。当一种文化开始寻找好的解释(good explanations),允许这些解释接受批评,并在更好的解释出现后替换旧解释,进步便由此启动。观察与实验极其重要,却不会替人制造理论。知识经由创造性的猜想与纠错而增长。

这本书的野心很大。全书十八章穿越认识论、量子物理、计算、数学、生物演化、人工智能、审美、文化、政治制度与环境思想。把这些主题连接起来的是一个主张:解释性知识是一种真实的物理力量;除了自然规律,它的潜在影响范围没有已知的上界。

因此,书名中的“无穷”并不表示人类必然永生,或者终有一天会知道一切。它意味着问题不会被最终穷尽,解释没有最后版本,改善也不必存在终极阶段。任何地方,只要开放式创造知识的条件刚刚形成,那里就出现了一个“无穷的开始”。


1. 好的解释很难被随意改动

多伊奇哲学的核心单位是好的解释。一个解释如果真正说明了现象为何如此,它的细节会受到问题本身的约束。随意改变其中的重要部分,解释就会失效。多伊奇把这种性质称为难以改变(hard to vary)

古代关于季节的故事可以说明相反情况。如果冬天来自某位神的悲伤,我们几乎可以任意修改故事,同时继续迎合相同的观察:换一位神、换一种情绪、换一个事件顺序。这样的灵活性看起来像强大的解释力,实际正是弱点,因为现象本身没有选中故事里的任何细节。

地轴倾斜造成季节的解释完全不同。倾角、公转、阳光入射角与南北半球相反的季节彼此约束。改变一个部分会在其他地方产生后果,并常常直接与观察冲突。它的组成部分通过对现实的说明连接起来,而不是为了保护结论被临时拼装在一起。

这个标准也适用于实验室之外。“团队缺乏主人翁意识”是一种容易复述、也容易修改的组织解释。更强的分析会指出哪些决策权、激励、信息流和反馈延迟造成了某次具体失败。它仍可能出错,但至少为批评提供了可以着力的位置。

好的解释并不等同于复杂解释、数学模型、准确预测或专家共识。公式可能预测一种模式,却没有说明其机制;简单观念也可能拥有极大的解释范围。关键检验在于:它是否无需任意调整便解决了问题,并且经受住了严肃的改进尝试。


2. 知识通过猜想与批评增长

多伊奇发展了卡尔·波普尔的可错论认识论(fallibilist epistemology)。我们无法从重复观察中直接推导一般理论。经验提供问题、约束和检验,人创造猜想。恒星理论不能从光点中直接读出,正如人类不可能仅从数千年观察物体坠落的经验里抽取出飞行理论。

知识创造的循环没有受保证的第一步,也没有最终的合格证书:

  1. 我们在已有知识中遇到问题。
  2. 我们提出可能解决问题的解释。
  3. 我们通过论证、实验和比较批评这些方案。
  4. 我们消除已经发现的错误,并继承新的问题。

证据不可或缺,因为它帮助我们在解释之间做出选择。但一个理论通过检验,并不会因此成为确定无疑或得到终极证明的真理。每个解释都应继续接受批评,包括最深层的物理理论,以及我们用来判断理论的哲学规则。

这也是多伊奇同时拒绝权威主义与激进怀疑主义的原因。权威主义的问题在于,没有任何个人、传统、数据集或方法永不犯错。怀疑主义的问题则是把可能出错当成否认改善的理由。我们可能始终有错,却仍能知道一个解释解决了另一个解释没有触及的问题。

可错性因此具有生产力。科学共同体的理性来自有组织的纠错,而非个体科学家的纯粹无误。同行评议、重复实验、开放数据、对抗式讨论和和平的制度变革,都在让错误更容易被发现和清除时体现价值。它们的正当性来自可纠正性,而不是声望。


3. 解释的范围

人类接触证据的范围既局部又狭窄,最好的解释却能远远超出它。从进入望远镜的光线出发,我们推断出人类永远不会到访的恒星的成分与历史;从岩石和基因留下的痕迹出发,我们重建没有任何观察者在场的远古事件。理论把此时此地一次微小的互动,与跨越巨大时空尺度的结构连接起来。

多伊奇以这种解释范围挑战经验主义把心智视作被动接收器的图景。仪器并非只是扩大感官。仪器读数只有进入解释理论之后才获得意义。射电望远镜、粒子探测器或显微镜之所以让我们感知现实,是因为多层理论告诉我们信号代表什么,以及怎样纠正其中的误差。

解释还会跨越领域。计算把物理机器与抽象算法连接起来;演化既解释生物适应,也有助于理解非遗传观念如何延续;认识论则把科学发现与政治制度联系起来,因为两者都依赖发现并纠正错误。

这不意味着一个理论可以把现实的所有层次还原成粒子物理。多伊奇承认抽象事物的实在性。数字、程序、制度和知识需要物理载体,但更高层次的解释仍然不可替代。即使知道计算机中每个原子的位置,也不能自动解释它正在运行哪种算法,以及程序为什么能解决问题。

解释范围既值得惊叹,也要求纪律。广泛理论必须通过在批评下解决相互关联的问题来赢得其范围。仅仅把同一套流行词汇贴到许多主题上,只会得到缺乏解释深度的表面广度。


4. 向普适性的跃迁

全书反复出现的一个深层模式,是向普适性的跃迁(jump to universality)。许多系统会逐渐改进,直到某种特定架构跨过阈值,原则上能够完成一整类变换。

通用计算机是最清晰的例子。只要给予足够的时间、存储和正确程序,它就能执行任何其他物理计算机可以执行的计算。普适性并不表示每种计算都容易、迅速或已经有人知道如何完成。它意味着面对一个新的可计算任务时,我们无需再为它制造一种原理上完全不同的机器。

多伊奇认为,人是普适解释者(universal explainers)。人类并未被限制在基因编码的那一小组知识里,只能适应某个祖先环境。我们能够创造关于黑洞、素数、灭绝生物、宪法以及尚不存在的技术的解释。人依然有限、易错,也依赖文化,但我们可能理解的范围并未绑定在固定的认知生态位上。

语言和文字进一步放大这种普适性。它们使解释性知识能够离开单个心智而延续,跨越世代接受批评,并与远方的观念重新组合。一个动态文化会成为分布式纠错系统,而它未来能够拥有的知识清单无法提前列出。

普适性不能与全能混淆。自然规律仍区分可能与不可能的变换;计算复杂度、能量、时间与缺失的知识都是真实约束。多伊奇提出的是一个更细致的主张:或许不存在一张固定目录,预先列出人类式解释原则上永远无法触及的主题。


5. 知识改变什么可以成为资源

人们谈论资源时,常把“有用”当成物质内部固有的属性。多伊奇强调,资源是物质加上改造它的知识。对没有核物理学的社会来说,铀不是能源;沙子也只有在半导体理论与工艺出现之后,才成为计算时代的关键材料。

这种观点改变了人与环境的关系。人类的生存从来不是依赖永远维持某个自然给定的生态位。衣物、农业、卫生、医学、建筑和能源系统,让人能够生活在身体并未通过生物演化适应的环境中。知识创造新选择,也同时创造新责任。

限制依然真实。自然规律无法谈判,局部材料、能源、时间、生态系统与制度能力都可能成为硬约束。多伊奇的重点在于,我们很少能提前知道某个表面限制究竟是根本不可能,还是缺少知识的症状。短缺有时可以通过替代、提高效率、循环利用、发现新来源,或者一种尚未被提出的变换来解决。

这也解释了人在他世界观中的特殊位置。知识可以产生与承载它的物质质量极不相称的物理影响。一小段编码后的观念可以穿越世界,重组工厂、调动能源、治愈疾病,也可以摧毁城市。人在宇宙尺度上具有意义,因为人能创造改变物理事件可能性的解释。

这个主张带有明确的伦理锋芒。更大的影响范围既会放大纠错,也会放大错误。扩展解法的同一种普适性也能扩展伤害,因此,能够接受批评的制度必须与技术力量共同成长。


6. 乐观主义是一套关于问题的理论

多伊奇的乐观主义常被误读为“所有事情自然都会变好”的预测。他真正提出的是一种方法论。问题不可避免,因为每个解法都会改变处境并暴露新的问题。进步取决于我们把问题视作可以通过新知识解决,除非有好的解释证明它确实不可解决。

他把这种态度压缩成两句极短的话:

“问题不可避免。问题可以解决。”

第一句阻止乌托邦主义。任何政策、技术、领袖或社会安排都无法终结历史。一种解法可能产生副作用,不公平地分配成本,或者暴露更深的困难。第二句阻止宿命论。当下的失败不能证明永远不可能;“现在没有已知解法”也不能解释为什么解法不可能存在。

因此,乐观主义要求积极的纠错。它追问缺少什么知识,哪些假设遮蔽了其他可能,怎样用实验区分方案,以及方案失败时制度如何恢复。当悲观主义悄悄把“我们不知道怎样做”变成“这件事做不到”,它便带来了认知上的危险。

这套哲学也拒绝把预测当成长远规划的基础。未来知识无法提前获知;如果我们已经拥有其内容,它就是当前知识。我们可以建设财富、科学能力、冗余、开放沟通以及允许快速批评的制度。韧性来自创造新回应的能力,而这些回应无法被任何旧计划提前写完。


7. 静态社会与动态社会

多伊奇认为,人类历史上的大多数社会都是静态社会(static societies):它们变化缓慢,是因为文化知识以压制变异的形式传递。仪式、禁忌、权威与惩罚让行为保持足够忠实,使社会得以复制自身。这样的文化可以包含复杂知识,同时让有意改善变得异常困难。

动态社会(dynamic society)维护批评传统。人们可以讨论、挑战并有意识地改变习俗与制度,而无需摧毁整个社会秩序。启蒙运动的重要性在于强化了这种“元传统”:没有任何思想来源高于批评,改善也不需要最终权威。

民主在这个解释中体现为纠错机制。它最重要的成就不是稳定选出唯一英明的统治者,而是让社会可以和平撤换政府并修改政策。评价一种政治制度时,应当观察它在错误暴露之后如何反应。

这个区分也适用于实验室和公司。一个组织可能不断赞美创新,却惩罚报告错误假设的人。它在现代工具之下仍保持静态文化。真正动态的组织保留信息渠道,让不方便的证据能够改变计划、激励机制和领导决策。

只有宽容还不够。批评必须与解释相连,替代方案也需要在接触现实的过程中改善。动态制度既需要开放,也需要记忆:有用知识应当延续,其中有缺陷的部分则可以被替换。


8. 为什么物理、审美与文化会出现在同一本书里

这本书的跨度最初可能令人迷惑。关于数学无穷、埃弗里特量子理论、自由选择、花朵之美、文化演化和可持续性的章节,似乎属于几本不同的书。多伊奇把它们视作对同一个世界观的检验。

量子理论是他最有争议的物理学案例。他捍卫埃弗里特解释或多重宇宙解释,主张量子现象应当被解释成现实的客观特征,而不能被压缩成预测观察结果的规则。数学提供了有限程序在解释上通向无限结构的例子。审美问题则让他追问,美是否包含客观问题与改进,而非完全由个人偏好决定。

文化延伸了演化故事。基因包含由生物变异和选择创造的知识,却不理解自己编码的内容。人类创造力引入了解释性知识:观念可以依据内容受到批评,被有意识地重新设计,并以改变选择环境本身的方式传播。

把这些内容统一起来的是实在论。多伊奇反对仅仅因为当前解释薄弱,就把困难领域宣布为虚假、主观或不可知。黑洞、数学抽象、道德进步与美提出不同的问题,但每个问题都邀请更好的解释,而不是退回权威或相对主义。

读者无需接受每个组成部分,仍可以从整体架构中受益。多重宇宙解释、客观审美以及部分文化主张仍有争议。可错论同样适用于多伊奇自己的体系。它的力量在于让分歧产生知识:一个替代解释更好地解决了什么问题?哪些细节无需任意修补也能经受批评?


9. 对科研与AI的当代延伸

多伊奇写作本书时,现代基础模型和今天的智能体式 AI 系统尚未出现。以下内容是对他思想框架的延伸,并非他直接针对当前技术提出的主张。

科研:从分数走向解释性进步

基准、消融实验、排行榜和统计检验都是有价值的批评工具。当分数取代问题本身,它们在认识论上就会变得单薄。模型可能依靠数据集伪影、更大规模或狭窄的评估优势提高综合指标,却没有说明机制,也没有解决底层失败。

多伊奇的框架提示一个更强的研究问题:这项结果让什么新解释成为可能? 有价值的贡献会指出方法为何工作、在哪里失败、哪些假设重要,以及什么观察会迫使我们修改理解。即使某个经验系统暂时难以得到完整的机制分析,只要实验能清晰区分相互竞争的解释,它依然可以带来解释性进步。

AI:预测、解释与创造力

现代语言模型是非凡的预测器,也是实用工具。基准成功、流畅输出与广泛任务覆盖本身,并不能判定一个系统是否正在以多伊奇所说的方式创造解释性知识。这是一个开放的科学与哲学问题,不能从单一行为直接贴上结论。

他的思想让我们把注意力转向过程:系统能否形成一个没有被直接提供的问题?能否提出细节受到现实约束的解释,让它们接受批评,识别错误的问题框架,再创造一个更好的框架?能否把批评迁移到其他领域,同时不只是为旧答案辩护?

这也会改变人机工作流的设计。当 AI 助手能够缩短“猜想—批评”循环,同时保持主张可检查,它最有价值。它可以生成替代方案、搜索反例、运行实验并暴露矛盾。人的贡献仍包括问题选择、判断、责任,以及建设让错误可以安全暴露的制度。

机器人:开放世界需要会纠错的智能体

机器人领域让固定成功与开放式知识之间的差距非常可见。一套策略可以掌握基准分布,却在物体、目标或社会情境稍有变化后失败。扩大经验规模或许能缩小差距,但真实部署仍会不断产生训练中没有出现的情形。

一个会纠错的机器人需要的不只是鲁棒性。它还要把意外识别为问题,生成可能原因,寻找有信息量的证据,主动求助,并修改自己的表征。这仍是一项研究计划,而不是已经完成的架构。多伊奇的哲学提供了一个有用标准:评价智能时应观察它创造和纠正解释的能力,而不只是统计已有能力清单中的任务数量。


10. 如何批判性地阅读这种乐观主义

这本书的力量部分来自它把主张推进到多数读者不会立即接受的位置。这种野心也制造了几组富有生产力的张力。

第一,“难以改变”是评价解释的标准,不是自动发现正确解释的机械算法。人们可能对哪些细节属于任意添加存在分歧,科学共同体也可能长期保留优雅的错误。社会条件、激励与权力会影响哪些批评能够被听见。

第二,把恶归类为可解决的问题,并不会让解法在技术或政治上变得容易。一些冲突包含利益不兼容、延迟伤害、协调失败,以及奖励现状的制度。知识不可或缺,但实施还可能需要信任、权威、资源和道德勇气。

第三,无限进步是一个关于自然规律所允许之可能性的主张,不是对历史结果的保证。文明可以摧毁知识、封闭批评,也可能来不及行动。乐观主义因此带来责任,因为可以避免的灾难不能交给命运处理。

最后,本书对客观进步的信心,也要求读者像对待物理问题一样,认真解释道德与审美改善的标准。多伊奇把这些领域开放给理性批评,却没有为每个领域提供完成版理论。

当这些张力被视作新的问题情境,它们反而加强了这本书。一套关于可错性的哲学,理应产生它无法预先回答的问题。


结语:作为纠错传统的“无穷”

《无穷的开始》归根结底是一本关于“开始”的书。一个好的解释会开启更大的问题空间;一台通用机器会开启无需设计者逐项列举的能力集合;动态社会会开启制度可以被有意改善的历史;当困难被当作知识问题,而非命运判决,乐观主义便由此开始。

封闭取向开放取向
保护解释免受修改让错误更容易暴露
从权威或原始数据中直接获得知识创造猜想并以批评检验
把现有资源视作固定清单追问新知识可以创造什么可能
预测一个没有问题的终极状态预期新问题,并改善解决问题的手段
优化已经接受的基准用结果深化解释
通过压制变化维持制度通过和平纠错维持社会

如果把本书与詹姆斯·卡斯的《有限与无限的游戏》并读,两本书给出了彼此补充的“无限”。卡斯追问如何让参与和可能性延续,多伊奇追问解释性知识如何持续增长;一本强调让游戏继续,另一本强调让错误得到纠正。它们共同为科研、组织与技术提出一个严格原则:让未来保持开放,也不断提高我们理解未来的能力。

因此,这本书最有用的问题并不是进步会不会自动发生,而是:我们是否正在建设这样的文化——一个严重错误能够成为更好解释的起点?


延伸阅读