A Letter to the Public: Beyond Results, There Is a History of Their Making
English
Word count: 2,778
To everyone,
We are entering an age in which the production of intellectual artifacts is becoming dramatically easier. A paragraph can be drafted in seconds. An image can be generated on demand. Code, summaries, diagrams, alternative formulations, and even long sequences of apparently coherent argument can now be produced at a speed that would have been difficult to imagine only a few years ago.
There is much to welcome in this. Greater generative capacity can lower barriers to expression, allow a person to explore unfamiliar fields, accelerate comparison and revision, and give individuals access to intellectual tools that once required large organizations or unusually abundant time. It would be a mistake to treat every increase in generative capacity as a decline in knowledge.
But every expansion of capacity changes the location of scarcity.
When finished intellectual products were expensive to produce, the product itself carried some information about the process behind it. A long technical book did not prove seriousness, truth, or quality; a lengthy paper could still be wrong, derivative, or empty. Yet the scale and form of a work often functioned as weak signals of time, sustained attention, accumulated reading, and repeated revision.
That relationship is rapidly weakening.
A polished report may now be generated very quickly. A long manuscript can be assembled with unprecedented speed. At the same time, a work that has developed through months or years of questioning, failed attempts, counterarguments, revisions, formal experiments, and changes of judgment may arrive before the public in exactly the same visible form: a finished PDF.
Two results may look similar while having radically different histories.
The important change, then, is not simply that machines can generate more text. It is that the visible result is becoming a weaker description of the process that produced it.
When generation becomes abundant, the history of generation becomes more important.
This problem belongs to a much older story about knowledge. Vannevar Bush, writing in 1945, argued that the usefulness of the human record depended not only on storage but on the possibility of extending, consulting, and following it. His imagined “memex” was memorable not because it merely stored documents, but because it preserved associative trails among them.[^1]
That insight is more profound than a technical proposal about information retrieval. Knowledge does not live only in isolated objects. It also lives in relations among objects: what followed what, what answered what, what changed because of what, what was rejected, what was retained, and which earlier path made a later path intelligible.
The philosophy of science reached similar conclusions from another direction. Karl Popper emphasized conjecture and criticism; Thomas Kuhn made scientific change and the historical transformation of standards central to our understanding of science; Imre Lakatos explicitly tied the appraisal of research programmes to their histories. Lakatos put the point sharply: “Philosophy of science without history of science is empty; history of science without philosophy of science is blind.”[^2]
These traditions differ greatly, but they share a lesson that matters now: knowledge is not adequately described by its surviving propositions alone. A conclusion has a history. A model has a history. A distinction has a history. Sometimes that history is epistemically important because it tells us which alternatives were considered, which objections were answered, which assumptions were discovered, and where uncertainty remains.
Artificial intelligence makes this old insight newly urgent because it changes the economics of intellectual production.
The cost of producing a plausible expression is falling much faster than the cost of producing a sound judgment.
It is becoming cheap to generate ten possible arguments. It remains difficult to know which one deserves to survive.
It is becoming cheap to generate a formal model. It remains difficult to judge whether the model clarifies the phenomenon, merely redescribes it, or quietly imports assumptions that the evidence cannot support.
It is becoming cheap to generate an answer. It remains difficult to identify the question that is actually worth asking.
It is becoming cheap to generate citations, summaries, and literature-shaped prose. It remains difficult to determine whether the cited sources support the precise claims being made, whether important counterevidence has been ignored, and whether an apparent consensus is only an artifact of selection.
The scarcity is therefore moving. Generative capacity becomes abundant; judgment, orientation, criticism, provenance, and historical intelligibility become relatively scarcer.
This is why I do not think the most useful distinction for the coming knowledge environment will be simply “human-made” versus “machine-made.” Serious intellectual work will increasingly involve mixed processes. A person may formulate a problem. A model may propose an answer. The person may reject its assumptions. Another model may produce a different formalization. Evidence may force both to be revised. A later conversation may reveal that the original problem was badly framed. The resulting theory may contain sentences drafted by machines, structures designed by humans, mathematical objects suggested interactively, sources located computationally, and final judgments that emerged only after repeated disagreement.
What, then, tells us what kind of intellectual work occurred?
Counting keystrokes will not. Counting machine-generated words will not. Looking only at the final artifact will increasingly fail as well.
The more relevant object is the trajectory: the sequence of questions, transformations, dependencies, criticisms, decisions, and revisions through which the artifact became what it is.
Computer science already has a useful language for one part of this problem: provenance. The W3C PROV framework defines provenance in terms of the entities, activities, and people involved in producing a piece of data or a thing, and explicitly notes that provenance can support judgments about quality, reliability, and trustworthiness.[^3]
This is important because provenance shifts our attention from the object alone to the relations that produced it. We ask not only “What is this?” but also “What did it derive from? What activities changed it? Which agents participated? Which earlier entities were transformed into later ones?”
A similar movement is already visible in digital media. The Coalition for Content Provenance and Authenticity has developed Content Credentials because increasingly powerful creation and editing technologies make the provenance of media relevant to transparency, understanding, and trust. Its specification emphasizes information about how an asset was created and changed.[^4]
The analogy with scholarship should be used carefully. A philosophical argument is not a photograph. Intellectual development cannot be reduced to cryptographic metadata, and a signed history cannot certify that an idea is true. Provenance is not a substitute for criticism.
But that limitation strengthens rather than weakens the argument. The purpose of recording a trajectory should not be to let infrastructure declare, “This work is trustworthy.” It should be to preserve enough of the relevant history that people can make better judgments for themselves.
A record of generation should support judgment, not replace it.
This distinction matters because any infrastructure of trust can become an infrastructure of control. A system that records every draft, every private doubt, every conversation, and every abandoned idea would damage the freedom that inquiry requires. Intellectual life needs places where one can be confused, wrong, speculative, unfinished, and even foolish without creating a permanent public identity around every attempt.
So I am not arguing for total logging.
I am arguing for selective reconstructability.
Important intellectual work should be able, where appropriate, to leave enough of a record that its major development can later be understood. A reader may not need every discarded sentence, but it can matter that a central hypothesis was abandoned after a specific objection. A future researcher may not need a complete transcript of every discussion, but it can matter that a formal model changed because its original assumptions were found to be too strong. A community may not need access to private notes, but it can matter that a concept developed across several dated working papers before it acquired its final form.
This kind of record could connect versions, papers, datasets, models, major conceptual changes, evidential revisions, rejected hypotheses, and the relationships among them. It could identify where a claim became narrower, where an analogy was demoted from explanation to heuristic, where a source changed the direction of an argument, or where an AI-generated proposal was accepted, transformed, or rejected through human judgment.
The point is not bureaucratic completeness. The point is to preserve the difference between a result and the path by which it became a result.
This matters first for trust, but trust is only one consequence.
It also matters for the recognition of value. When we see only a final result, we tend to locate value in the result alone. Yet inquiry often creates value before it reaches a publishable endpoint. A failed model can reveal the structure of a problem. A negative result can close an attractive but misleading route. An objection can force a theory to become more precise. A sequence of revisions can show future researchers where the real difficulty lies. An abandoned distinction can prevent years of conceptual confusion from being repeated.
If those trajectories disappear, later communities inherit the answer while losing part of the learning that made the answer possible.
It matters for circulation as well. Knowledge circulates more effectively when later users can understand not only what was claimed, but under what assumptions, against which alternatives, and with what unresolved limitations. A formula detached from the history of its construction is easy to transport and easy to misuse. A concept detached from the problem that called it into existence can become a slogan. A theory detached from its failed alternatives can appear more universal than it ever was.
And it matters for reproduction. To reproduce intellectual work is not always to obtain the same number from the same code. In conceptual and theoretical inquiry, reproduction can mean reconstructing the conditions under which a judgment was made and seeing whether another investigator, facing the same alternatives and evidence, finds the trajectory intelligible.
There is also a social dimension.
Scholarly communities do not merely store statements; they create social facts around them. They decide, often informally, which ideas are associated with which traditions, which contributions count as important, which failures disappear from memory, and which narratives become standard. Robert Merton’s classic work on priority in science showed how deeply recognition and originality became institutionalized within scientific life.[^5]
A richer historical record should not intensify the race to declare conceptual ownership. I do not think every idea needs a proprietor, and independent discovery is a normal feature of intellectual life. Nor should infrastructure manufacture a social identity of “the first person” around every concept.
But refusing to turn ideas into property does not require us to become indifferent to historical fact.
If an argument appeared in a dated document, that is part of history. If several people reached similar ideas independently, that too is part of history. If a later narrative falsely erases earlier work, the existence of a reliable record gives a community something better than competing assertions. It allows claims about the past to remain answerable to evidence.
This is an important distinction: preserving history is different from owning ideas.
A healthy infrastructure would therefore neither award conceptual sovereignty nor erase provenance. It would make it easier to say: here are the surviving records; here are their dates; here are the relations among them; here is what changed. The community can then reason about attribution, influence, independence, and significance with a stronger evidential basis.
The need extends beyond scholarship in the narrow sense. Organizations, public institutions, technical standards, policy frameworks, collaborative projects, and communities all produce social facts whose present form can conceal their developmental history. A rule may look inevitable after it has stabilized. An institution may forget the alternatives it once considered. A standard may survive while the reasons for its design disappear. A collective may remember a decision but lose the history of disagreement that made the decision legitimate.
In each case, the present object can become misleading when detached from the relations that generated it.
This is why the question of knowledge infrastructure is, in the end, a question about what kind of historical memory a high-generation society wants to possess.
The twentieth century built extraordinary infrastructures for storing and retrieving finished knowledge. Libraries, journals, citation systems, databases, repositories, and eventually the Web greatly expanded access to the inherited record. Bush’s problem was how to navigate an overwhelming store of knowledge. Our problem includes his, but adds another layer: when the production of plausible new material accelerates dramatically, we must also preserve enough information to distinguish among different histories of production.
The next generation of repositories may therefore need to store more than artifacts. They may need to represent relationships among artifacts and the transformations that connect them. Version control is one beginning. Persistent identifiers are another. Open notebooks, preregistration, registered reports, data provenance, computational environments, and transparent revision histories each capture parts of the same broader intuition: the route to a result can matter to our understanding of the result.
No single system will solve this. Nor should one system become the universal judge of intellectual legitimacy. Different domains require different levels of traceability. Mathematics, ethnography, experimental biology, philosophy, historical research, software engineering, and artistic practice generate evidence and judgment in very different ways.
What can be shared is a principle.
Where the history of generation matters to interpretation, trust, reuse, or accountability, our institutions should make it possible to preserve that history.
This principle also changes how we think about AI disclosure.
A binary label such as “AI used” or “AI not used” may be useful, but it is increasingly insufficient. There is a profound difference between asking a system to produce a finished manuscript from a short instruction and using systems repeatedly within a year-long research process in which questions are reformulated, arguments contested, formalizations tested, sources checked, structures redesigned, and claims revised.
Both processes involve AI. The label alone cannot describe the difference.
A more mature culture of disclosure would ask what roles different agents played. Which parts involved generation? Which involved verification? Which involved human judgment? Which models or tools were used for exploration? Which claims were materially revised after criticism? What remained uncertain? What was abandoned?
Again, the purpose is not to calculate a percentage of “human authorship.” Intellectual responsibility cannot be measured by word count. The purpose is to make the structure of the process more intelligible.
And there is another reason to care.
Future generations will study this period.
They will ask how knowledge changed when humans began to think, write, model, and argue routinely with generative systems. If we preserve only final outputs, we will leave them a distorted archive. They will know what was published, but not how human and machine agency were actually entangled in its production. They may be unable to distinguish a one-shot generation from a long dialogical inquiry, or an automated synthesis from a theory that underwent repeated human criticism and reconstruction.
We have an opportunity to leave a better record.
That record should remain modest. It should protect privacy. It should permit intellectual experimentation. It should not turn every act of thought into a performance. It should not create a new bureaucracy of authenticity in which only fully logged minds are trusted.
But it should recognize that the history of making has become part of the information environment in which knowledge is judged.
The central challenge of the AI era is therefore not only how to generate more knowledge-like material. We are already learning how to do that extraordinarily well.
The harder challenge is how to preserve the conditions under which abundance can still support understanding.
That requires criticism. It requires judgment. It requires institutions capable of remembering revisions rather than only celebrating outputs. It requires records that can show relationships without pretending that relationships prove truth. It requires enough historical continuity that future users can tell the difference between a conclusion and a conclusion that survived a history of serious challenge.
Knowledge has always been more than its final sentences.
Every mature idea carries traces of alternatives that failed, distinctions that changed, evidence that resisted, people who disagreed, tools that enabled new views, and judgments that selected one path among many.
For much of history, those traces were difficult to preserve. In the coming era, preservation is technically easier, yet conceptually more necessary.
We should use that possibility wisely.
We do not need to preserve everything.
We do need to preserve enough that knowledge retains a history.
Because beyond results, there is a history of their making.
And that history, too, is part of what a knowledge society should learn how to value.
With respect and hope,
Wanhong Huang
[^1]: Vannevar Bush, “As We May Think,” The Atlantic Monthly, July 1945. Bush wrote that a useful scientific record “must be continuously extended, it must be stored, and above all it must be consulted,” and later developed the idea of associative “trails.” A historical web edition is preserved by the W3C: https://www.w3.org/History/1945/vbush/.
[^2]: Imre Lakatos, “History of Science and Its Rational Reconstructions,” in The Methodology of Scientific Research Programmes, ed. John Worrall and Gregory Currie. Cambridge University Press. See https://www.cambridge.org/core/books/abs/methodology-of-scientific-research-programmes/history-of-science-and-its-rational-reconstructions/9D334AB29EE0B59916FACE08D80DF23E. For Kuhn’s account of scientific development and change, see the Stanford Encyclopedia of Philosophy entry: https://plato.stanford.edu/entries/thomas-kuhn/.
[^3]: W3C, PROV-Overview: An Overview of the PROV Family of Documents, 2013: https://www.w3.org/TR/prov-overview/. The W3C PROV Primer also emphasizes reconstruction of how something was generated: https://www.w3.org/TR/prov-primer/.
[^4]: Coalition for Content Provenance and Authenticity, Content Credentials: C2PA Technical Specification, version 2.4. The specification states that, amid increasingly powerful creation and editing techniques, provenance is critical to transparency, understanding, and trust. See https://spec.c2pa.org/specifications/specifications/2.4/specs/C2PA_Specification.html.
[^5]: Robert K. Merton, “Priorities in Scientific Discovery: A Chapter in the Sociology of Science,” American Sociological Review 22, no. 6 (1957): 635–659.
中文
中文字符数：3,638
给公众的一封信：结果之外，还有生成的历史
诸位：
我们正在进入一个结果越来越容易出现的时代。
一段文字、一张图像、一份分析、一段代码，甚至一本结构完整的书，都可以以过去难以想象的速度被生成、组合和重新表达。人工智能降低了许多知识生产活动的技术门槛，也扩大了个人探索问题、比较观点、尝试形式化和修改表达的能力。
这本身并不是一件坏事。恰恰相反，更多的人能够参与创造，更多的问题能够被尝试，更多原本因为资源不足而无法展开的思考有机会获得表达，这可能是知识生产方式的一次重要扩展。
但每当一种能力变得丰富，另一种东西往往会开始显得稀缺。
今天逐渐稀缺的，也许不是“结果”，而是结果的生成历史。
过去，我们当然也不能仅凭一本书的厚度、一篇论文的长度，判断它是否可靠。一千页文字从来不自动意味着一千页思想。但一个复杂成果至少往往隐约携带着某些时间信号：阅读、积累、修改、讨论、失败以及重新组织。
这种信号正在迅速减弱。
今天，两个外表几乎相同的成果，可能拥有完全不同的历史。
一个可能来自一次很短的自动生成；另一个可能来自一年甚至更久的问题意识，经历许多次讨论、反论、失败的模型、被放弃的路线、证据修正、概念重构，最后才形成现在的形态。
可当这一切都结束以后，我们看到的可能只是两个相似的 PDF。
于是，一个非常重要的关系开始断裂：
我们所看见的结果，越来越不足以告诉我们这个结果是怎样成为它自己的。
我认为，这是人工智能时代值得公共讨论的一个基础问题。
知识从来不只是最终留下来的那几句话。
波普尔将知识的发展理解为猜想与批判之间不断发生的过程；库恩让我们看到科学对象、问题和判断标准本身具有历史；拉卡托斯则提醒我们，研究纲领需要在时间中理解，一个理论的意义并不完全存在于某个静止的最终版本中。
这些传统彼此并不相同，但它们至少共同指向一件事：
知识的发展具有过程。
一个结论之所以成为今天的结论，往往因为它经历过一些事情。
它曾经遇到某个反例。
它曾经有一个更大胆的版本，后来被证据迫使收缩。
它曾经依赖一个漂亮的形式模型，后来发现那个模型把太多假设偷偷带进了问题。
它曾经把两个概念混在一起，直到一次反论迫使它们被区分。
它也可能曾经失败。
而失败本身，也可能是知识。
如果我们只保存最后那个看起来最整洁的版本，那么我们保存的是结论，却可能同时丢失了理解结论所需要的一部分历史。
这种损失在生成能力变得廉价之后，会变得更加严重。
因为“能够生成”与“能够判断”不是同一种能力。
能够迅速给出十种理论并不意味着能够判断哪一种值得留下。
能够生成一个方程，不意味着知道这个方程是否真正对应了问题。
能够找到一种漂亮的类比，不意味着这种类比没有遮蔽重要差异。
能够写出完整的论证，也不意味着已经找到了它最强的反论。
在一个生成能力越来越丰富的时代，真正稀缺的可能逐渐变成另外一些东西：
判断，批评，选择，修订，来源，以及能够让后来的人重新理解这些判断如何发生的历史。
所以，我并不认为未来最重要的问题只是“这是不是 AI 写的”。
这个问题太粗糙了。
未来大量严肃的知识工作，很可能本来就是人与机器共同参与的。
一个人提出问题，机器给出方案；人发现其中的错误，再提出反论；机器帮助比较新的路径；人重新定义问题；一个形式模型被构造出来，又因为假设太强而被放弃；新的材料改变了原有判断；最后，整个理论结构被重新组织。
在这样的过程中，到底哪一句是“人的”，哪一句是“机器的”，也许并不是最有意义的问题。
更有意义的问题是：
这项知识工作经历了怎样的生成、批判、判断和转变？
这也是为什么，我越来越认为，我们需要重新思考“信任”。
信任从来都不是一个孤立对象自身携带的固定属性。
我们之所以相信一个研究成果，往往不仅因为它现在呈现成什么样，也因为我们知道它与什么证据发生过关系，经过怎样的检验，由哪些机构、方法、共同体和记录支撑，它是否接受过批评，以及后来的人是否能够重新检查其中的一部分过程。
换句话说，信任本来就具有关系性，也具有历史性。
只是过去，我们经常可以把这层历史隐藏在最终作品的背后。
现在，也许不行了。
计算机科学中有一个很有用的词：provenance，来源与生成谱系。
W3C 的 PROV 框架把一个对象的来源理解为参与其生成的实体、活动与行动者之间的关系，并明确指出，这些信息能够帮助人们判断质量、可靠性和可信度。
近年的数字内容基础设施也开始关注类似问题。C2PA 等体系尝试保存图像、音视频等内容经历过怎样的创建和修改。它的一个重要思想是：来源记录本身不替人决定某项内容“可信”还是“不可信”，它提供历史，使人拥有更好的判断条件。
我认为，知识也需要类似的转变。
但知识的生成史比文件元数据更加复杂。
我们已经拥有非常重要的基础设施：版本号、时间戳、DOI、Git 历史、预印本、开放仓库、数据存档。
它们已经保存了大量过去很容易消失的信息。
然而，一串 Version 1、Version 2、Version 3，并不自动告诉我们一个理论为什么发生变化。
第二版为什么和第一版不同？
是因为出现了新证据？
是因为发现了逻辑漏洞？
是因为一个反论？
是因为价值判断发生变化？
还是因为一个模型虽然形式上更优美，却被发现不适合原来的对象？
这些东西很少能够从文件本身恢复。
因此，我认为我们需要的并不是“记录一切”。
思想不应该成为全天候监控的对象。
人需要私人笔记，需要尚未成熟的猜想，也需要说错话、走错路而不必永久公开的空间。
真正值得追求的，是一种可重建性。
我们可以不保存每一句话，却保存重要的变化。
一个重要的问题何时出现。
一个核心概念如何改变。
一条研究路线为什么被放弃。
一个模型经历了怎样的修改。
一篇工作论文与后来一本书是什么关系。
哪些反论真正改变了理论。
哪些判断由人作出。
人工智能在哪些阶段参与了探索、整理或生成。
哪些来源后来被发现不足以支持原来的主张。
这样的记录不是日记。
它更接近一种知识发展的谱系。
而它的价值也不仅仅是防止误解。
它能够改变我们认识知识价值的方式。
如果只看最后结果，我们很容易把价值理解成最终交付出来的对象。
但生成过程本身也在创造价值。
一条失败的路线可能告诉后来的人什么地方走不通。
一个被放弃的理论可能让一个更好的问题出现。
一次反论可能使原来过于宏大的主张变得更谨慎、更准确。
一个长期没有答案的问题，也可能在很多年后因为新的技术、新的材料或新的关系而重新获得意义。
如果这些过程完全消失，后来的人就必须重复许多已经发生过的探索。
知识的再生产因此不仅依赖结论的传播，也依赖历史的保存。
这件事还关系到另一个容易引起争议的问题：思想的“优先性”。
科学社会学很早就注意到，学术共同体存在关于发现、承认和优先权的制度。默顿曾经详细讨论过科学中的 priority。
但我并不认为未来的知识基础设施应该成为一种更加精密的“谁是第一”的认证机器。
思想没有那么简单。
相似的问题可以在不同地方独立出现。
一个概念的形成可能依赖许多前人的工作。
一些思想并不存在清晰的单一出生时刻。
如果我们把生成历史记录的目的变成争夺首创，那么新的基础设施很可能只是把旧的竞争结构技术化。
历史记录还有另一种用途。
我们可以不主动制造“某个人是首创者”的社会事实，但仍然保存历史事实本身。
如果一份材料在某个时间已经存在，那么这件事情应该能够被后来的人检查。
如果未来出现某种叙事，声称一件事情从未发生，或者把一个复杂的共同发展过程重新包装成某一个主体的独占成就，那么社会至少应该仍然有机会回到历史材料。
基础设施的任务不必是宣布：
“这个思想属于谁。”
它只需要让后来形成的社会事实，仍然能够受到历史事实的约束。
这一区别非常重要。
它意味着记录不是为了占有。
保存也不是为了控制。
恰恰相反，保存生成历史，可以让知识更容易进入公共讨论，因为我们不必把信任压缩成对某个作者、机构或品牌的单一信赖。
人们可以看见轨迹。
可以自己判断。
可以重新解释。
甚至可以在未来的知识条件下重新评价过去曾经被认为正确的东西。
从这个意义上说，生成轨迹也可能成为一种公共资源。
它不仅服务于作者。
它服务于后来的人。
服务于研究共同体。
服务于知识的再利用。
也服务于未来对过去进行重新认识的可能性。
这让我想到范内瓦尔·布什在《As We May Think》中关于“trail”的想象。
他设想未来的人类知识系统不只是储存文献，还能够保存人在文献之间走过的关联路径。后来我们拥有了超文本，也拥有了规模远超当时想象的信息网络。
但今天，也许我们需要重新提出这个问题。
未来需要保存的可能不仅是：
“从哪一份文献跳到了哪一份文献。”
还包括：
一个思想曾经怎样走到这里。
这是比引用网络更丰富的一种轨迹。
它包含关系。
包含时间。
包含修改。
包含判断。
也包含尚未解决的问题。
人工智能让我们第一次如此清楚地看到这一需求，并不是因为过去的知识没有历史，而是因为过去我们还能把许多历史信号粗略地寄托在最终成果之中。
如今，结果与过程可以被迅速分离。
一旦这种分离成为常态，我们就需要新的制度去重新连接它们。
我并不知道这种基础设施最终应该长什么样。
它可能由开放仓库发展而来。
可能与版本管理结合。
可能需要新的出版制度。
可能需要机器可读的 provenance。
也可能需要一种尚未形成的知识记录伦理。
它甚至可能不是一个单一平台，而是一组相互连接的制度与协议。
但有一点已经越来越清楚：
未来的知识基础设施不能只关心我们生产了什么。
它还需要给我们留下理解“它是怎样形成的”的可能。
因为当结果越来越丰富，历史就不应该越来越贫乏。
当生成越来越容易，判断就会越来越重要。
而当机器与人共同进入知识生产，我们尤其需要知道的，不只是最终留下了哪些句子，而是那些句子之前发生过什么。
结果之外，还有生成的历史。
那段历史，并不是作品之外无关紧要的废料。
有时，它正是我们理解作品、建立信任、认识价值，并让新的知识继续生成的条件。
所以，我想把这封信留给所有参与创造、传播、阅读、使用和保存知识的人。
无论我们是研究者、学生、教师、工程师、编辑、开发者、艺术家，还是只是偶然提出一个问题的人，我们都正在进入同一个新的知识环境。
在这个环境中，我们会拥有越来越强大的生成能力。
而我们需要共同学习的，也许是另一件事：
怎样让生成留下历史。
谨致敬意与希望，
Wanhong Huang