<?xml version="1.0" encoding="utf-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Tangent Space</title>
    <link>https://blog.viva-la-vida.me/</link>
    <description>Changxu&apos;s Words</description>
    <language>en</language>
    <lastBuildDate>Wed, 18 Nov 2026 09:03:22 +0000</lastBuildDate>
    <atom:link href="https://blog.viva-la-vida.me/rss.xml" rel="self" type="application/rss+xml" />
    
    <item>
      <title>(Don&apos;t) Measure Every Rock</title>
      <link>https://blog.viva-la-vida.me/2026/11/15/don-t-measure-every-rock</link>
      <guid isPermaLink="true">https://blog.viva-la-vida.me/2026/11/15/don-t-measure-every-rock</guid>
      <pubDate>Sun, 15 Nov 2026 00:00:00 +0000</pubDate>
      <description>&lt;p&gt;Recently I watched a &lt;a href=&quot;https://www.bilibili.com/video/BV1MbFXz5Esa?vd_source=60849680a6b7258a13c47771728b9e1f&quot;&gt;video&lt;/a&gt; talking about the &lt;a href=&quot;https://en.wikipedia.org/wiki/Coastline_paradox&quot;&gt;Coastline Paradox&lt;/a&gt;. First noticed by Lewis Fry Richardson in 1951, the problem asks why a coastline does not have a single scale-independent length. The measured perimeter depends on the method and scale that you do the measurement. Since a landmass has features at all scales, the smaller scale you use for measurement, the longer the measured coastline becomes.&lt;/p&gt;

&lt;p&gt;The observation later helped motivate Benoit Mandelbrot’s formalization of fractal geometry. The property that matters here is simpler than the full theory: the finer the ruler you measure a coastline with, the more boundary you find. In the mathematical idealization, the measured length can grow without bound, even though the area it encloses stays finite.&lt;/p&gt;

&lt;p&gt;This reminds me of a heuristic I have learned from productionizing software systems: for an established system, problems at different scales are often different problems. Demanding more is like measuring with a finer ruler: more boundary appears, and that new boundary is new terrain, not the old terrain repeated.&lt;/p&gt;

&lt;p&gt;Let’s take building a distributed system as an example. Building the basic feature in a real production system is already a serious project. You need to deal with authentication, persistence, schema, deployment, monitoring, and unglamorous plumbing that exists before the feature does anything interesting. And once it exists, every one of these is a harder version of that same system:&lt;/p&gt;

&lt;ol&gt;
  &lt;li&gt;Reliability: 99.95 -&amp;gt; 99.99% -&amp;gt; 99.999%.&lt;/li&gt;
  &lt;li&gt;Latency (p99): 500ms -&amp;gt; 200ms -&amp;gt; 100ms&lt;/li&gt;
  &lt;li&gt;Scope: one region -&amp;gt; many regions -&amp;gt; global&lt;/li&gt;
  &lt;li&gt;Conditions: ideal inputs-&amp;gt; spikes, retries, backfills -&amp;gt; slow, duplicated, or unavailable dependencies
…&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Just as a coastline exposes more details when measured at a smaller scale, a production system exposes new failure modes, bottlenecks, and coordination challenges as its reliability, latency, traffic, and organizational expectations become more demanding (there will be another essay that talks about why there will never be a perfect project planning and you will have to deal with these problems after delivering the MVP).&lt;/p&gt;

&lt;p&gt;But the Coastline Paradox also has another useful lesson. Even if the measured perimeter of an island may increase indefinitely as the measurement scale gets smaller, the total area of the island can still have an upper bound. That feels aligned with our real-world intuition: the boundary may be infinitely detailed, but the land itself is still finite.&lt;/p&gt;

&lt;p&gt;For software projects, a naive analogy is that you can always invent more requirements, each of which may require substantial engineering effort. You can almost always find a harder version of the same problem: more reliable, more scalable, lower latency, more fault tolerant, more globally available. From a technical point of view, these problems can be complex, challenging, and rewarding to solve. In that sense, the problem space can feel fractal&lt;sup id=&quot;fnref:1&quot;&gt;&lt;a href=&quot;#fn:1&quot; class=&quot;footnote&quot; rel=&quot;footnote&quot; role=&quot;doc-noteref&quot;&gt;1&lt;/a&gt;&lt;/sup&gt;: each improvement reveals more structure at the next scale.&lt;/p&gt;

&lt;p&gt;But the value of the system is not measured by the length of that boundary. It is measured by the area it encloses: the customer value, product impact, and real user needs the system actually serves. Notice the asymmetry the coastline hides in plain sight. As you measure finer, the perimeter grows without bound (there is always another rock), but the enclosed area converges on a fixed, finite number. The boundary never reaches a point that says stop; the area does. This, I think, is why capable engineers over-invest: not from poor judgment, but because they are taking their stopping signal from the boundary, the one side of the system that has no natural end. If your stopping question is “is there more boundary I could harden?”, the answer is permanently yes. The discipline is to take the stopping condition from the area instead, not “is there more I could do”, but “is there more real value left to enclose?”, and that question usually has a finite answer.&lt;/p&gt;

&lt;p&gt;Another observation is that the same project can look very different depending on who is observing it, what role they play, and what measurement scale they care about. Some people look at the area of the island. Others are more interested in counting every rock along the coastline. There is nothing inherently wrong with either perspective. They are simply using different maps for different purposes. A real estate investor may care about the total area, zoning potential, and development value of the island. An ecologist may care about the shoreline, tide pools, erosion patterns, and small habitats that only appear at a much finer scale. Someone whose home is built right on the coast may care about every curve of the boundary, because that is where their daily life happens.&lt;/p&gt;

&lt;p&gt;The same is true in software. A leader may look at whether the system creates enough customer value to justify the investment. An infrastructure engineer may look at the edge cases, failure modes, and reliability gaps exposed at the boundary. A customer support team may care about the handful of users who hit unusual states. A security engineer may care about the one small crack that others would ignore. These are not necessarily conflicting views. They are different measurement scales applied to the same system.&lt;/p&gt;

&lt;p&gt;In that sense, choosing a measurement scale is also a form of alignment. Different ways of observing the same goal would result in different tasks, so collaborators need to agree not only on the goal, but also on the granularity at which the problem should be examined. Are we trying to estimate the island’s area, or are we trying to trace the coastline rock by rock? Both can be valid tasks, but they require different levels of attention and lead to different definitions of “done”.&lt;/p&gt;

&lt;p&gt;This becomes even more important when working with AI. Before handing over execution, it is useful to align on the measurement scale: the level of detail, the acceptable approximation, the expected rigor, and the amount of reasoning the task deserves. Problems at different scales may all require serious attention, even when they look similar from a distance. A model, left to itself, does what the engineer does: it takes its stopping signal from the boundary, because the boundary always offers more to examine. An unbounded thinking budget on such a problem will always find another rock, because nothing in the task tells it where the area ends. The fix is to hand the model the area: name the decision the output feeds, so it measures toward that and stops when the value is enclosed, rather than tracing a boundary it was never asked to walk&lt;sup id=&quot;fnref:2&quot;&gt;&lt;a href=&quot;#fn:2&quot; class=&quot;footnote&quot; rel=&quot;footnote&quot; role=&quot;doc-noteref&quot;&gt;2&lt;/a&gt;&lt;/sup&gt;.&lt;/p&gt;

&lt;div class=&quot;footnotes&quot; role=&quot;doc-endnotes&quot;&gt;
  &lt;ol&gt;
    &lt;li id=&quot;fn:1&quot;&gt;
      &lt;p&gt;The full fractal picture includes self-similarity, that the same structure recurring at every scale. I’m not relying on that property here. In a production system, finer scale tend to surface genuinely new problems rather than smaller copies of the old ones. What this essay borrows from the coastline is narrower, that finer measurement surfaces more boundary, while the enclosed area stays bounded. &lt;a href=&quot;#fnref:1&quot; class=&quot;reversefootnote&quot; role=&quot;doc-backlink&quot;&gt;&amp;#8617;&lt;/a&gt;&lt;/p&gt;
    &lt;/li&gt;
    &lt;li id=&quot;fn:2&quot;&gt;
      &lt;p&gt;The hard pill to swallow is that, without real human experience in a domain, you often do not know what scale of problem you are looking at. When you enter a new field, it is hard to tell whether the issue in front of you is a tiny rock on the coastline, a meaningful change in the shoreline, or a feature that changes the shape of the entire island. You also may not know what level of detail is appropriate, which details are essential, which are nice to have, and which only create noise. This is one reason AI collaboration still (and may always) require human judgement (more in this &lt;a href=&quot;/2026/07/12/leaving-the-room&quot;&gt;essay&lt;/a&gt;). Someone needs to decide what kind of map is useful for the task. Sometimes the most important work is not asking AI to think harder, but making sure someone understands what scale of thinking the problem actually deserves. Someone has to know which map matters. &lt;a href=&quot;#fnref:2&quot; class=&quot;reversefootnote&quot; role=&quot;doc-backlink&quot;&gt;&amp;#8617;&lt;/a&gt;&lt;/p&gt;
    &lt;/li&gt;
  &lt;/ol&gt;
&lt;/div&gt;
</description>
      
      <category>Thinking</category>
      
      <category>AI</category>
      
    </item>
    
    <item>
      <title>Leaving the Room</title>
      <link>https://blog.viva-la-vida.me/2026/07/12/leaving-the-room</link>
      <guid isPermaLink="true">https://blog.viva-la-vida.me/2026/07/12/leaving-the-room</guid>
      <pubDate>Sun, 12 Jul 2026 00:00:00 +0000</pubDate>
      <description>&lt;p&gt;AI makes the generation of ideas cheap. What it does not make cheap (directly) is verifying the ideas against reality. That asymmetry is what this essay is about.&lt;/p&gt;

&lt;p&gt;There are two basic types of “reality”:&lt;/p&gt;

&lt;ul&gt;
  &lt;li&gt;Physical world, as the objective reality (or what Searle called the “Brute Fact”).&lt;/li&gt;
  &lt;li&gt;Human consensus (or what Searle called the “Institutional Fact”), is the collective recognition (social rules) that are accepted by a group of people&lt;sup id=&quot;fnref:1&quot;&gt;&lt;a href=&quot;#fn:1&quot; class=&quot;footnote&quot; rel=&quot;footnote&quot; role=&quot;doc-noteref&quot;&gt;1&lt;/a&gt;&lt;/sup&gt;&lt;sup id=&quot;fnref:2&quot;&gt;&lt;a href=&quot;#fn:2&quot; class=&quot;footnote&quot; rel=&quot;footnote&quot; role=&quot;doc-noteref&quot;&gt;2&lt;/a&gt;&lt;/sup&gt;. It gives an object, action, or person a particular meaning or status.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Language, or the description of reality, is a projection of a human’s understanding of reality. The AI model can capture the structure of the language perfectly as long as there is enough data. However, printing letters that can follow the same structure of human language does not mean it can affect reality as expected. If you, like me, believe that some part of the great goal is to transform reality, then bear with me: the usefulness of whatever machine generates, depends on the cost and speed of the reality verification. Coding is an area where verification is the cheapest. For bounded programming tasks with executable tests, many claims can be checked automatically within seconds. However, even with the test verification only verifies what its author thought to assert, as a cache of anticipated failures based on the author’s understanding of the reality. The true verification comes with running the code in production to battle against both types of realities (for example, how the code performs on real hardware, and how users feel about the product). Mathematics shows a partial version of the same transition: once a statement and proof have been formalized, a proof assistant can check them mechanically, though the cost of formalization remains high. Beyond coding and Math, a legal opinion, a medical diagnosis, a business strategy, their verification all require the gradual expansion of reality. It takes a couple of finance quarters to verify the business strategy, and it takes a human body’s course of disease to verify the medical diagnosis. The destination of the verification is always the reality by itself, and the key separation between language and reality, determines that the verification cannot be fully done within language itself.&lt;/p&gt;

&lt;p&gt;And among all the ways verification exits language, the hardest case, the one worth examining, is when the only available verifier is human first-hand experience.&lt;/p&gt;

&lt;p&gt;In a famous 1982 thought experiment, philosopher Frank Jackson imagined Mary, a scientist who knows every physical fact about color (wavelengths, optics, the neuroscience of vision) but has lived her whole life in a black-and-white room. The question: when she finally steps outside and sees red, does she learn something new?&lt;/p&gt;

&lt;p&gt;Philosophers have argued about the answer for forty years. But set the debate aside and notice what happens if we recast the story: the language model is Mary, a Mary whose room has no door. It has read every description of red ever written: it can recite the wavelengths, the associations, the metaphors. And it will never step outside, because its room is the language itself. We, meanwhile, are all people who have seen red. Framed this way, the machine’s fluency about color is neither new nor frightening. The interesting moment is when it gets one wrong: when it confidently misnames a color it has never seen. Who catches the error? Only someone who has been outside the room. This is quite the inversion the thought experiment hides, that the value of first-hand experience is not merely that it produces knowledge but that in some domains it supplies the final standard against which claims are checked. The experience is the verifier of last resort. Most physical reality verification is, in fact, being instrumented. Blood panels check diagnoses, strain gauges check bridges. So the claim is not that human experience monopolizes verification. The claim is narrower and harder: there is a class of domains where the standard of correctness is human experience: whether the dish is too salty, whether the pain has eased, whether the prose moves you. The problem is not that instruments lack precision; it is that they have nowhere to stand. You could build a delight-detector, but you could only calibrate it against human reports. If the answer bottoms out in a human report, then the human has not been removed from the loop, only moved to its end. When the instrument verifies that something should be classified as red, it is executing a cached human calibration.&lt;/p&gt;

&lt;p&gt;The more interesting case is the areas where verification appears to happen entirely within language. It becomes an even harder problem especially for human consensus. The verification loop for a lot of human consensus is intrinsically what Merton called, a “self-fulfilling prophecy”. A bank is solvent as long as enough people believe it is; the act of verifying ‘this bank is failing’, by withdrawing, is what makes it fail. On March 9, 2023, Silicon Valley Bank faced approximately $42 billion withdrawal requests in a single day, with another $100 billion pending for the next day, a bank run that moved at the speed of group chats&lt;sup id=&quot;fnref:4&quot;&gt;&lt;a href=&quot;#fn:4&quot; class=&quot;footnote&quot; rel=&quot;footnote&quot; role=&quot;doc-noteref&quot;&gt;3&lt;/a&gt;&lt;/sup&gt;. This is where AI gets really dangerous. Physical reality will punish the wrong language. If the model provides a wrong calculation to a bridge design, the bridge will break, and the physical reality would verify what is generated. Human consensus, however, has no immunity against the language machine because its verification is the belief. In reflexive social domains, a widely believed description can change the object until it becomes true. This critical danger existed before the introduction of AI, however, the cost of the generation of the language that can form human consensus, was very high. It used to take printing presses, parties, churches, newspapers, and decades, to form a new consensus. This high cost was a rough proof-of-work: the consensus that survived had survived collision with reality itself. The open question, how a society re-prices its consensus verification when language is free, deserves its own essay.&lt;/p&gt;

&lt;p&gt;Whatever machine generates is only as useful as the verification behind it. It depends on the cost and speed of the reality verification. If that is the bottleneck, then the human work of the coming years is not to compete with machines at generating ideas (&lt;a href=&quot;/2026/11/15/don-t-measure-every-rock&quot;&gt;ideas, like coastlines, can be subdivided endlessly&lt;/a&gt;), it is to drive down the cost of verifying them. There are three ways to do it.&lt;/p&gt;

&lt;p&gt;The first is to design for verifiability. Coding is not cheap to verify by accident; decades of engineering (type systems, unit tests, continuous integration, etc.) deliberately compressed the cost of checking a program from days to seconds. A lean, well-bounded abstraction is exactly this: an artifact whose claims are few and sharp enough for reality to check quickly. Other domains have barely begun this work. Build the correct abstraction (which assumes simplification), align others on it (an abstraction no one shares verifies nothing, and also see the next paragraph), and iterate fast against reality.&lt;/p&gt;

&lt;p&gt;The second is to tend to the consensus. Consensus behaves like a cache of verification: you spend money without re-verifying what money is, sign contracts without re-deriving contract law, because society verified these once and cached the result. A language machine that pollutes consensus is poisoning that cache, and a society whose cache fails falls back to every person verifying everything alone, which no one can afford. The cache does not maintain itself; it is rebuilt, continuously, by people through social processes, and that work cannot be delegated to the very machines it guards against.&lt;/p&gt;

&lt;p&gt;The third is to grow verifiers. If experience is the verifier of last resort&lt;sup id=&quot;fnref:3&quot;&gt;&lt;a href=&quot;#fn:3&quot; class=&quot;footnote&quot; rel=&quot;footnote&quot; role=&quot;doc-noteref&quot;&gt;4&lt;/a&gt;&lt;/sup&gt;, then the point of education shifts from transferring knowledge alone (machines transfer knowledge for almost free now), towards giving people with enough first-hand experience with reality that they can check what machines say. Teach people to leave the room.&lt;/p&gt;

&lt;p&gt;Cheap idea generation was never the goal. It was the bottleneck, and the bottleneck has moved. Verification is ours to build.&lt;/p&gt;

&lt;div class=&quot;footnotes&quot; role=&quot;doc-endnotes&quot;&gt;
  &lt;ol&gt;
    &lt;li id=&quot;fn:1&quot;&gt;
      &lt;p&gt;Because there are different groups of people, there could be different consensus realities in the world. &lt;a href=&quot;#fnref:1&quot; class=&quot;reversefootnote&quot; role=&quot;doc-backlink&quot;&gt;&amp;#8617;&lt;/a&gt;&lt;/p&gt;
    &lt;/li&gt;
    &lt;li id=&quot;fn:2&quot;&gt;
      &lt;p&gt;Even though they rely on human recognition, it does not mean an individual can arbitrarily determine such reality. &lt;a href=&quot;#fnref:2&quot; class=&quot;reversefootnote&quot; role=&quot;doc-backlink&quot;&gt;&amp;#8617;&lt;/a&gt;&lt;/p&gt;
    &lt;/li&gt;
    &lt;li id=&quot;fn:4&quot;&gt;
      &lt;p&gt;To be more clear, in the SVB example, consensus alone does not create the underlying vulnerability, but it acts as an accelerant and trigger within an already vulnerable financial structure. &lt;a href=&quot;#fnref:4&quot; class=&quot;reversefootnote&quot; role=&quot;doc-backlink&quot;&gt;&amp;#8617;&lt;/a&gt;&lt;/p&gt;
    &lt;/li&gt;
    &lt;li id=&quot;fn:3&quot;&gt;
      &lt;p&gt;To be precise about the boundary of this example: machines are not condemned to the room. I am borrowing the architecture of Jackson’s thought experiment, not its argument about consciousness. A camera gives a machine an empirical channel to the world; it does not by itself settle whether the machine experiences color. &lt;a href=&quot;#fnref:3&quot; class=&quot;reversefootnote&quot; role=&quot;doc-backlink&quot;&gt;&amp;#8617;&lt;/a&gt;&lt;/p&gt;
    &lt;/li&gt;
  &lt;/ol&gt;
&lt;/div&gt;
</description>
      
      <category>Thinking</category>
      
      <category>AI</category>
      
    </item>
    
  </channel>
</rss>