<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Daniel on Medium]]></title>
        <description><![CDATA[Stories by Daniel on Medium]]></description>
        <link>https://medium.com/@daniel6651?source=rss-3846c41c646d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*_RJVlF8FmZrcBXoH</url>
            <title>Stories by Daniel on Medium</title>
            <link>https://medium.com/@daniel6651?source=rss-3846c41c646d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 16:59:09 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@daniel6651/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The Suffering of Specification]]></title>
            <link>https://medium.com/@daniel6651/the-suffering-of-specification-00f605c6d9a1?source=rss-3846c41c646d------2</link>
            <guid isPermaLink="false">https://medium.com/p/00f605c6d9a1</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Daniel]]></dc:creator>
            <pubDate>Sun, 20 Sep 2026 22:23:33 GMT</pubDate>
            <atom:updated>2026-09-20T22:23:33.044Z</atom:updated>
            <content:encoded><![CDATA[<p>Like many others, I tried Karpathy’s LLM wiki.</p><p>I implemented it several times. And I had to throw them all away.</p><p>It was too easy to use.</p><p>Where was the burden? Where was the suffocating procedure? Where was the suffering?</p><blockquote><em>All conditioned things are impermanent.</em></blockquote><blockquote><em>All code is conditioned.</em></blockquote><blockquote><em>All code is impermanent.</em></blockquote><blockquote><em>I code, therefore I suffer.</em></blockquote><blockquote><em>I suffer, therefore I am.</em></blockquote><p>I realized fairly quickly that the concept needed more layers of suffering.</p><p>The obvious next step is, of course, to add a taxonomy. So I did.</p><p>But then it struck me: I could not possibly do this the easy way. What would be the point?</p><p>As one does in these situations, I turned to the Bhagavad Gita for guidance. Reading the line “The mind is restless and difficult to restrain, but it is subdued by practice,” I realized what was missing.</p><p>I needed a formal specification.</p><p>But not just a simple PRD. That would be far too easy.</p><p>The suffering needed to manifest itself in the specification itself. A high-level goal became structure, structure became increasingly detailed specifications, and eventually implementation plans. Naturally, all of this had to be grounded in a thorough study of the field.</p><p>This seemed appropriate.</p><p>But then another problem became apparent.</p><p>The formal specification naturally called for a formal method of reasoning.</p><p>What could actually be known? What should be doubted? If there was doubt, the doubt itself had to become part of the process. Then the language itself had to be doubted.</p><p>A requirement could not simply be a requirement; it had to be distinguished from a goal, constraint, claim, concept, heuristic, workflow, or type.</p><p>We now had a formal system of prompts.</p><p>We were no longer specifying the system. We were specifying the reasoning by which the system was to be specified.</p><p>Which, naturally, required specifying the reasoning by which that reasoning was to be specified.</p><p>Which was, of course, the goal.</p><p>The transcendent was now obvious.</p><p>I started iterating.</p><p>I threw different AI agents at each other. Agents reviewed agents. Agents questioned agents.</p><p>Agents questioned their own reasoning.</p><p>I pushed it beyond sanity and made it produce layer after layer of specification.</p><p>Eventually, the AI started to complain.</p><p>“This is enough. Let’s start coding.”</p><p>I knew then that I was getting somewhere.</p><p>The taxonomy grew quickly. But I soon realized that taxonomy was merely the layman’s approach.</p><p>The aim was suffering, not chaos.</p><p>A taxonomy alone is like a feather out of a bird: technically related, but hardly sufficient to describe the bird.</p><p>What emerged was no longer merely a taxonomy. It was an epistemology, with ontology, facets, domains, and types.</p><p>Then Sophocles entered the picture. Or perhaps it was someone else.</p><p>“Trust dies, but mistrust blossoms.”</p><p>The words arose in my mind, and they made the next step obvious: we needed a stringent system for controlling what was allowed into the wiki.</p><p>More suffering.</p><p>An ingestion pipeline with sanitization, classification, promotion, and review checkpoints followed.</p><p>The same principle obviously applied to the underlying remote filesystem. Surely the storage layer deserved its own suffering as well.</p><p>Crash-safe source capture.</p><p>Transactional journaling.</p><p>Recovery.</p><p>Consistency.</p><p>The usual things.</p><p>Then, one night, while reading Tractatus Logico-Philosophicus, a sentence caught my attention:</p><p>“Language disguises thought.”</p><p>What could that mean for our wiki?</p><p>I had to read the book again.</p><p>The one-dimensional representation of multidimensional ideas could not possibly do them justice.</p><p>The answer was obvious.</p><p>A graph.</p><p>Graph representation. Graph search. Graph intelligence.</p><p>Another layer of suffering emerged.</p><p>Any knowledge produced by man or machine raises another question: what happens when the knowledge disagrees?</p><p>Dialectical tension seemed inevitable.</p><p>Fortunately, Bohr had already provided the answer.</p><p>“The opposite of a correct statement is a false statement. But the opposite of a profound truth may well be another profound truth.”</p><p>Clearly, the wiki would have to accommodate contradictory truths. Alternative facts, perhaps.</p><p>Some research into the field gave little more than a glimpse of what lay ahead.</p><p>By this point, the specification had become a system of its own.</p><p>So, after driving the AI to the brink of insanity with overspecification, it was finally time to implement.</p><p>It was becoming difficult to tell where the specification ended and the system began.</p><p>Mutatis mutandis.</p><p>At some point I began to suspect that the specification was no longer describing the system.</p><p>The specification was the system.</p><p>Eventually, however, something strange happened.</p><p>It worked.</p><p>The current version is functional.</p><p>Does it add any significant value over the plain LLM wiki?</p><p>Probably not.</p><p>But that was never really the point.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=00f605c6d9a1" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>