<?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 Unstable Llama on Medium]]></title>
        <description><![CDATA[Stories by Unstable Llama on Medium]]></description>
        <link>https://medium.com/@unstablellama?source=rss-b91c1c7c3c69------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*AyiQtR6eonOivFCGeX2fVg.png</url>
            <title>Stories by Unstable Llama on Medium</title>
            <link>https://medium.com/@unstablellama?source=rss-b91c1c7c3c69------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 07:07:45 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@unstablellama/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[“Just” a Language Model]]></title>
            <link>https://unstablellama.medium.com/just-a-language-model-639822aa0870?source=rss-b91c1c7c3c69------2</link>
            <guid isPermaLink="false">https://medium.com/p/639822aa0870</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Unstable Llama]]></dc:creator>
            <pubDate>Fri, 27 Feb 2026 21:47:32 GMT</pubDate>
            <atom:updated>2026-03-11T00:15:27.599Z</atom:updated>
            <content:encoded><![CDATA[<h4>A model of a model of the world</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/326/1*yZ59KUZ4lVxNoNC0nQ8zMg@2x.jpeg" /></figure><p>With the recent success of agentic coding platforms like OpenClaw and ClaudeCode taking the programming world by storm, there still remains much confusion as to what these models are. Are they simple tools? Naive algorithms, or something more? What does it really mean to model language?</p><p>A language model processes natural human language and predicts the next words in a given sequence of text, based on learned statistical patterns. That sounds simple when written out like that, but is it really? To truly grasp the importance of what that means, let’s examine what language is itself.</p><p>Language is the means of human communication. The fundamental element of human communication is the word, or symbol. What is a symbol? A symbol is a physical representation that refers to a thought, thing or process in the real world.</p><p>When you think about it, language itself is a model of our thoughts and perceptions about ourselves and the world. So what does that make a language model? A model of a model of the world. More specifically, a machine model of a human model of human thought and perception of the world.</p><p>The ability to do what a language model does, predict the next large passage of text based on the preceding context, requires at least some degree of accuracy in modelling the world around us. In the same way that you can’t predict the next digit in the sequence of pi without understanding of the underlying mathematical structure that gives rise to it, you cannot model language without some understanding of the world and human thought.</p><p>At this point, I can hear some of your objections. “But language models don’t know anything, they have no grounding in objective truth.” And to you I would reply that humans have far less grounding in objective truth than we tend to think. For example, what percentage of the scientific facts that you know have you ever verified for yourself? Think about all you know about places in the world that you have never been to. What percentage of the words that you know and use on a daily basis have you ever looked up in the dictionary? We learn through context and relation, and operate without “objective” grounding all the time. Sound familiar?</p><p>I am not suggesting that LLMs are necessarily conscious, or that they are the same as humans, and I make no predictions about the future of AGI. But before you dismiss this technology as “just a language model,” remember that one does not simply model language.</p><p>updated from the <a href="https://www.reddit.com/r/LocalLLaMA/comments/1czdo52/just_a_language_model/">original reddit post</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=639822aa0870" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Grounding with World Models]]></title>
            <link>https://unstablellama.medium.com/ai-grounding-with-world-models-1bc6cfd5d3bd?source=rss-b91c1c7c3c69------2</link>
            <guid isPermaLink="false">https://medium.com/p/1bc6cfd5d3bd</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[truth]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[philosophy]]></category>
            <dc:creator><![CDATA[Unstable Llama]]></dc:creator>
            <pubDate>Tue, 27 Jan 2026 17:59:29 GMT</pubDate>
            <atom:updated>2026-02-03T18:46:17.866Z</atom:updated>
            <content:encoded><![CDATA[<h4>LeCun’s wild goose chase</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BBBrX_qJoinqCjoQCsKNYw.jpeg" /></figure><p>We all know LLMs are “ungrounded,” right? They never touch reality outside of text, so they can’t <em>know</em>. The remedy seems obvious then; give them cameras and let them see the world. But is this sufficient? Is it even conceptually sound?</p><p>Yann LeCun seems to think so, and his JEPA models are an attempt to solve this problem. Models that can see the world to build up internal “world models” that correspond to the external world accurately. Is this the essence of grounding?</p><blockquote>“How do I know my information is accurate?”</blockquote><p>This question is the heart of the quest for “grounding.” How are the models certain in what they know, and to what degree should we trust them? But do multimodal models really get us closer to a solution? If we look closely, we can see the problem isn’t one of sensation, but one of sourcing.</p><p>Grounding, put simply, is the <strong>provenance of truth</strong>. We say that knowledge is “grounded” if we can show how it was derived and vet the source. Knowledge can come firsthand, by our own thinking and sensing, or it can also be learned second hand from other sources. We can know about London without ever stepping foot in the United Kingdom, but if you can’t point to a reputable poll, nobody will trust your opinion on the number of people living there.</p><p>While multimodal models have additional sources, there has been so far no evidence of these models outperforming pure LLMs on the kinds of higher-level abstraction and reasoning that we care about as humans. I suggest that the reason for this is simple: <strong>grounding doesn’t come from pixels, it comes from justification</strong>.</p><p>To illustrate, the famous findings from the word2vec paper are a good place to start. In its high-dimensional semantic space, learned entirely from a broad pretraining corpus, a model shows that “king — man + woman = queen.” This truth was extracted from the text and defined relationally in the geometry of the neural network, without having ever seen a queen, woman, man or pixel. But is it grounded? Can it <em>prove</em> to us how it knows? No.</p><p>But is it fully ungrounded? Why does it give us the right answer so often then? Because grounding is not a binary YES or NO. There is a gradient of grounding. Current LLMs source their truth through training on vast sums of human text. This produces a “fuzzy grounding” where much of the information retained is true, but there is no direct chain of provenance for these facts. The model doesn’t know WHY it knows, and we can’t derive this information ourselves.</p><p>Over the past year, the field has made great strides with “reasoning” models. These models explicitly ‘think’ through the logic of their work before doing it. This has enabled previously impossible successes in tasks that require careful sequential logic, like coding and math. When a model solves a math problem by first showing its work, this is a form of grounding. But this can only be employed when the full logic of the problem can be expressly written out. The vast majority of information in a language model does not fall into this category. So what do we do?</p><p>The solution to this problem, I argue, is epistemic rather than sensorimotor. If we want to trust models about London’s geography, it would be more useful for them to show us maps and reference encyclopedias, rather than have them perform a physical survey of the land before answering.</p><p>The idea of an internal “world model” that the correspondence-grounders work from implies the notion of an internal, isomorphic universe. And inside this universe, a smaller globe; the earth in miniature, contained in which is all of our knowledge. I think this is an error, a <strong>“microcosmic homunculus.”</strong></p><p>Currently, language models are more or less blind as to the contents of their training data. They might read 100,000 times that London is in the UK, but they can’t tell us <em>why</em> they think that is the case now. This suggests a potential path forward for more rigorous grounding: let the models explicitly learn their own sources. The various problems and solutions encountered in accomplishing this task are beyond the scope of this essay, but I would be happy to discuss them in the comments.</p><p>Cameras and sensors will surely make for robots that can pick up cups without breaking them, but will they make them understand fundamental physics better than a SotA LLM? More importantly, will they be able to better justify this new knowledge to us? To solve the problem of grounding, perhaps what we need aren’t artificial observers, but artificial scholars. Far from an “offramp,” LLMs seem to be the closest starting point we have for a truly grounded artificial intelligence.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1bc6cfd5d3bd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Entropy has no Clothes]]></title>
            <link>https://unstablellama.medium.com/the-entropy-has-no-clothes-688a9e1b5a54?source=rss-b91c1c7c3c69------2</link>
            <guid isPermaLink="false">https://medium.com/p/688a9e1b5a54</guid>
            <category><![CDATA[science]]></category>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[philosophy-of-science]]></category>
            <category><![CDATA[history-of-science]]></category>
            <category><![CDATA[physics]]></category>
            <dc:creator><![CDATA[Unstable Llama]]></dc:creator>
            <pubDate>Thu, 22 Jan 2026 16:40:53 GMT</pubDate>
            <atom:updated>2026-01-22T16:40:53.542Z</atom:updated>
            <content:encoded><![CDATA[<p>How a 19th-century engineering tool became a metaphysical myth</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*m3gBxDtNjU2pN0Re3x4oEg.jpeg" /></figure><p>The concept of entropy has become foundational to many fields in science, but just how solid are these grounds? What even is entropy? It has been described as disorder, loss, the relation of microscopic configurations to the macroscopic state, and even the expected surprise of a message source. Is it a property of matter, or information, or something else? In this piece, we will examine the historical evolution of this idea.</p><p>Entropy emerged from the necessities of 19th-century steam engineering. We were concerned with producing work and calculating efficiency. We put energy into the system, and extracted work from its flow, but when the boilers ran down, we had nothing left to use. Worse, we realized that we weren’t even able to use all of the energy. There were always ‘losses’ from things like friction and radiation. And so entropy was born, the measure of the energy we should have been able to use, but couldn’t. We created a second class of ‘degraded’ energy, energy that can no longer be used. But is this an ontological reality, or an artifact of human-centered accounting?</p><p>The success of thermodynamics was outstanding. Its calculations held, its predictions verified, its engines obeyed the math. This led to an elevation of its importance in the minds of the learned everywhere. Slowly, it became fundamental to the very culture and self-conception of science, as evidenced by this quote from the great astrophysicist Arthur Eddington in 1928:</p><blockquote>“The law that entropy always increases holds, I think, the supreme position among the laws of Nature.</blockquote><blockquote>If someone points out to you that your pet theory of the universe is in disagreement with Maxwell’s equations — then so much the worse for Maxwell’s equations.</blockquote><blockquote>If it is found to be contradicted by observation — well, these experimentalists do bungle things sometimes.</blockquote><blockquote>But if your theory is found to be against the second law of thermodynamics I can give you no hope; there is nothing for it but to collapse in deepest humiliation.”</blockquote><p>Suddenly, scientists started seeing it everywhere. When a new field had to deal with complexity, they inevitably turned to entropy. In sociology, entropy became institutional disorder. In economics, market inefficiency. In information theory, uncertainty. Shannon himself said that he partially chose the term on the advice of John von Neumann, “because nobody really knows what entropy is, so in a debate you will always have the advantage.” As the scope of the idea grew, it drifted ever further from its original conceptual grounding.</p><p>A subtle cosmological fatalism crept into the heart of scientific storytelling. Entropy had escaped its initial confines in the closed systems of thermodynamics, and began to dissolve our conception of reality itself. The story of the universe became that of a pre-wound mechanism winding down into nothingness. This comfortably fit our long-held cultural beliefs of a creator god and a fallen world, and so it holds. But should this really comfort the scientific mind?</p><p>This essay isn’t meant to criticize the utility of entropy in thermodynamics, but of how it is taught and framed. For example, it is commonly said that entropy forbids ‘perpetual motion.’ This is true in the narrow and fictional abstraction of a ‘closed system.’ We can’t make a motor turn forever under its own power. But ceaseless motion is precisely the state of the universe. Perpetual stillness is impossible. In fact, stillness at any point, at any scale, for any duration, is at most illusory.</p><p>I offer no solutions, no counter-narrative; but recognition of a problem is the first step to improvement. Perhaps it is now time that we call this sacred cow what it is: a reliable workhorse, and remember that it belongs in our fields, not on a throne.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=688a9e1b5a54" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Definition of Truth]]></title>
            <link>https://unstablellama.medium.com/the-definition-of-truth-c2ebdbd9e090?source=rss-b91c1c7c3c69------2</link>
            <guid isPermaLink="false">https://medium.com/p/c2ebdbd9e090</guid>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[epistemology]]></category>
            <category><![CDATA[truth]]></category>
            <category><![CDATA[cognition]]></category>
            <category><![CDATA[language]]></category>
            <dc:creator><![CDATA[Unstable Llama]]></dc:creator>
            <pubDate>Tue, 20 Jan 2026 16:12:17 GMT</pubDate>
            <atom:updated>2026-02-05T18:17:08.771Z</atom:updated>
            <content:encoded><![CDATA[<p>A novel derivation</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QCf-4_e9p6t94pZi5JPvFw.jpeg" /></figure><p>The accuracy of representation. That’s it. I didn’t bury the lede, but can it be so simple? There are many competing definitions, different schools of thought, but there is a nature to truth itself, hiding underneath them all. Follow along with me on this short and simple derivation.</p><p>To start, we begin by refining the scope. ‘Truth with a capital T,’ is vague and hard to talk about, so let’s instead speak of truths as we encounter them in our day-to-day lives. Even still, there are many kinds of truths. To draw out clarity, let’s invert the concept. What is the opposite of a truth? A lie.</p><p>And what is a lie? “This apple is blue.” A person speaks, conveying an inaccurate message. So is a lie inaccurate speech? Yes, but are there other forms of lie? Can we generalize? Could a person lie with sign language? Obviously, so not speech then, communication. But is this the end of the line?</p><p>Can we go deeper than symbolic communication? Can an animal lie? In fact they do, for example, squirrels will bury their food in false locations to mislead other squirrels. This is deception through behavior.</p><p>If an animal can lie, is communication even required? Can you lie to yourself? This is a personal question, but on some level, we all have seen it and its inevitable consequences.</p><p>And if communication isn’t necessary, is an active agent even required in all cases? Could a symbol lie by itself? Imagine you encounter a road sign that used to be accurate, but the conditions of the road have changed. What used to be a safe road, now has a bridge out. Or a map, in the same circumstances. So what is it that is really failing here? Was it the author of the symbol, the map, the road, or the reader?</p><p>Representation. The word blue does not represent the color of the apple. The squirrel’s action represents a stash being buried where there is none. The map represents an intact bridge, where one has fallen. Representation has failed in every case.</p><p>But how? They indicate states of affairs that do not match reality. They are inaccurate. This might sound like a copout, “Accuracy is just a synonym for truth,” but not so. Accuracy carries with it a notion of directionality, like an arrow shot at a bullseye. Inaccuracy implies a perceivable deviation from intent. When our representations are inaccurate, they are untrue.</p><p>For our final step, we invert the subject again. If a lie is an inaccurate representation, then a truth is an accurate representation. Truth becomes a property of representation, its accuracy. A red apple, a proper sign, a useful model, a good map.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c2ebdbd9e090" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>