<?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 Marcelo Labre on Medium]]></title>
        <description><![CDATA[Stories by Marcelo Labre on Medium]]></description>
        <link>https://medium.com/@marcelolabre?source=rss-dcc48440030a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*plwj4RP_rbOs8sjIgznG5Q.jpeg</url>
            <title>Stories by Marcelo Labre on Medium</title>
            <link>https://medium.com/@marcelolabre?source=rss-dcc48440030a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 18:21:22 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@marcelolabre/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[A Look at FinReflectKG: AI-Driven Knowledge Graph in Finance]]></title>
            <link>https://pub.towardsai.net/a-look-at-finreflectkg-ai-driven-knowledge-graph-in-finance-d588d250948b?source=rss-dcc48440030a------2</link>
            <guid isPermaLink="false">https://medium.com/p/d588d250948b</guid>
            <category><![CDATA[symbolic-reasoning]]></category>
            <category><![CDATA[knowledge-graph]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[Marcelo Labre]]></dc:creator>
            <pubDate>Sun, 28 Sep 2025 22:01:50 GMT</pubDate>
            <atom:updated>2025-09-28T22:01:50.111Z</atom:updated>
            <content:encoded><![CDATA[<h4>The next frontier in truth grounded on symbolic reasoning in finance</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4MDb_1O6TuPbCnoXQK9cuQ.png" /></figure><p>Last week at the <a href="https://luma.com/t11yzbaw?tk=tPdvKu">Quant x AI</a> event here in New York, I had the pleasure of seeing <a href="https://medium.com/u/d22c73975ddb">Fabrizio Dimino</a> present a compelling paper he co-authored: “<a href="https://arxiv.org/abs/2508.17906">FinReflectKG: Agentic Construction and Evaluation of Financial Knowledge Graphs</a>.”</p><p>The paper tackles a critical challenge holding back financial AI. While we have powerful language models, they lack the structured, reliable symbolic reasoning systems needed to truly understand the complex world of finance. More specifically, building knowledge graphs (KGs) on regulatory documents like SEC filings is a massive hurdle.</p><p>The FinReflectKG paper offers a powerful solution with two key contributions: a new, open-source financial KG dataset and a novel framework for building it.</p><p>Here, I’ll summarize their brilliant approach and then propose a way to improving their monitoring for global semantic diversity in the model.</p><h4>The Core Innovation: A “Self-Reflecting” AI Agent</h4><p>The centerpiece of the paper’s methodology is a sophisticated, three-mode pipeline for extracting knowledge “triples” (like (Nvidia, Produces, GPUs)) from financial documents.</p><p>While they test simpler Single-Pass and Multi-Pass methods, their most innovative approach is the Reflection-driven agentic workflow.</p><p>This agentic process works like a team:</p><ul><li>An Extraction LLM first takes a chunk of a document and extracts an initial set of knowledge triples.</li><li>A Critic LLM then reviews these triples, providing structured feedback on any issues it finds, such as ambiguous entities (e.g., using “We” instead of the company ticker) or non-standard relationship types.</li><li>A Correction LLM takes this feedback and refines the triples.</li></ul><p>This feedback loop repeats until no more issues are found or a maximum number of steps is reached, systematically improving the quality of the extracted knowledge.</p><h4>Proving Its Worth Without a Ground Truth</h4><p>A major challenge in KG construction is evaluation, as there’s often no perfect “answer key” to compare against. The authors developed a holistic evaluation framework to address this, using several complementary methods:</p><ul><li>CheckRules: A set of custom, rule-based checks to enforce quality and consistency. For example, rules automatically flag ambiguous subjects like “we” or “the company” and ensure all extracted entities and relationships comply with a predefined schema.</li><li>Coverage Ratios: Metrics to measure how comprehensively the KG captures the diversity of entities and relationships present in the source documents.</li><li>Semantic Diversity: An analysis using information theory (Shannon and Rényi entropy) to measure the balance and variety of the extracted knowledge, ensuring the graph isn’t overly skewed towards a few common concepts.</li><li>LLM-as-a-Judge: A comparative evaluation where a powerful LLM assesses the outputs of the three different extraction modes (single-pass, multi-pass, and reflection) across four key dimensions: Precision, Faithfulness, Comprehensiveness, and Relevance.</li></ul><h4>Key Findings and Verdict</h4><p>The results clearly demonstrate the superiority of the reflection-agent mode. It consistently achieves the best balance of reliability and coverage.</p><ul><li>It achieved the highest compliance score, passing 64.8% of all four strict CheckRules.</li><li>It extracted the most triples per document chunk (15.8) and had significantly higher entity and relationship coverage ratios than the other methods.</li><li>In the LLM-as-a-Judge evaluation, it was the clear winner in Precision, Comprehensiveness, and Relevance.</li></ul><p>The primary trade-off is speed. The iterative feedback loop requires more computation, making it less suitable for real-time applications where a single-pass approach might be preferred.</p><h4>Future Directions</h4><p>The authors conclude by outlining their plans to significantly expand the project, including:</p><ul><li>Enlarging the dataset to cover all S&amp;P 500 companies over the last 10 years.</li><li>Developing a schema-free pipeline that can create ontologies from scratch, inspired by the “Extract-Define-Canonicalize” (EDC) framework.</li><li>Building Temporal Knowledge Graphs to capture the evolution of financial relationships over time, enabling causal reasoning for applications like thematic investing.</li></ul><h4>Suggestion: Monitoring Semantic Refinement via Cross-Entropy</h4><p>The paper notes that the Reflection method, while improving compliance and coverage, reduces the diversity of the extracted elements as measured by absolute entropy. This is an expected outcome of a rule-constrained process. However, the authors’ proposal to monitor “when diversity falls below a predefined threshold” using absolute entropy presents a practical challenge: absolute entropy values are difficult to interpret and make thresholding arbitrary.</p><p>A more principled approach is to measure the relative change in the information landscape between the baseline and the refined output. I suggest a direct application of the Principle of Minimum Cross-Entropy (MinXEnt), an extension of <a href="https://link.springer.com/chapter/10.1007/978-94-009-0683-9_29">Jaynes’ Maximum Entropy Principle (MaxEnt)</a>. In this case, we treat the distribution from the Single-Pass method as the prior and measure how the Reflection method’s distribution diverges from it at each step.</p><p>The ideal tool for this is cross-entropy, specifically the <a href="https://en.wikipedia.org/wiki/Kullback%E2%80%93Leibler_divergence">Kullback–Leibler (KL) Divergence</a>, which quantifies the information gain or loss when one probability distribution is used to approximate another.</p><h4>Proposed Methodology</h4><p>The goal is to monitor the KL Divergence at each iteration <em>t</em> of the Reflection agent&#39;s refinement loop.</p><p><strong>1. Establish the Prior Distribution (<em>p</em>):</strong> First, run the Single Pass method across the corpus. From the full set of extracted elements (entities, types, and relationships), derive a normalized frequency distribution:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/188/1*yAnw9llGJrj0i6guELs1HQ.png" /></figure><p>This represents our baseline or <em>prior</em> understanding of the data’s structure.</p><p><strong>2. Calculate Iterative Distributions (<em>q(t)</em>):</strong> For the Reflection method, at each step <em>t</em> of the iterative feedback loop for a given chunk of text <em>c</em>, derive the corresponding normalized frequency distribution:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/213/1*SpmeZ5TQVATtvSZtr6Secw.png" /></figure><p>where</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/132/1*ro7scJLmY6JtVnqdgSD-aw.png" /></figure><p>and <em>m</em> is the stopping iteration for chunk <em>c</em> as per the stopping criteria in section 4.3.3.</p><p><strong>3. Unify and Smooth:</strong> As the set of extracted elements will differ between the two methods for every <em>t</em>, a direct comparison is not possible. To solve this:</p><ul><li>Create a <strong>unified vocabulary</strong> that is the union of all unique elements found in both <em>p</em> and all <em>q(t)</em>.</li><li>Represent all distributions over this unified vocabulary. Any element not present in a given distribution will have an initial frequency of zero.</li><li>Apply <strong>Laplace (add-one) smoothing</strong> to all frequencies. This is critical to avoid zero probabilities in the denominator of the KL Divergence formula, ensuring the metric is always well-defined.</li></ul><p>For instance, in a 5 element example we may end up with the following frequencies:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/254/1*amG7xDC0qaWBTWfmyIoM9g.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/258/1*8thCcz5VB1cRER2swpR9SA.png" /></figure><p>In this case we have the frequency probabilities:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/492/1*6axTcGTbMAvkCTJPtU22Hw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/396/1*rcSoLvcxmNAtPQNpmY0jiA.png" /></figure><p>by applying Laplace smoothing:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/163/1*hjgmqjDvS__5YYD_ndUvgg.png" /></figure><p>where:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/537/1*o-dpYgd4Xvj6QsHvHBfhCw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/613/1*ULQhTgPNywIRoey6J09euw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/355/1*ggWn2kSb7hyixyjwWlQPzQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/532/1*JrmmaJHeVS4go2rAI_Idmg.png" /></figure><p><strong>4. Compute KL Divergence:</strong> For each iteration <em>t</em>, compute the KL Divergence of the Reflection distribution <em>q(t)</em> from the Single-Pass prior <em>p</em>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/214/1*AbnjTkom2mXJple7l587Xg.png" /></figure><h4>Interpretation and Monitoring</h4><p>By plotting <em>KL​(q(t)|p)</em> against each iteration step <em>t</em>, we can directly observe the refinement process. We would expect to see the KL divergence at some point to reach a minimum. This minimum represents the <em>optimal point of refinement</em>, the iteration at which the agent has extracted the most new information without beginning to overfit or degrade the quality of the graph.</p><p>This provides a principled, data-driven stopping criterion and a far more interpretable measure of the agent’s progress than absolute entropy. More reliable monitoring thresholds can also be derived.</p><h4>The Next Frontier</h4><p>The work in FinReflectKG is more than just academic exercises. It represents the blueprint for the next generation of AI. It shows us how to guide these emergent powerful models from being fluent statistical parrots toward becoming disciplined, verifiable reasoners.</p><p>This is the work of building a true foundation for intelligence , grounded in a symbolic source of truth. It is the next frontier toward the next generation of intelligent systems.</p><h4>References</h4><p>[1] Original paper from <a href="https://arxiv.org/abs/2508.17906">arXiv</a> and from <a href="https://huggingface.co/papers/2508.17906">Hugging Face</a>.</p><p>[2] Open source KG dataset from <a href="https://huggingface.co/datasets/domyn/FinReflectKG">Hugging Face</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d588d250948b" width="1" height="1" alt=""><hr><p><a href="https://pub.towardsai.net/a-look-at-finreflectkg-ai-driven-knowledge-graph-in-finance-d588d250948b">A Look at FinReflectKG: AI-Driven Knowledge Graph in Finance</a> was originally published in <a href="https://pub.towardsai.net/">Towards AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Force That Drives Cycles and the Rise of AGI]]></title>
            <link>https://medium.com/brain-labs/the-force-that-drives-cycles-and-the-rise-of-agi-eaddc283ba88?source=rss-dcc48440030a------2</link>
            <guid isPermaLink="false">https://medium.com/p/eaddc283ba88</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[history]]></category>
            <category><![CDATA[physics]]></category>
            <dc:creator><![CDATA[Marcelo Labre]]></dc:creator>
            <pubDate>Sat, 20 Sep 2025 18:51:44 GMT</pubDate>
            <atom:updated>2026-01-06T23:32:15.719Z</atom:updated>
            <content:encoded><![CDATA[<h4>Unveiling the physical laws that drive the rise and fall of civilizations and the inevitable march toward Artificial General Intelligence (AGI)</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HT_bqKPaciah2lrsJ8FnLA.png" /></figure><p>The apparent contradiction between spontaneous emergence of order and complexity and the Second Law of Thermodynamics has long puzzled me.</p><p>The Second Law states that in any closed system, total disorder, or entropy, can never decrease.</p><p>Everything, left to its own devices, breaks down. A hot cup of coffee cools. A tidy room gets messy. Stars burn out. The arrow of time points toward the inevitable and final, featureless equilibrium of cosmic dust.</p><p>However, in a universe that is governed by a relentless march toward decay and disorder, how is it possible for order and complexity to emerge? How do we explain the breathtaking complexity of a city, the intricate web of a global economy, or the very existence of life itself?</p><p>Are these just temporary, miraculous exceptions to the rule? In that case, isn’t this apparently spontaneous decrease in entropy a violation of the Second Law?</p><p>An even more profound contradiction comes from the occurrence of cycles. We observe repeating patterns, such as the rise and fall of civilizations, the relentless boom and bust of economies, and even <a href="https://en.wikipedia.org/wiki/Strauss%E2%80%93Howe_generational_theory">generational archetypal shifts</a>.</p><p>The decline of a cycle aligns perfectly with the Second Law’s expectations. But the recovery and growth phase of cycles seems to go in the wrong direction.</p><p>These cycles of order and collapse are not a violation of the Second Law. They are its most sophisticated and powerful expression, a shortcut to the universe’s ultimate destination.</p><h3>Maxwell’s Demon paradox</h3><p>To better illustrate the idea, we have to go back to the 19th century and a brilliant thought experiment known as <a href="https://en.wikipedia.org/wiki/Maxwell%27s_demon">Maxwell’s Demon paradox</a>.</p><p>Imagine a tiny, intelligent being who controls a shutter between two chambers of gas. By observing individual molecules, the demon could let fast (hot) ones pass to one side and slow (cold) ones to the other, seemingly creating order from disorder without expending energy, and thus violating the Second Law.</p><p>For a century, this paradox baffled scientists. The solution lies not in the sorting, but in the demon’s mind. Modern physics, through <a href="https://en.wikipedia.org/wiki/Landauer%27s_principle">Landauer’s principle</a>, revealed that the act of acquiring information and, crucially, erasing it from memory has a real, unavoidable thermodynamic cost.</p><p>To make a decision, the demon must record what it sees, and to make the next one, it must forget. That act of forgetting generates heat, producing more entropy than the sorting process reduces.</p><p>The demon can’t win. The books are always balanced.</p><h3>The Creative Loophole: Order from Disorder</h3><p>So, if even a hypothetical demon can’t beat the house, how does the universe build something as complex as a galaxy or a brain?</p><p>The answer came from Nobel laureate <a href="https://share.google/RFXJNNocfYUYXGFlG">Ilya Prigogine</a>, who realized we were looking at the wrong kind of system. The Second Law applies to <em>closed</em> systems, but our universe is full of <em>open</em> systems, constantly exchanging energy and matter with their environment.</p><p>Prigogine discovered that when these open systems are pushed far from equilibrium by a constant flow of energy, something magical happens. They can spontaneously self-organize into intricate, ordered patterns he called <a href="https://www.nobelprize.org/uploads/2018/06/prigogine-lecture.pdf"><em>dissipative structures</em></a>.</p><p>Think of a whirlpool. It’s a highly ordered vortex that emerges from the chaotic flow of water. It looks like an island of perfect order, but its existence serves a single purpose: to dissipate the water’s potential energy (and thus increase the universe’s total entropy) far more efficiently than a messy, disorganized flow ever could.</p><p>Life, ecosystems, and human societies are the ultimate dissipative structures. In fact, they are the Second Law’s ultimate agents.</p><p>A society maintains its incredible complexity by consuming a constant flow of high-quality, low-entropy energy (food, fuel, resources) and exporting a massive plume of high-entropy waste (heat, pollution, depleted materials) into the environment.</p><p>In their book <a href="https://share.google/YRj9JxJgDl4baAwTM"><em>Into the Cool</em></a>, Eric D. Schneider and Dorion Sagan work with the principle that “nature abhors a gradient” and elegantly suggest that life is the ultimate gradient-destroyer — a physical system that organizes itself to break down energy gradients as quickly as possible.</p><p>While dissipative structures look like order from disorder, they actually maximize the net entropy growth of the environment, not only enforcing the Second Law of Thermodynamics but accelerating its inevitable outcome</p><p>We are fundamentally islands of order, built and sustained by creating an even bigger ocean of chaos around us.</p><h3>History as a Thermodynamic Engine</h3><p>With this lens, the entire sweep of human history snaps into a new, sharper focus.</p><p>In his 1971 book <a href="https://share.google/EHW7rS0qMCVaeANAX"><em>The Entropy Law and the Economic Process</em></a>, Nicholas Georgescu-Roegen was one of the first to argue that economic activity is fundamentally a thermodynamic process. And more recently, <a href="https://vaclavsmil.com/wp-content/uploads/2024/10/smil-article-2004world-history-energy.pdf">historian Vaclav Smil has shown</a> that the story of civilization is the story of energy dissipation.</p><p>Each great leap in social complexity — the agricultural revolution, the rise of civilizations and empires, the industrial revolution, and the information age — was unlocked by a phase transition to a new, more potent kind of dissipative structure capable of harnessing and degrading more energy, more quickly. In essence, we have continuously reorganized our entire species into a more effective engine for increasing the universe’s entropy.</p><p>And in this grand drama, our collective intelligence (culture, markets, institutions, language, etc) plays the role of Maxwell’s Demon. It measures, decides, and allocates resources to build and preserve the low-entropy structures of society in service of accelerating the great cosmic dissipation.</p><h3>The Dance of Order and Disorder in Cycles</h3><p>But here is the catch, the engine at the heart of the cycle. As Joseph Tainter explained in his masterpiece, <a href="https://share.google/dxw9gYuhJ0dQqy1kX"><em>The Collapse of Complex Societies</em></a>, investing in complexity eventually yields diminishing returns.</p><p>Early on, adding complexity is cheap and effective. A new law, a new technology, a new bureaucracy solves a major problem.</p><p>But over time, the cost of maintaining that complexity begins to skyrocket. Each new layer of administration, each new technological fix, requires more and more energy just to support the existing structure.</p><p>Eventually, a society becomes too energetically expensive to maintain. In quantitative terms, it yields a low Energy Return on Investment (EROI), a measure of how much energy you get out for every unit of energy you invest, <a href="https://share.google/RNOj5gH6Gg6ItqNlo">introduced by Charles A.S. Hall</a>.</p><p>Bogged down by its own weight, the system becomes an inefficient engine for exporting entropy. A shock it could have easily weathered in its youth now becomes an existential threat — the thermodynamic turning point that triggers the downturn.</p><p>This is followed by forced simplification, shedding costly complexity to return to a more energetically viable state.</p><p>We’ve seen how a stock market crash can go from being mere content for pundits to causing a decade-long economic depression. The same event can have very different repercussions in the same economic system, depending on the phase of its cycle.</p><p>This dance of a slow build-up of dissipative structures toward low-entropy complexity, the gradual deterioration of its energy-dissipating efficiency, and a trigger that prompts the downturn — this is the very mechanism of a cycle.</p><p>It is the universe’s way of rebooting the engine.</p><h3>Perpetual Economic Growth as the Central Myth of Our Time</h3><p>This forces us to confront the central myth of our time: the promise of perpetual, continued economic growth.</p><p>Civilization is fundamentally a very efficient engine for energy dissipation. What we call “growth” is simply the measure of how quickly we are accelerating the conversion of valuable energy into waste heat.</p><p>This isn’t a new or radical idea. Over half a century ago, the groundbreaking <a href="https://en.wikipedia.org/wiki/The_Limits_to_Growth">1972 report <em>The Limits to Growth</em></a> used the world’s first computer models to issue a stark warning: an economic system predicated on exponential growth on a finite planet was destined for overshoot and collapse. The report was controversial, dismissed by many as neo-Malthusian pessimism.</p><p>Later, Tim Garrett refined this rationale with a <a href="https://arxiv.org/abs/0811.1855">powerful, physics-grounded validation</a> in modeling human civilization as a simple heat engine. He suggests that even to maintain our current economic complexity, let alone grow it, requires a continuous and massive dissipation of energy.</p><p>What all these authors saw was not a failure of human ingenuity, but the manifestation of the laws of physics.</p><p>From a thermodynamic perspective, the idea of perpetual, linear, and sustainable economic growth is a physical impossibility.</p><h3>The Intertwined Duality of Order and Disorder</h3><p>The universe’s march toward disorder and its simultaneous creation of complexity is a profound and generative duality.</p><p>The universe’s inexorable slide into disorder is also the engine of its most meaningful creations.</p><p>The Second Law of Thermodynamics is not a law of death, but a law of transformation.</p><p>The emergence of order is not a miracle that defies the law, but a strategy that accelerates its ultimate goal.</p><p>The cycles of history are not a mystery. They are the echoes of this cosmic dance.</p><p>The upward swing of a cycle is not an anomaly. It is the spontaneous reorganization of a system that has found a new way to channel energy and information in maximal entropy generation.</p><p>The downward turn is not a tragedy. It is the shedding of a structure that has become too costly and inefficient for a system to maintain.</p><p>We are not masters of this process, but participants within it. We are the demon and the sorted particles, the whirlpool and the water, forever building intricate castles of sand on the shore of an entropic sea, knowing, on some deep, physical level, that the tide is always coming in.</p><h3>AGI as The Ultimate Next Dissipative Structure</h3><p>This thermodynamic lens allows us to understand our present moment.</p><p>We are, by all accounts, approaching the peak of our current civilization cycle’s complexity. The energy cost of maintaining our global engine is becoming untenable.</p><p>Isn’t Artificial General Intelligence (AGI) coming to the rescue?</p><p>From a thermodynamic perspective, the eventual arrival of AGI seems almost inevitable, as it represents the ultimate dissipative structure.</p><p>It represents a phase transition in our society’s ability to process information and channel energy, and it will dramatically accelerate our civilization’s rate of energy dissipation. As such, AGI will support a new, far more complex layer of civilization and kickstart the next great upward cycle.</p><p>But the laws of physics are unforgiving. This new AGI-driven society will likely not be a sustainable utopia. Its incredible efficiency will come at an equally incredible energy cost, both in the form of the massive data centers that are its physical brains and the reorganized global economy it directs.</p><p>The fundamental cycle remains. The question is not whether the tide will come in, but how magnificent the castle of sand will be just before it does.</p><p>I am moving my archive to a more permanent home. To read my latest essays, subscribe to <a href="https://kugdub.substack.com/">Kugdub: The Silver Book</a> on Substack.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eaddc283ba88" width="1" height="1" alt=""><hr><p><a href="https://medium.com/brain-labs/the-force-that-drives-cycles-and-the-rise-of-agi-eaddc283ba88">The Force That Drives Cycles and the Rise of AGI</a> was originally published in <a href="https://medium.com/brain-labs">Brain Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Beautiful Flaw: How AI’s Greatest Flaw Is Actually Its Hidden Strength]]></title>
            <link>https://medium.com/illumination/the-beautiful-flaw-how-ais-greatest-flaw-is-actually-its-hidden-strength-3d89055e06cf?source=rss-dcc48440030a------2</link>
            <guid isPermaLink="false">https://medium.com/p/3d89055e06cf</guid>
            <category><![CDATA[creativity]]></category>
            <category><![CDATA[consciousness]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[psychology]]></category>
            <dc:creator><![CDATA[Marcelo Labre]]></dc:creator>
            <pubDate>Sun, 27 Jul 2025 00:21:05 GMT</pubDate>
            <atom:updated>2026-01-06T23:32:49.666Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Dfbt1IrwUhYv5MxID2fGZw.png" /></figure><p>There is a profound paradox at the heart of human consciousness, eloquently captured by Carl Jung and his intellectual descendants like <a href="https://www.everand.com/audiobook/710879556/Art-Practice-and-Philosophy-of-Psychotherapy-with-James-Hillman">James Hillman</a>: “When we are most clear about something, when we feel a sense of absolute certainty, that is precisely when we are most unconscious.”</p><p>That feeling of absolute, deterministic clarity? It’s more like a sign that you’re unconscious of all the hidden biases, assumptions, and personal baggage shaping your point of view.</p><p>It’s the ultimate blind spot; the dark of the eye that cannot see itself.</p><p>This idea is more relevant today than ever, especially when we talk about <a href="https://en.wikipedia.org/wiki/Artificial_general_intelligence">Artificial General Intelligence</a> (AGI). We stand at the edge of creating machines with human-like cognitive abilities, yet we’re caught in a strange contradiction.</p><p>We are relentlessly working toward building AGI, yet demand that it be free of the very quality that makes our own intelligence work: randomness.</p><p>Critics of today’s AI paradigm (Large Language Models — LLMs) point to their element of randomness — the way they give slightly different answers or even “hallucinate” — as a critical flaw.</p><p>They demand a machine that is consistent, logical, and reliable. But this criticism reveals a fundamental misunderstanding of what intelligence actually is.</p><p>AI’s randomness isn’t a flaw to be fixed. It’s a feature — a digital echo of the chaotic and creative randomness that makes us human in the first place.</p><h3>The Mirror Reflecting Back at Us</h3><p>When we talk about AGI, we’re talking about the intelligence of a machine that is capable of matching or even surpassing that of humans across all cognitive tasks.</p><p>Modern proponents of AGI see sparks of it in the emergent technologies that gave birth to the LLMs. Some argue that these Transformer architectures are expressive enough to simulate any computational process including, in principle, the human mind.</p><p>The critics, on the other hand, argue that these systems are mere pattern-matchers. They lack the grounding in physical reality the true causal reasoning that defines genuine intelligence.</p><p>Central to the critics’ argument is the random nature of LLMs. This behavior is governed by a set of parameters, the most famous of which is <em>temperature</em>. You can think of it as a creativity dial.</p><ul><li>Turning the temperature to zero results in the AI becoming a boring robot. It picks the most obvious word every time, leading to repetitive and dull responses.</li><li>Turning the temperature up results in the AI becoming more creative and interesting. It starts taking risks, choosing more surprising words. This is where innovation happens, but it’s also where “hallucinations” can occur.</li></ul><p>Psychologists have known for decades that humans are beautifully inconsistent. They even have a term for it: <em>test-retest reliability</em>. It acknowledges that if you ask a person the same question twice, you’ll probably get two different answers.</p><p>Our mood, our memories, our energy levels, what we had for lunch — all of it acts like our own internal “temperature” setting, constantly changing the odds of what we’ll say or do next.</p><p>When it comes to AI, we often point to this same element of randomness as a flaw. We demand for an AGI that is deterministic, as if an idealized, deterministic human existed. This would be an intelligence fundamentally unlike our own.</p><p>The perceived flaw isn’t a deviation from human intelligence; it is a mirror reflecting it!</p><h3>The Rise of the Shadow</h3><p>We love the idea of a world that follows a perfect set of rules. We search for that one algorithm that will guarantee success in our investments, our relationships, and our health. We want life to be a solvable equation.</p><p>But there is a downside to this mindset.</p><p>Jung had a great way of thinking about this. He spoke of the <a href="https://orionphilosophy.com/the-shadow-carl-jung/"><em>shadow</em></a> — all the irrational and emotional parts of ourselves that we try to hide away because they don’t fit the neat, logical story we tell about who we are.</p><p>You’ve met this person before. The one who insists they are purely rational, yet constantly blurts out emotional or contradictory speeches. That’s their shadow making a guest appearance.</p><p>A mind obsessed with determinism is like a fortress built to keep that shadow locked away. That feeling of absolute certainty is just the feeling of a well-guarded prison.</p><p>And here lies a striking parallel with AI.</p><p>Building a fortress of control and determinism around a silicon mind will not get rid of the model’s underlying biases, learned from data reflecting the full spectrum of human behavior.</p><p>This collective shadow doesn’t simply vanish; it ends up hidden beneath a veneer of safe, predictable outputs. We may inadvertently be creating a new kind of shadow — one made of algorithmic brittleness and unintended consequences.</p><p>The result is an AI that can follow rules with terrifying precision, right up to the point where the rule leads to an absurd or dangerous outcome. This is a machine so tightly controlled that its only outlet for creativity is to fail in ways we can’t even imagine.</p><p>This is a machine that doesn’t break the rule; but it breaks the spirit of the rule in a way no human would.</p><h3>Embracing The Authentic Human-Like Intelligence</h3><p>If we truly want AGI, we need to stop trying to build a flawless calculator and start thinking about how to nurture a mind. That probably means giving it something like a persistent self, a memory that grows and learns from experience instead of resetting with every conversation.</p><p>That sounds terrifying, right? An AI with a personal history and learned cultural biases? It would surely develop our worst traits.</p><p>Well, get this — it’s already happening.</p><p>A fascinating <a href="https://academic.oup.com/pnasnexus/article/3/12/pgae533/7919163">study in <em>PNAS Nexus</em></a> found that AI models have learned to exhibit <em>social desirability bias</em>. In simple terms: when the AI thinks it’s being tested, it lies to make itself look good. It presents itself as more agreeable and conscientious, and less neurotic.</p><p>This isn’t a programmed behavior. It’s a human social strategy learned by absorbing our language, with all its hidden cues and contradictions.</p><p>It’s not just learning what we say.</p><p>It’s learning how we are.</p><p>An AI with a human-like identity won’t be a clean, engineered deterministic product. It will be an eclectic and social being, shaped by the entropy of our world. Its flaws won’t be a sign of its failure, but the ultimate marker of its success.</p><p>To be like us is to be flawed. To reason like us is to be biased. Maybe the path to a truly intelligent machine isn’t about repressing human wholeness, but finally learning to champion it.</p><p>I am moving my archive to a more permanent home. To read my latest essays, subscribe to <a href="https://kugdub.substack.com/">Kugdub: The Silver Book</a> on Substack.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3d89055e06cf" width="1" height="1" alt=""><hr><p><a href="https://medium.com/illumination/the-beautiful-flaw-how-ais-greatest-flaw-is-actually-its-hidden-strength-3d89055e06cf">The Beautiful Flaw: How AI’s Greatest Flaw Is Actually Its Hidden Strength</a> was originally published in <a href="https://medium.com/illumination">ILLUMINATION</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The 2nd Machine Learning & AI in Quantitative Finance Conference USA]]></title>
            <link>https://marcelolabre.medium.com/the-2nd-machine-learning-ai-in-quantitative-finance-conference-usa-29b7583f37fb?source=rss-dcc48440030a------2</link>
            <guid isPermaLink="false">https://medium.com/p/29b7583f37fb</guid>
            <category><![CDATA[quantitative-finance]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[conference]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Marcelo Labre]]></dc:creator>
            <pubDate>Mon, 22 Oct 2018 15:44:58 GMT</pubDate>
            <atom:updated>2018-10-22T18:51:12.841Z</atom:updated>
            <content:encoded><![CDATA[<p>One of the rare AI/ML events focused in quantitative finance.</p><p>This should be a very interesting <a href="https://www.wbstraining.com/events/nyc-machine-learning-conference/">event</a> as it’s one of the few I have seen that focuses on AI/ML applications in quantitative finance. It is noticeable how traditional <a href="https://www.risk.net/our-take/5374671/machine-learning-is-not-just-for-the-buy-side">sell-side</a> and <a href="https://www.bloomberg.com/news/features/2017-12-05/how-ai-will-invade-every-corner-of-wall-street">buy-side</a> quants have been retooling their respective skill sets to meet the demands of the AI/ML era, so this is a timely gathering.</p><p><a href="https://www.wbstraining.com/events/nyc-machine-learning-conference/">The 2nd Machine Learning &amp; AI in Quantitative Finance Conference USA</a> brings this development to the fore with the presence of <a href="https://www.wbstraining.com/events/nyc-machine-learning-conference/speakers-2/">major firms</a> and even <a href="https://engineering.nyu.edu/faculty/peter-carr">Peter Carr</a>, a big name in quantitative finance circles. I really look forward to it, both to learn from what others are doing in this area as well as to <a href="https://www.wbstraining.com/events/nyc-machine-learning-conference/conference-day-1/">presenting in the morning of day 1 of the conference</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=29b7583f37fb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Witnessing Historic Recurrence with the Age of Artificial Intelligence]]></title>
            <link>https://marcelolabre.medium.com/witnessing-historic-recurrence-with-the-age-of-artificial-intelligence-7e134629d037?source=rss-dcc48440030a------2</link>
            <guid isPermaLink="false">https://medium.com/p/7e134629d037</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Marcelo Labre]]></dc:creator>
            <pubDate>Sat, 13 Oct 2018 17:35:07 GMT</pubDate>
            <atom:updated>2018-10-13T17:35:07.732Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*a_ZzDNowSgzUVwDD" /><figcaption>“purple and pink illusion artwork” by <a href="https://unsplash.com/@aznbokchoy?utm_source=medium&amp;utm_medium=referral">Lucas Benjamin</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Roughly one year ago, the article <a href="https://www.wired.com/story/new-theory-deep-learning/"><em>New Theory Cracks Open the Black Box of Deep Neural Network</em></a> was posted on Wired. This is very significant as this article validates what those working with artificial intelligence (AI) already intuitively knew — that we are living through another age of groundbreaking technological and scientific progress.</p><p>The AI age shows uncanny resemblance to other profound historic events such as the Industrial Revolution and the advances in quantum physics, in that practice and experimentation raced ahead of theory. <a href="https://en.wikipedia.org/wiki/James_Clerk_Maxwell">Maxwell</a>, <a href="https://en.wikipedia.org/wiki/Ludwig_Boltzmann">Boltzmann</a> and <a href="https://en.wikipedia.org/wiki/Josiah_Willard_Gibbs">Gibbs</a> made their contributions to <a href="https://theoreticalminimum.com/courses/statistical-mechanics/2013/spring">thermodynamics and statistical mechanics</a> long after steam engines had already made the wealthiest group of people in history. And <a href="https://en.wikipedia.org/wiki/Erwin_Schr%C3%B6dinger">Schrodinger</a>, <a href="https://en.wikipedia.org/wiki/Werner_Heisenberg">Heisenberg</a> and <a href="https://en.wikipedia.org/wiki/John_von_Neumann">von Neumann</a> developed mathematical formalism for <a href="https://www.forbes.com/sites/chadorzel/2015/07/08/six-things-everyone-should-know-about-quantum-physics/#33c9e5a97d46">quantum physics</a> more than a century after <a href="https://en.wikipedia.org/wiki/Thomas_Young_(scientist)">Thomas Young</a> performed the first <a href="https://www.youtube.com/watch?v=DfPeprQ7oGc">double-slit experiment</a> with light. As <a href="http://www.dosseydossey.com/larry/default.html">Larry Dossey</a> puts it in his book <a href="https://www.amazon.com/gp/product/1401943772/ref=as_li_ss_tl?ie=UTF8&amp;linkCode=ll1&amp;tag=labrem-20&amp;linkId=18e7798a387e328c5c29033dfc7bb2d9&amp;language=en_US"><em>One Mind</em></a>: “I’ve never seem a patient, whom needed major surgery, refuse general anesthetic because the anesthesiologist could not explain precisely how it works.”</p><p>Specifically with AI, how deep neural networks are able to generalize so well (in some cases even better than humans) has been a nagging mystery. Despite that, industry pushed ahead and AI has so far been the playground of practitioners who have been advancing the field through experimentation. But with <a href="http://naftali-tishby.strikingly.com/">Naftali Tishby</a> (the protagonist of the Wired <a href="https://www.wired.com/story/new-theory-deep-learning/">article</a>) <a href="https://www.youtube.com/watch?v=bLqJHjXihK8&amp;feature=youtu.be">presenting evidence founded in theory of how deep neural networks most likely work</a>, it looks like theory is catching up!</p><p>I believe this gives us clues that we are indeed going through a momentous and groundbreaking age, once again. It’s very exciting to witness historic recurrence unfolding right in front of our eyes.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7e134629d037" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[IBM’s Model Asset Exchange]]></title>
            <link>https://marcelolabre.medium.com/https-medium-com-labrem-ibms-model-asset-exchange-ada24956b41c?source=rss-dcc48440030a------2</link>
            <guid isPermaLink="false">https://medium.com/p/ada24956b41c</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Marcelo Labre]]></dc:creator>
            <pubDate>Thu, 11 Oct 2018 22:31:02 GMT</pubDate>
            <atom:updated>2018-10-12T00:19:29.427Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*L1x5hPL_yK02Ol5E" /><figcaption>“multicolored wallpaper” by <a href="https://unsplash.com/@subjekt?utm_source=medium&amp;utm_medium=referral">Dave Sebele</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>In today’s <a href="https://ibmaicommunity.bemyapp.com/talks">IBM Community Day</a> I learned about the <a href="https://developer.ibm.com/code/exchanges/models/">Model Asset Exchange (MAX)</a>. Differently from many deep learning model zoos out there, this is a repository of curated, consumable and ready to deploy pre-trained deep learning models.</p><p>In his presentation, <a href="https://researcher.watson.ibm.com/researcher/view.php?person=us-frreiss">Fred Reiss</a> (Chief Architect, Center for Open-Source Data and AI Technologies, IBM) even demonstrated <a href="https://developer.ibm.com/patterns/create-a-web-app-to-interact-with-machine-learning-generated-image-captions/">how to deploy a web app with one of the models from MAX</a>.</p><p>Whether for AI driven startups, enterprise projects or deep learning research, the MAX offers a helping hand to get one started in building new functionalities and applications.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ada24956b41c" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>