<?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 Carlos Araya on Medium]]></title>
        <description><![CDATA[Stories by Carlos Araya on Medium]]></description>
        <link>https://medium.com/@carlos.araya?source=rss-8a09ab0e80f6------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Y-T6i36wLJ_etycmGuUHAw.jpeg</url>
            <title>Stories by Carlos Araya on Medium</title>
            <link>https://medium.com/@carlos.araya?source=rss-8a09ab0e80f6------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 11:34:02 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@carlos.araya/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[Solving the AI Alignment Problem]]></title>
            <link>https://medium.com/@carlos.araya/solving-the-ai-alignment-problem-28be3aaf0a5c?source=rss-8a09ab0e80f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/28be3aaf0a5c</guid>
            <category><![CDATA[lean]]></category>
            <category><![CDATA[programming-languages]]></category>
            <category><![CDATA[isla]]></category>
            <category><![CDATA[ai-alignment]]></category>
            <category><![CDATA[belief-systems]]></category>
            <dc:creator><![CDATA[Carlos Araya]]></dc:creator>
            <pubDate>Tue, 15 Sep 2026 02:43:41 GMT</pubDate>
            <atom:updated>2026-09-15T03:04:13.424Z</atom:updated>
            <content:encoded><![CDATA[<p>An aligned agent is one for which the forbidden situations are inaccessible — not improbable</p><p>By Carlos Araya</p><p>— -</p><p>Solving the alignment problem — the problem of ensuring that an AI agent’s reasoning and conclusions follow explicitly stated desires and intentions, with guarantees — has become the central challenge of deploying AI agents responsibly.</p><p>Here we describe a method that solves the alignment problem <em>by construction</em>. It follows naturally from our research on building formal artificial belief systems with a reflexive equation of the form:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/754/1*GVHjV6e-_0rICAwOm0Opqw.png" /><figcaption>Fig 1: Reflexion</figcaption></figure><p>The idea is simple: for a given agent <strong><em>K</em></strong>, its beliefs and its reasoning must live inside the semantic space of its Ethics and Values. Any solution to this equation therefore conforms to an explicitly established set of ethics and values.</p><p>Modal logics formalize the concepts of necessity and possibility. For years they have been the approach of choice for analyzing obligations, permissions, and prohibitions. What a given entity <em>should</em> do and what it <em>might</em> do, expressed in formal terms, is the basis of our proposal. As we show below, an aligned agent is one for which the forbidden situations are inaccessible — not improbable.</p><h4>A little context</h4><p>In the 1670s and 1680s, Gottfried W. Leibniz proposed the existence of certain abstract concepts he called <em>worlds</em>: descriptions of alternative situations the world could be in. He built the foundation of his modal metaphysics, and the idea of alternative worlds, while working through problems of divine creation, necessity, and contingency.</p><p>In 1959, the logician Saul Kripke used this intuition to propose a semantic method for analyzing modal theories, known as <em>possible worlds semantics</em>. Under Kripke’s method, a sentence is necessary, possible, or impossible depending on whether the worlds in which it is true are all worlds, some worlds, or none.Building on these ideas, we developed the <strong><em>Aleph</em></strong> logical system specifically to represent belief systems — systems that can hold beliefs and reason with them over finite-state systems, such as computers.</p><p>In Aleph, every proposition is a <em>space</em>: the set of possible worlds where it holds. ℵ is every world; ∅ is none. Because every domain is finite, ∧ is intersection, ∨ is union, and entailment is containment of possible worlds. By construction, Aleph is sound and complete with respect to the states of any finite-state system, and it is decidable.</p><p>Here is the equivalence above, rephrased in Aleph as a combination of set and logic operations:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/936/1*68ypW7srnpKvLXn_A4ChcA.png" /><figcaption>Fig 2: Reflexion in Aleph</figcaption></figure><p>In this equivalence, <strong><em>T</em></strong> — the theory — represents the ethics. The subjunctive conditionals can be used to represent values, among other actions: for a given conditional, if <strong><em>K</em></strong> entails <strong><em>N — </em></strong><em>necessary preconditions</em>, and <strong><em>P </em></strong>is consistent with <strong><em>K</em></strong> <strong><em>— </em></strong><em>possible preconditions</em>, then the conclusion <strong><em>C</em></strong> must be part of every solution to K. Finally, <strong><em>B</em></strong> represents the contingent beliefs — those that correspond to observations that may or may not be the case. <strong><em>B</em></strong> beliefs are acquired and incorporated through the operations described below.</p><p>As can be shown, the solution to these equations is a bifurcation of alternative fixpoints, each representing a different state of belief:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/936/1*eRPjaA3RdubtBs7zuK8iWw.png" /><figcaption>Fig 3: Fixpoints bifurcation</figcaption></figure><p>The important observation is that each fixpoint is contained inside the semantic space of the ethics, represented by <strong><em>T</em></strong>. In <strong><em>K = T ∧ Rᵢ</em></strong>, the theory <strong><em>T</em></strong> is immutable law: it removes worlds, and nothing can put them back. The behaviors <strong><em>Rᵢ </em></strong>are what the agent currently believes, held as one or more states (scenarios) — they include <strong><em>B</em></strong> together with the consequences of the conditionals and the ethics.</p><p>The conditionals — subjunctive, “if χ were the case, φ” — can help describe the value system: defaults, obligations, priorities, evidence. Values may yield to one another; laws never yield. Conditionals express dispositions and actions, not truths (Ramsey tests). They do not suffer the paradoxes of material implication found in first-order logics, and they survive the resolution of the equivalence.</p><h4>Why these equivalences model belief systems</h4><p>Aleph agents are formal belief systems in the sense of the AGM postulates. Every operation offered is a set operation with a semantics you can check:</p><p>- <strong>Expand</strong>, <strong><em>K + A = K ∧ A</em></strong>.<br>- <strong>Revise</strong>, <strong><em>K ∗ A</em></strong> — accept <strong><em>A</em></strong> and keep as much of <strong><em>K</em></strong> as you can (in agreement with the <strong>AGM</strong> postulates: Success, Inclusion, Consistency, Vacuity).<br>- <strong>Retract</strong>, <strong><em>K ÷ A</em></strong> — stop believing <strong><em>A</em></strong>; never a theorem of <strong><em>T</em></strong>.<br>- The <strong>fixpoint</strong> — the self-reflexive synonymy gives the least fixpoint of a system of conditionals, with mutual support excluding self-sustaining cycles.</p><p>In short: in Aleph, alignment is a statement about possible worlds, not about outputs. An aligned agent is one for which the forbidden situations are inaccessible — not improbable.</p><h4>ISLA: a language for belief systems</h4><p>We have developed a programming language for handling belief systems, called <strong>ISLA</strong>, the Information Spaces Language for Agents. <strong>ISLA</strong> is a functional programming language and an interactive model assistant, built on handling the space of possible states of a finite-state system.</p><p><strong>ISLA</strong> stands to models as <strong>LEAN</strong> stands to proofs. In <strong>LEAN</strong>, a program is a proof: it establishes a theorem by proof-theoretic construction. In <strong>ISLA</strong>, a program is a model: it constructs the space of worlds in which its judgments hold, and it computes with that space. A theorem in ISLA is validated when its space is the whole universe, and refuted by the counterexample space that remains. <strong>LEAN</strong> is proof-theoretic; <strong>ISLA</strong> is model-theoretic.</p><p><strong>ISLA</strong> uses <strong>LLMs</strong> for what they are good at — processing language — and only at the doors. Its core is a deterministic, explainable, and provable engine: belief systems over finite possible worlds, computed by highly scalable deduction and derivation methods — fixpoints, revision, and certified constraint closure — where every answer is an exact set with a witness: the model. It is built for mission-critical applications, where a refusal must carry its reason and a result must be certifiable. An immutable theory keeps the agent aligned by construction.</p><p>— -</p><p><em>Write to us at </em><a href="/service/mailto:info@singularities.com"><em>info@singularities.com</em></a><em> for more information about ISLA.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=28be3aaf0a5c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Reasoning and the Limitations of Humans and Machines]]></title>
            <link>https://medium.com/@carlos.araya/reasoning-and-the-limitations-of-humans-and-machines-670a7dab7ae1?source=rss-8a09ab0e80f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/670a7dab7ae1</guid>
            <category><![CDATA[formal-logic]]></category>
            <category><![CDATA[agi]]></category>
            <category><![CDATA[finite-state-machine]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[reasoning]]></category>
            <dc:creator><![CDATA[Carlos Araya]]></dc:creator>
            <pubDate>Sat, 26 Oct 2024 17:57:37 GMT</pubDate>
            <atom:updated>2024-10-27T15:05:45.885Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1019/1*xSNgioFxWL0dNkEitHiH5Q.png" /></figure><p><strong>Summary:</strong> Integrating formal reasoning capabilities with various information representation and processing methods is a promising pathway toward achieving a more General Artificial Intelligence. Serving as an amalgam to bridge the gap between different paradigms is one of the objectives of a novel logical system called Aleph, which offers a framework for modeling agents that can reason, make decisions, and adapt while considering resource constraints.</p><p>This research illuminates specific challenges and offers insights on leveraging the strengths of generative models, constraint satisfaction, and others to produce smarter solutions to various problems.</p><p>There is a growing consensus that true Artificial General Intelligence (AGI) requires more than pattern recognition and data-driven learning capabilities. It needs the ability to reason, reflect, and make decisions like human cognitive processes, what psychologist Daniel Kahneman calls “System 2” thinking.</p><p>Applying force from a bad position or using inadequate tools represents poor choices. In this article, from the lens of experience in developing a formal system for intelligent agents, we will attempt to describe the challenge that intelligence represents for machines and how to take advantage of the new and wonderful strengths of various methods that, although not developed for reasoning, are new and useful opportunities for this purpose.</p><p><strong>The Need for Reasoning in AI</strong></p><p>Generative models often fail tasks that require deep reasoning, understanding complex relationships, precision, and, more importantly, adaptation to new information. Recent research, such as the GSM-Symbolic paper by Mirzadeh et al., highlights these limitations and underscores the challenges of large language models in mathematical reasoning and problem-solving.</p><p>According to our experience, the challenge of reasoning in the context of artificial agents with System 2 capabilities, as recently suggested, is as much about rationality and decidability as it is about the resources required to respond on time. The challenge is a product of intelligence&#39;s inherent limitations, not of any intelligent entity itself, as discussed below.</p><p>Addressing this challenge with a formal logical system called “Aleph,” aimed at intelligent agents that can reason, decide, adapt, and serve as the basis for solutions to practical problems, has led us to three fundamental design principles: the requirement of intelligence, the potential scope of the apparatus, and the tools for scalability:</p><p><strong>The Intelligence Challenge</strong></p><p>To meet the requirements of a formal belief system, as described by the epistemological postulates known as AGM — the gold standard for rational agents — the system must be capable of reflecting on its knowledge to determine the consequences of its actions.</p><p>The AGM postulates propose how rational belief representations — consistent and omniscient representations that contain all the consequences of their own beliefs — should be updated in light of new information. Evidently, these postulates do not describe a human, according to Kahneman. However, the reasons that distance us from this will also distance machines: limitations. We will analyze how the limitations will lead AGI solutions to seek a balance between comprehensive reasoning and their computational resources.</p><p><strong><em>We offer a non-anthropocentric definition of an intelligent entity: one that reacts rationally under time and resource constraints.</em></strong></p><p>Adapting to change according to the AGM postulates involves determining the largest consistent subsets of what is believed and what is implied by the new information. Very often, any agent will face impractical computational demands, leading to unfeasible calculations, delays, and imprecision, as well as learning and using heuristics as described by System 1 and System 2.</p><p>Our experience tells us that this requirement demands <em>O(2(2^n)) </em>consistency checks when dealing with subsets (cliques) of up to <em>n</em> basic pieces of information. Interestingly, the Blue Brain Project reports that the brain thinks in up to 11 dimensions when trying to establish a clique of neurons to produce a signal.</p><p>We conclude that the ability to tackle the challenge of these calculations limits the capacity of any intelligent system. However, in this complexity lies an opportunity. Beyond the prodigious results in natural language from generative methods, the results in determining protein structures are impressive. Our intuition is that systems capable of improving sequence prediction, such as Liquid AI, can help predict candidates for these cliques, whose consistency can then be checked with formal methods. The more powerful these capabilities are, the larger the <em>n</em> to which the artificial entity can aspire, and therefore, the better the quality of the solutions produced.</p><p>From Aleph&#39;s implementation perspective, generative methods are part of the research areas that seek to complement the collection of diverse strategies for solving this problem.</p><p><strong>The Capabilities of the Apparatus</strong></p><p>To ensure practicality and feasibility on modern digital computers, the Aleph logical system was specifically designed to deal with what is possible to achieve with a digital computer. Accordingly, the semantics of the logical language is any finite-state system based on the semantic concepts of possible worlds used in formal philosophy.</p><p>The resulting finite-state model system has semantics bounded precisely by the number of individual concepts and terms, which produces a finite semantic information space — Aleph is sound and complete with respect to any finite-state system. Therefore, the system can symbolically handle any finite-state system&#39;s possible states (or “worlds”), providing a rich and intuitive semantic foundation.</p><p>Additionally, the finite nature ensures that reasoning processes are decidable and, in many cases, efficiently executable, which is crucial for real-time decision-making.</p><p>In this way, Aleph provides an intuitive and understandable way of representing information spaces and offers efficient and concise abstract symbolic handling through the algebraic operations represented by the logical operators.</p><p><strong>Scalable by Extension</strong></p><p>Under certain criteria, axiomatic systems have the very promising property of being formally and naturally extensible with logical theories about other concepts, such as time, legal, business, or scientific. Intuitively, a formal logical theory can be seen as the language of expressions and theorems generated by its axioms and inference rules. Including other theories under these criteria will produce a broader and richer language.</p><p>In the case of Aleph, its confinement to digital computers allows implementing systems that can be integrated with other mathematical software systems, such as constraint satisfaction, optimization, machine learning of biases and heuristics, etc., in addition to generative models, to take advantage of the merits offered by such methods.</p><p>Finally, calculating solutions to model change leads to multiple solutions, representing an opportunity for symbolic calculations over complex state spaces that can be parallelized, improving efficiency. This possibility, and the bounding of the size of the cliques mentioned above, allows the response time of the implementations to be bound to what is possible and timely.</p><p><strong>Conclusion</strong></p><p>This article highlights the importance of merging learning with reasoning and other techniques to address the inherent limitations in AI systems and emphasizes rationality under resource constraints.</p><p>Integrating formal reasoning with generative models represents a significant step toward achieving a more General Artificial Intelligence. The possibility of joining diverse problem-solving method capabilities gives us optimism that this will play a crucial role in increasingly useful applications.</p><p>If you would like an extended version of this article, including the formal description of Aleph, please feel free to contact me at carlos.araya@singularities.com.</p><p><strong>References:</strong></p><ul><li>Mirzadeh, I., Alizadeh, K., Shahrokhi, H., Tüzel, O., Bengio, S., &amp; Farajtabar, M. (2024). <em>GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models</em>. arXiv:2410.05229.</li><li>Kahneman, D. (2011). <em>Thinking, Fast and Slow</em>. Farrar, Straus and Giroux.</li><li>Grady, Pat, and Matt Miller. (2023). <em>Generative AI’s Act II</em>. Sequoia Capital. Retrieved from <a href="https://www.sequoiacap.com/article/generative-ais-act-o1/">https://www.sequoiacap.com/article/generative-ais-act-o1/</a></li><li>Gärdenfors, P. (1988). <em>Knowledge in Flux: Modeling the Dynamics of Epistemic States</em>. The MIT Press.</li><li>Blue Brain Project. <a href="https://www.epfl.ch/research/domains/bluebrain/">https://www.epfl.ch/research/domains/bluebrain/</a></li><li>Reimann, M. W., Nolte, M., Scolamiero, M., Turner, K., Perin, R., Chindemi, G., Dłotko, Paweł, Levi, R., Hess, K., &amp; Markram, H. (2017). <em>Cliques of Neurons Bound into Cavities Provide a Missing Link between Structure and Function</em>. <em>Frontiers in Computational Neuroscience</em>, <em>11</em>, 48. Retrieved from <a href="https://doi.org/10.3389/fncom.2017.00048">https://doi.org/10.3389/fncom.2017.00048</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=670a7dab7ae1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Symbiosis Unleashed: Amplifying AI with the Integration of Logic and Generative Large Language…]]></title>
            <link>https://medium.com/@carlos.araya/symbiosis-unleashed-amplifying-ai-with-the-integration-of-logic-and-generative-large-language-140b2a4c9f3a?source=rss-8a09ab0e80f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/140b2a4c9f3a</guid>
            <dc:creator><![CDATA[Carlos Araya]]></dc:creator>
            <pubDate>Tue, 27 Jun 2023 04:22:46 GMT</pubDate>
            <atom:updated>2023-06-27T04:22:46.351Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/720/1*3dPurUvsfzXoBeqWYwQ9zg.png" /><figcaption>Solving problems the human way: Theory of Mind</figcaption></figure><h3>Symbiosis Unleashed: Amplifying AI with the Integration of Logic and Generative Large Language Models</h3><p>As we delve deeper into the AI era, we are confronted with a growing need for supporting software solutions to real-world business and engineering problems. In response to this need, we propose the marriage of formal logic, particularly a new innovative modal logic called Aleph, with generative Large Language Models (LLMs).</p><p>This alliance is more than just a merger; it’s a symbiosis that aims to give AI a considerable boost, enabling it to effectively address the solution of complex problems that exist in the real world. Through the use of logic, we propose a promising strategy to generate robust, explainable, rational, scalable solutions that leverage LLMs without reasoning errors.</p><p>Starting with the core benefits of this symbiotic integration, we elaborate on its strong mathematical foundation, ability to handle transitions, adherence to the natural Theory of Mind paradigm, and compatibility with industry-level purpose-specific problem solvers.</p><p>The Aleph logic, we argue, is an invaluable tool for this purpose. With its clear semantics, impressive expressiveness, and natural expansiveness, Aleph grants AI unprecedented abstraction in problem-solving strategies.</p><p>We illustrate how the union of LLMs and Aleph can simplify tasks like ontology engineering, the abstractions of processes, secure valid deductions, and offer explanations to users.</p><p>We then explore the roots of Aleph, its evolution to represent and reason about all potential states of a problem, and its unique characteristics. Additionally, we exemplify practical applications of Aleph in sectors such as e-commerce, supply chain optimization, and fulfillment processes, underlining its effectiveness and broad-ranging potential.</p><p>In summary, we aim to make a persuasive argument for the synergistic relationship between logic and LLMs. We are convinced that this collaboration could herald the advent of a new age of AI applications — more logical, explainable, scalable, robust, and, most importantly, attuned to the complexities of real business and engineering scenarios.</p><h3>Understanding the Aleph Logic</h3><p><strong>Aleph</strong> is a <em>modal</em> logic specifically designed for business and engineering applications targeting an <strong><em>agent-based</em></strong><em> approach </em>(Artificial General Intelligences). It<strong> </strong>offers a formal language that extends first-order logic by incorporating modal concepts like necessity and contingency.</p><p>However, <strong>Aleph</strong>’s most significant advancement is in its well-defined semantics, rooted in the potential states of business or engineering systems. This logic offers a mathematical framework for modeling solutions where the manipulated entities are not <strong><em>numbers</em></strong> but rather the <strong><em>spaces</em></strong> of problem states. Aleph is <em>sound </em>and<em> complete</em> in this regard.</p><h3>Embracing the Benefits of Symbiosis</h3><p>The union of Aleph modal logic and generative LLMs presents four key benefits:</p><p>1. A <strong>strong mathematical foundation</strong> that allows robust concepts such as belief systems, behaviors, ethics and obligations, self-reflection, and more.</p><p>2. To build <strong>agents</strong> that not only represent specific situations but also capture transitions that depend on previous situations. In fact, Aleph’s agents fulfill the AGM postulates<a href="#_ftn1">[1]</a>, ensuring logical principles for rationality and coherent belief revision.</p><p>3. To solve problems using a natural <strong>Theory of Mind</strong> approach. This cognitive ability to understand and attribute mental states and intentions to others and ourselves<a href="#_ftn2">[2]</a>, is a successful evolutionary capability that facilitates analysis and adaptation to complex situations involving multiple agents in a modularized manner.</p><p>4. The integration of logic-based systems with <strong>specialized industry-level problem solvers</strong>, such as theorem provers and deductive technologies, constraint satisfaction, and optimization solvers, allows the modeling and solving of large-scale problems intuitively.</p><p>The Aleph logic’s clear semantics, remarkable expressiveness, natural expansiveness, and problem-solving strategies that are reminiscent of how humans represent the world and deal with problems through the modeling with equations and variables offer a new high level of unparallel abstraction.</p><h3>The Future with Aleph and Generative Language Models</h3><p>There’s a promising opportunity to unite formal logic with the advancements in LLMs, paving the way for more explainable, scalable, robust, and rational AI business and engineering applications.</p><p>The potential benefits of this combination are numerous:</p><p>1. LLMs can help to streamline the laborious task of <strong>ontology engineering</strong>, typically a prerequisite for large logical systems, by helping to discover concepts and relationships in corpora of different sorts, including documentation, manuals, process descriptions, and business transactional data.</p><p>2. The relationships between <strong>formal languages </strong>and<strong> automata</strong> and the Internet of Things sensors could help bring the LLM’s capabilities to learn from observation and create automata representing abstractions of diverse processes.</p><p>3. Logics are essentially equivalent to the language containing all the theorems inductively generated from their axioms and inference rules<a href="#_ftn3">[3]</a>. If the methods of LLMs are constrained to such language, we will ensure that the deductions carried out are valid. The real value of this characterization comes from the following.</p><p>4. An often-overlooked feature of logic is that they are inherently extensible with other theories. If the base logic, like Aleph, has certain properties, like soundness and completeness, incorporating a new logic dealing with other concepts, like time, sortals, or business theories, a larger and richer symbolic language will emerge with the same property, as long as the logic axioms and inference preserve the property<a href="#_ftn4">[4]</a>. This unique capability enables a system like Aleph to be the ideal foundation for constructing more potent AI theories.</p><p>5. LLMs can help synthesize descriptions of the steps carried out by long deductive and calculating computations to provide linguistic or graphical explanations to users.</p><p>6. In addition, generative methods can gradually learn to choose the most promising deductive and calculating steps in large search spaces, similarly as we become experts in certain domains. This could become a fantastic opportunity now that proof methods are starting to dominate search strategies in the struggle with the complexity and practicality of real-life situations due to an overwhelming abundance of intricate details<a href="#_ftn5">[5]</a>. The approach has already been explored in mathematics with some good results<a href="#_ftn6">[6]</a>.</p><p>7. Finally, a similar approach could help mitigate the computational irreducibility problem, as pointed out by Stephen Wolfram<a href="#_ftn7">[7]</a>, by avoiding exploring all possible branches to find viable solutions.</p><p>LLMs are a testament to how new approaches can overcome several limitations. <strong>Aleph</strong> doesn’t break any theoretical limitations, but since it is tailored to represent and reason over business problem states of computer applications, and since it can seamlessly complement other technologies such as LLMs, theorem provers, constraint-satisfaction, and linear programming tools, it is in the right direction for more expressive, manageable, expandable, and scalable solutions.</p><h3>Aleph: A Deep Dive into its Foundations</h3><p><strong>Aleph</strong> was developed to represent and reason about all possible states that a business, science, or engineering problem can assume, commonly known as the “possible worlds” of the problem. This approach diverges from the conventional focus on a single state managed by most programming languages.</p><p>The notion of “possible worlds” is a semantic concept that can be traced back to the work of Philosopher Gottfried Leibniz in the 17th century, who proposed the existence of an infinite number of possible worlds, each representing a distinct way the world could be. Saul Kripke and David Lewis proposed the idea of using it as a framework to comprehend the semantics of modal statements in the 20th century. Kripke suggested that possible worlds could be considered abstract entities, each representing a comprehensive and internally coherent representation of how the world could be, whereas Lewis considered them concrete entities.</p><p>To illustrate, consider mathematics, where a numeral such as ‘4’ denotes the property shared by all sets with exactly four elements, known as the number four<a href="#_ftn8">[8]</a>. In <strong>Aleph</strong>, a statement like ‘Year = 2023’ precisely denotes what is shared by all the “worlds” in which Year equals 2023. Aleph ensures a complete correspondence between statements and the space constituted by the possible states where those statements hold. Technically speaking, the <strong>Aleph</strong> logic is both <em>sound</em> and <em>complete</em> regarding the universe of “possible worlds” within a given finite-state problem or system — referring to one with a defined yet potentially extensive range of possible states.</p><p><strong>Aleph</strong> is a significant leap forward from previous efforts of using modal logic in AI. It introduces several innovative features to tackle practical, real-world problems effectively following straightforward procedures:</p><p>1. The universe of problem states is generated using typed structures to describe the problem <strong>domains</strong> (data) and functions and variables to reference the domain values.</p><p>2. The problem spaces are described and operated using <strong>statements</strong> composed of operations like <strong>and</strong><em> </em>(intersection), <strong>or</strong> (union), <strong>not</strong> (complement), and <strong>necessity</strong> (for differentiating whether a given statement is true across all states).</p><p>3. Find parameters using <strong>machine learning </strong>or<strong> </strong>other methods<strong>.</strong></p><p>4. Model solution using reflective <strong>equivalences</strong> (‘<em>equations</em>’) as in finance, engineering, and economy.</p><p>5. Specify <strong>goals</strong> using maximizations and <strong>reflection</strong>.</p><p>6. <strong>Aleph</strong> finds the roots of the equation — <strong>singularities</strong>,<strong><em> </em></strong>using advanced<strong> proprietary</strong> and <strong>off-the-shelf </strong>algorithms.</p><h3>Practical Applications</h3><p>Aleph logic and implementation have been tried in several applications, such as <em>client</em> <em>agents</em> for e-commerce, problem specification, supply chain optimization, and fulfillment processes.</p><p><strong>Aleph</strong> logic expressions seamlessly translate into the representations employed by various problem-solving methods. Therefore, <strong>Aleph</strong> can be used to automate the problem description required by these methods and analyze their outputs, facilitating result improvement through interactive processes. These automated interactions include relaxing constraints, adjusting parameters, and performing belief revisions to produce new more satisfying descriptions. The decision-making process carried out by <strong>Aleph</strong> during these interactions is guided not only by business-defined strategies and policies but also by learned insights. Aleph implementation uses algorithms based on Market-Basket Analysis, Shapley values, and Behavioral Economic principles. The amalgamation of abstraction and potency offered by <strong>Aleph</strong> represents a remarkable leap forward for automating programming tasks and the reliable automation of human-machine interactions.</p><p>With artificial entities possessing beliefs, behaviors, and personalized rational or biased adaptation capabilities, it is exciting how natural it is to model problems using <strong><em>agents</em></strong>. These <em>agents</em> represent various problem actors or entities representing clients, knowledge workers, decision-makers, physical entities, or organizations. <strong>Singularities</strong> has developed methods to learn and represent behaviors and decision patterns suitable for analysis, explanation, and, more interestingly, for dynamic adaptation within a dynamic flow of interactions.</p><h3>A Brief History</h3><p>The term <strong>Aleph</strong> was coined in 2012 to name the efforts of creating the logic system and platform described in this document and was inspired by Borges’s ‘<em>Aleph</em>’ book. Singularities was founded in 2014 in California to bring <strong>Aleph’s</strong> application to the market. The beginnings of <strong>Aleph</strong> can be traced back to the 1980s and 1990s when the modal logic Z was proposed for knowledge representation and reasoning around the frame problem<a href="#_ftn9">[9]</a>,<a href="#_ftn10">[10]</a>. The last ten years of research in <strong>Aleph</strong> done by Singularities significantly improved the logic theory and semantics and delivered a scalable implementation to target business and engineering applications.</p><p><a href="#_ftnref1">[1]</a> <a href="https://plato.stanford.edu/entries/logic-belief-revision/">https://plato.stanford.edu/entries/logic-belief-revision/</a></p><p><a href="#_ftnref2">[2]</a> David Premack proposed that consciousness emerges when an individual is not only aware of others having thoughts and intentions (theory of mind) but also when they are aware of their own mental states and thoughts.</p><p><a href="#_ftnref3">[3]</a> Alfred Tarski’s Syntactic Approach.</p><p><a href="#_ftnref4">[4]</a> Kurt Gödel’s Incompleteness Theorem.</p><p><a href="#_ftnref5">[5]</a> <a href="https://people.eng.unimelb.edu.au/pstuckey/PPDP2013.pdf.">https://people.eng.unimelb.edu.au/pstuckey/PPDP2013.pdf.</a></p><p><a href="#_ftnref6">[6]</a> <a href="https://openai.com/research/improving-mathematical-reasoning-with-process-supervision">https://openai.com/research/improving-mathematical-reasoning-with-process-supervision</a>.</p><p><a href="#_ftnref7">[7]</a> <a href="https://mathworld.wolfram.com/ComputationalIrreducibility.html.">https://mathworld.wolfram.com/ComputationalIrreducibility.html.</a></p><p><a href="#_ftnref8">[8]</a> Proposed by German mathematician and logician Gottlob Frege.</p><p><a href="#_ftnref9">[9]</a> Frank Brown, ed. Proceedings of Workshop on The Frame Problem in AI. Elsevier, 1987.</p><p><a href="#_ftnref10">[10]</a> Carlos Araya. “On the Knowledge Representation Capabilities of a Modal Logic”. In: Proceedings of 1994 Florida AI Research Symposium, 1994.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=140b2a4c9f3a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>