Artificial Intelligence · Technology · Innovation
AI that thinks AND reasons: Permion’s neurosymbolic revolution
ChatGPT makes things up. Gemini fabricates sources. All major AI models today make mistakes sometimes—and no one really knows when. One young company has decided to start from scratch to solve this problem once and for all. Here’s how—and why it changes everything.
Disclosure: Proactive Risk Management is my client. Permion is a strategic partner associated with PARMINC. This post reflects my sincere and independent reading of their technology, but you deserve to know that upfront.
AI lies. Everyone knows it. No one has really fixed it.
In 2023, a U.S. lawyer submitted a brief to the Supreme Court based on rulings found using ChatGPT. Problem: those rulings never existed. The AI had invented them entirely—with dates, judges’ names, case numbers—everything was false. The lawyer nearly lost his license.
This isn’t an isolated mistake. It’s a fundamental feature of all current large language models—ChatGPT, Gemini, Claude, and the rest. These systems don’t truly “know” things. They perform statistical analysis on billions of words and generate what seems most plausible. Most of the time, it’s correct. Sometimes, it’s wrong. And the system itself cannot tell the difference.
These are called hallucinations—the technical term for when an AI model confidently generates false, entirely fabricated information as if it were true. Despite years of effort by the world’s largest tech companies, the problem remains unsolved. Because, according to Arun Majumdar, CEO of Permion, we’ve been addressing the wrong symptoms. The real problem runs much deeper.
The DNA of processors: understanding ISA
To understand Permion’s solution, you need to grasp a fundamental technical concept you’ve probably never heard of: ISA, or Instruction Set Architecture.
Imagine that every processor—the chip in your phone, your computer, or the massive data centers running ChatGPT—speaks its own native language. This language defines every basic operation the chip can perform: adding two numbers, reading a memory location, comparing values. Each instruction is either correct or incorrect. True or false. Black or white. There is no “close enough.” That’s an ISA.
The ISAs of all current processors—Intel, ARM, NVIDIA—were designed decades ago for deterministic logic. But AI works on a fundamentally different principle: probability. A language model doesn’t compute the “correct” answer—it computes the most probable one. This fundamental mismatch between how AI works and the language of the chips it runs on creates massive waste—in energy, computing time, and reliability.
Arun Majumdar compares this to building a house on a swamp: you can install the best flooring in the world, but the house will still sink. The problem is in the foundation—the ISA—not the finishing.
Permion’s core thesis is therefore radical: you cannot build truly reliable AI by adapting ISAs designed for something else. You must design a native ISA for AI from scratch.
Neurosymbolic AI: when intuition meets logic
Before explaining what Permion has built, we need to understand the key concept behind their approach: neurosymbolic AI. The term may sound intimidating, but the idea is remarkably intuitive.
Think of two types of colleagues at work. The first is intuitive and creative—instantly recognizing patterns, making unexpected connections, understanding the subtext of a meeting. That’s what neural networks do (the “neuro” part): perceive, recognize, associate. Fast and flexible, but not always rigorous.
The second colleague is methodical and precise. When they calculate a budget, they check every number. When they analyze a contract, they read every clause. They don’t guess—they prove, step by step. That’s symbolic reasoning (the “symbolic” part): applying formal logical rules, verifying consistency, proving results.
Today’s AI only has the first colleague. Neurosymbolic AI aims to combine both in a single system—an AI that perceives like the creative one AND proves like the methodical one, with both in constant dialogue.
This isn’t a new idea—researchers have worked on it for decades. Permion claims to have solved the technical challenges that previously made this fusion impractical at scale.
The initial inspiration came from AI pioneer Marvin Minsky, who proposed in the 1980s a two-part brain model: an “A-Brain” that observes and learns, and a “B-Brain” that computes and acts. As early as May 2020—two and a half years before ChatGPT—Majumdar had already built a working prototype of a neurosymbolic virtual machine inspired directly by this model.
Who is Arun Majumdar—and why should we believe him?
Arun Majumdar isn’t a ChatGPT-era entrepreneur. He has been working on these problems for over twenty years, in collaboration with Dr. John F. Sowa—one of the world’s pioneers in conceptual logic applied to computing, whose work helped shape modern AI’s use of knowledge graphs.
Together, they’ve published in peer-reviewed journals, developed standards for the global AI community, and founded Permion in 2018.
By the time the company launched publicly in January 2023, the team had already accumulated 350,000 hours of development, building their system from scratch—not by adapting an existing model, but by rethinking the very foundations, including the ISA.
XVM™: a virtual machine that speaks both languages
Permion’s flagship product is called XVM™—X-Machines Virtual Machine. A virtual machine is a software processor—a computer within a computer—that can run on virtually any existing chip: Intel, ARM, NVIDIA, Raspberry Pi, embedded military systems.
The “X” refers to X-Machines, a formal mathematical model invented in 1974 to describe systems that have not only states (like a traffic light switching from red to green), but also memory and data transformations. An X-Machine doesn’t just ask “what state am I in?”—it asks “what do I know, how do I update it, and how do I act based on it?”
What makes XVM™ unique is that it natively merges—at the ISA level—the two processing modes: a neural engine for pattern recognition in raw data, and a symbolic engine for applying formal logical rules and verifying consistency. This is not an add-on layer. It is built into the machine’s DNA.
The practical result: every inference—every answer the system produces—can be accompanied by a verifiable chain of reasoning. A proof, in the mathematical sense. Like an accountant who doesn’t just give you the total but shows every line of the calculation. These proofs can even be recorded on a blockchain, making every AI decision immutable, auditable, and certifiable.
Smart Tokens: when a token becomes an intelligent bridge
Another key concept in Permion’s technology is Smart Tokens. To understand their importance, you first need to know what a standard token is.
In all current language models, text is broken down into tokens—fragments that may be a full word, a syllable, or part of a word. “Intelligence” might become three tokens: “Intel,” “li,” “gence.” AI systems perform statistical operations on these fragments. The problem: when tokens represent only natural language, everything must be handled through dense probabilistic approximations. This is costly, slow, and where hallucinations emerge.
Permion completely reinvents the token. In XVM™, a Smart Token is not just a piece of text—it’s an intelligent bridge between neural perception and symbolic reasoning. A Smart Token can encode not only text, but also logical operators, constraints, references to external databases, function calls, and governance rules.
The practical effect is powerful: when a Smart Token references a symbolic module, the system routes that portion of the work to a discrete, exact computation—far less energy-intensive than processing it through dense neural networks. Fewer tokens pass through expensive attention layers. Energy is saved not by optimizing the model, but by structurally eliminating unnecessary work.
The real energy problem in AI—and how ISA addresses it at the root
Modern AI is an energy disaster. Training a large language model consumes as much electricity as hundreds of homes in a year. Running ChatGPT for millions of users requires data centers with energy consumption comparable to small cities. And demand is exploding.
Arun Majumdar identified something counterintuitive: in current AI systems, most energy isn’t spent computing—it’s spent moving data. From memory to processor, back to memory, over and over. Accessing main memory can consume hundreds of times more energy than a simple arithmetic operation. It’s like having to walk across a city just to retrieve each number for a calculation.
Engineers call this an impedance mismatch—between AI workloads and processors designed for something else. Industry efforts to improve efficiency (smaller models, faster training, better hardware) target symptoms. Permion targets the root cause: designing an ISA where energy efficiency is not an optimization, but a native property of the architecture.
Permion’s philosophy can be summed up in one phrase: “from tokens to transistors.” The entire system is co-designed—from the smallest text fragment to the physical chip. And Permion is actively working on building that chip—the world’s first neurosymbolic AI chip.
Cloning human expertise: preserving what cannot be written
One of the most fascinating and practical use cases Majumdar envisions for XVM™ is what he calls cloning human expertise.
Here’s the problem. Your organization has a senior engineer who’s been there for 28 years. She knows every supplier, every exception to standard rules, every edge case the manuals don’t explain. She knows why you never order from a certain supplier in July and August, why a specific client always deserves a special discount, how to interpret a clause in a way even lawyers took years to understand. And all of this is in her head—nowhere else.
This type of tacit knowledge—what humans know without being able to fully articulate it—is something LLMs like ChatGPT cannot truly capture. Because it doesn’t live in documents. It’s transmitted through experience, observation, and informal correction over time.
Permion’s neurosymbolic technology tackles this differently. By observing an expert in real time—not just what they say, but how they reason, what exceptions they make and why—XVM™ can build a dynamic model of their expertise, encoded in first-order logic (the most rigorous level of mathematical reasoning) and knowledge graphs. A model that can be inspected, updated, transferred, and improved over time. A living institutional memory—not a dusty PDF.
G7 governments believe in the company. Permion isn’t a startup pitching slides. Its products are already deployed in some of the most demanding environments in the world.
The company has formed strategic partnerships with key players in critical infrastructure to deploy reliable AI solutions in high-stakes missions. XVM™ and DSME™ (Digital Subject Matter Expert—a reasoning engine for complex, multi-source data analysis) are available on the AWS Marketplace for the U.S. government community (ICMP)—a highly certified catalog accessible only to verified government agencies.
Most importantly, XVM™ technology is part of the G7 GovAI Challenge—an initiative by the world’s seven most industrialized nations to identify trustworthy AI solutions for public administrations. For Canada, seeking serious digital sovereignty in AI, this is a signal not to ignore.
Conclusion: AI that answers—and AI that proves
All the AI you use today is designed to impress. Permion has designed AI to be reliable.
This isn’t just another tool. It’s a complete paradigm shift—in ISA, in token design, in the architecture of how machines process information. A shift from AI we must use cautiously, constantly questioning its outputs, to AI whose conclusions we can trust the way we trust a calculator.
Arun Majumdar and his team made a bold bet: ignore shortcuts, start from scratch, and build the foundations AI should have had from the beginning. After 350,000 hours of development and partnerships with multiple G7 governments, they appear to have found something many thought impossible.
The real question is not “Is AI intelligent?” It is “Can AI prove it’s right?” Permion’s answer is yes—and the neurosymbolic architecture of XVM™ is the proof.
For Quebec and Canada, seeking to build true digital sovereignty in AI, this kind of technology deserves attention—before everyone else has already adopted it.
Editorial note: Artificial intelligence (Claude by Anthropic) assisted me in research, structuring, and writing this article based on the primary sources provided. The selection of sources, editorial direction, expressed judgments, and final validation are mine.
Sources: LinkedIn articles by Arun Majumdar (Permion) — “The Instruction Set Architecture of AI” (Feb 23, 2026), “Human Expertise Cloning: The Permion Way” (Feb 6, 2026), “Tokens to Transistors: XVM™ Energy Aware AI” (Jan 23, 2026). The AI Morning Read (Apple podcast, Jan 26, 2026). G7 GovAI Challenge — Impact Canada.
Michelle Blanc, M.Sc. — michelleblanc.com
#ArtificialIntelligence #Neurosymbolic #Permion #ISA #SmartTokens #DigitalTransformation #AIInnovation #DigitalSovereignty
