The Cognoria Manifesto
One of my most profound moments with generative AI came in 2024 when I decided to build a Chrome extension. My programming experience is limited to Python, a world away from the complexities of frontend development and browser APIs. Yet, I created an AI text-refinement plugin, WriteSharp [1], and published it—not for profit, but for the thrill of seeing if it could be done. The code is fully open-source [2], and with AI as my collaborator, the entire project was finished in under a week.
I used Claude 3.5, the state-of-the-art model for coding at the time, to write the software. My role was to direct it using prompting skills I had learned, like using Anthropic's prompt generator to create expert personas for the LLM. I would instruct the AI to act as a UI/UX designer, then a frontend developer, and so on. My work became conducting this virtual team: feeding artifacts between conversations, testing code, and describing bugs.
A similar "aha" moment happened during my MBA. I'd taken a marketing class where I spent hours on a complex pricing case, cross-checking my analysis with a colleague. A year later, a friend in the same class asked if I could try solving it with my ChatGPT Plus account. The model generated a perfect, one-shot answer in minutes, matching the conclusions that had taken me hours of work to reach.
These experiences were a visceral demonstration of a new superpower at our disposal. Realizing that was the spark that ignited my much deeper journey into generative AI.
The Search for Depth
As a lifelong student of entrepreneurship, I'm always looking for the next opportunity. When the GenAI wave hit in 2024, the RAG-based chatbot was the most obvious play. I jumped in, assembled a small team, and we started building Stomps.io [3], which we planned to differentiate with a one-time license fee. But as we moved from business plan to build-out, we hit a wall of unexpected depth. The development became a constant stream of questions for which we only had surface-level answers: What's the right chunk size? What's the principled reason for setting temperature to 0? We were guessing.
Looking back, the most unsettling part was our complete lack of an evaluation framework for a system that was, by its nature, probabilistic. We had built a product without any real oversight—a classic symptom of operating without a solid foundation. We hadn't just taken on a simple software project; we had stumbled into the dual disciplines of Information Retrieval and AI Engineering, and we were completely out of our depth.
At the same time I was wrestling with the practical challenges of Stomps, I was trying to get my theoretical bearings. My curated Twitter feed pointed unanimously to one paper as the source of the current progress: "Attention Is All You Need" [4]. Reading it was a humbling experience; it assumed a fluency in foundational concepts like Feed-Forward Networks and optimization theory that I simply didn't possess. This gap in my knowledge sent me down a rabbit hole, working through chapters of the "Speech and Language Processing" textbook [5], "Neural Networks and Deep Learning" [6], and lectures from Stanford's CS224n course [7].
My studies then led me to the field of Mechanistic Interpretability through resources like Transformer Circuits [8]. Here was a fascinating path to reverse-engineer models that didn't require massive compute clusters, making it seem accessible. But even this path had a non-negotiable prerequisite: a deep mastery of Linear Algebra. That was the final click. I realized the only way to make a real "dent" in this field, from the application layer to the research front, was to grasp the first principles—the mathematical and architectural truths that make everything else possible.
A Search for Peers
Grasping those first principles felt like the right path, but I knew it would be a slow and isolating climb to do alone. By serendipity, I came across an article by Eugene Yan [9] detailing his AI paper club—a simple, powerful model for collective learning. It was the blueprint I was looking for, and it became the impetus for Cognoria. My guiding principle was simple: I wanted to build a community where I was the dumbest person in the room. I was hungry for an environment of peers who would push my understanding.
We started in early 2025. I facilitated discussions on "Attention Is All You Need," a chapter from Chip Huyen's "AI Engineering" [10], and brought in a guest to teach agentic RAG. The people who joined were smart and curious, but I sensed a gap in the shared intensity of our commitment. With no knock on them, I was doing both the logistical and the intellectual heavy lifting, and eventually, the momentum stalled. The vision had been right, but the initial foundation wasn't. The lesson was profound: a community like this can't be built on casual interest. It must be built on a shared, relentless hunger for understanding.
The Trillion-Dollar Gap
Zooming out from my own bubble, I see a paradox: despite the hype, the real-world adoption of Generative AI is still in its infancy. I see it in my own city of Calgary, and this isn't a local quirk; it's a global phenomenon. A recent KPMG survey from mid-2025 [11] confirmed that the biggest barrier leaders face is a gap in talent and skills. Similarly, a 2024 report from BCG [12] cited a lack of AI literacy and tech talent as two of the top three challenges for executives.
While other barriers like cost and security exist, this human-sized skills gap is the most fundamental. It's the real bottleneck, but it's also the real opportunity. McKinsey estimates that GenAI could add trillions to the global GDP [13], and that value will be captured by the people who can bridge the gap between the technology's potential and its practical application. Mastering these systems from first principles opens up at least three distinct paths.
1. Join the Vanguard
The top AI labs and robotics companies continue to hire at a blistering pace. One path is to join them on the front lines, contributing directly to the core of AI development—working on new architectures, training regimes, safety, and alignment to actively push the boundaries of what's possible.
2. Become the Leader
Another path is to stay in your current industry and become the indispensable expert who leads its AI transformation from within. Imagine GenAI streamlining the creation of complex P&IDs in engineering firms (under human oversight, of course) or new foundation models in healthcare that can interpret pathology, neuroimaging, and clinical text. My conviction is that the true unicorns—those who combine deep industry context with a fundamental expertise in GenAI—are the ones who will build the future of their industries.
3. Build the Future
The third path is to become an entrepreneur who, armed with a fundamental understanding, can see and build products others can't. A compelling example is Vikash Singh, whose work I recently came across [14]. While learning the fundamentals, he realized the highest quality data was locked away in textbooks and PDFs. This insight led him to build Marker, an open-source framework for extracting data that he monetized under his company, Datalab. In just a year, it became a seven-figure business used by top AI labs like Anthropic. His journey is proof: going deep on the fundamentals gives you the vision to identify the real challenges that surface-level practitioners miss, and the skill to actually build the solution.
Ultimately, this is about more than just upskilling. It's about being in a position to solve one of the single biggest bottlenecks to our next wave of technological progress: the need for a real, fundamental understanding to make a meaningful contribution.
Our Shared Mission
Recognizing the persistent need for people who truly understand the fundamentals, I am relaunching Cognoria. My ambition is to gather a small group of founding members to go on this journey together. I believe deeply in the saying that you are the average of the five people you spend the most time with. While the epicenter of AI is in hubs like San Francisco, not all of us can pack our bags and move. So, my question is: why not create our own hub of excellence, remotely?
This will not be a casual reading group; it will be a shared launchpad. Our first mission is to build the rocket fuel together through a structured, collaborative journey of the definitive "Deep Learning" book by Goodfellow, Bengio, and Courville [15]. This foundational text will be our guide, and we'll supplement it with other essential resources along the way.
The commitment required isn't all of your time, but it is a genuine investment of your focus. The expectation is that everyone will wrestle with the material, contribute insights, and be willing to both teach and learn. We'll have short, focused sync discussions once a week (to be decided by the group), but most of our work will be asynchronous. Who knows, we might even discover how to use the AI we're studying to accelerate our own learning. Ultimately, this community will be built on that shared, relentless hunger that was missing before.
A tweet by Zach Wilson [16] breaks down AI engineering into four levels. Levels 1 and 2 involve using APIs and building with tools like RAG. Many resources can get you there. Cognoria exists for the ascent to Level 3 (engineering reliable systems) and Level 4 (optimizing AI at scale). By aiming this high, mastery of the earlier levels becomes a natural byproduct. This journey is for those compelled to move beyond the surface and grasp the first principles that drive everything.
If that sounds like you, it would be an honor to take this journey together.
References
[1] WriteSharp on Chrome Web Store
[2] WriteSharp Open-Source Code
[4] “Attention Is All You Need”
[5] “Speech and Language Processing” by Jurafsky & Martin
[6] “Neural Networks and Deep Learning”
[8] A Mathematical Framework for Transformer Circuits
[9] Eugene Yan’s article on AI paper clubs
[10] Chip Huyen’s “AI Engineering”
[11] KPMG — Generative AI Survey Report (June 2025)
[12] BCG — AI at Work in 2024: Friend and Foe
[13] McKinsey — The economic potential of generative AI