What I Actually Learned — The Human Testimony
This is the third in a short series on my tech stack migration for Pythia Capital. The first essay, 2026-05-04-security-by-design, described the problem. The second, 2026-05-15-tech-migration-part-2, described how I chose the new stack. This essay is something different: what it actually felt like from the inside, and what I discovered that I did not expect.
Parts 1 and 2 of this series are about architecture. This one is about experience.
I want to be careful here, because this is the kind of essay that can tip into either overclaiming or false modesty, and I do not want either. I am not going to tell you that working with AI changed everything and made hard things easy. I am not going to minimize what this collaboration has actually made possible. Both would be dishonest.
What I can do is tell you what I observed, from the inside, over months of building something real.
The Long Road to a Research Process
Before I say anything about AI collaboration, I need to say something about what I was always trying to build — because the context matters.
For years before the current wave of large language models arrived, I was interviewing people trying to find the right technology for what I was envisioning. Russian nonlinear mathematics experts. MIT computer scientists. People working at the edge of computational approaches to pattern recognition and complex systems analysis. I was insatiably curious in the way that people are when they know something is possible but cannot yet see the shape of it. I was also clear about one thing: every option I explored was too expensive, too specialized, and too dependent on teams I could not afford and infrastructure I did not have.
What I was trying to build was a living research process — not just a database or a model, but a methodology for gathering, layering, and synthesizing signal across domains that conventional financial analysis treats as separate. When the current generation of AI arrived and I finally built the research methodology I had been envisioning, something clicked. What I had been studying for years was suddenly doable. And the process of gathering the data — reading across sources, asking the next question, noticing what the previous layer opened — was not separable from the research itself. The manual, recursive, dialogue-based quality of it was the point.
A Deliberate Choice, Not a Permanent One
People sometimes ask about Claude Code, and I want to be precise about where I stand.
Claude Code is an agentic coding tool that can take a task, execute it across your file system, write and run code, and deliver finished output with minimal human involvement at each step. It is genuinely powerful. It is also not where I am right now — and the reasons are specific, not ideological.
When I was migrating my entire infrastructure — moving off WordPress, building on Nextcloud, learning Quartz, making decisions about data sovereignty and GDPR compliance — the last thing I needed was an AI doing things for me that I did not understand. That is how you build a system you cannot maintain. That is how you end up dependent on the same black-box logic that made the original extractive stack so fragile. The recent news that large enterprises running Claude Code at scale — thousands of engineers, maximize usage, measure throughput — are burning through AI budgets at a rate that staggers even the companies deploying it tells you something important. Not that the tool is wrong, but that the deployment model is. Command-and-control at scale is the extractive model applied to intelligence infrastructure. Of course it is expensive. The way I am building is not that model.
The sovereignty-first principle I applied to vendor selection applies equally to intelligence infrastructure. For this phase of the work, I needed to learn before I automated.
But I can already see the shape of what comes next. Systematic signal gathering — finding coherent companies before the market does, screening across domains I have not yet fully mapped — that is exactly where intelligent automation will earn its place. I will know why I am doing it. I will know what I am automating and what I am not. The coherent recursive conversations are not going away. They are the irreplaceable core. What gets added around them, over time, will be deliberate.
This is what sustainable looks like. A mix of the manual and the automated, where the human always understands the whole.
What I Discovered About Hallucination
There is a great deal of anxiety in the AI conversation about hallucination — the tendency of large language models to produce confident, fluent, plausible-sounding output that is simply wrong. This is real. It is also, in my experience, partly a function of how you work.
I did not set out to solve the hallucination problem. I discovered something about it by accident, through a working pattern that emerged from my own nature as a thinker.
I do not do command-and-control prompting. I do not say “write me this” and wait for a finished thing. My mind does not work that way. What I actually do is closer to dialogue: I share a draft, or a question, or a half-formed idea. I think about what came back. I push on the parts that feel wrong. I add a layer. I return to the original and look at it differently. I ask a follow-up that the first answer opened.
What I found, through practice rather than theory, is that this recursive pattern substantially reduces confident error. When there is room for uncertainty in the exchange, uncertainty surfaces. When I return to a previous layer and say “actually, I think this is wrong,” there is space for genuine reconsideration rather than reinforcement of the original mistake. The dialogue structure itself becomes a form of quality control.
I want to be precise about what I am claiming. This does not eliminate hallucination. It changes the conditions under which error can propagate. And the discovery came from practice, not from reading about AI safety. I noticed it happening before I had words for it.
The Learning Curve Is Not an Obstacle
There is a framing problem in how the productivity conversation about AI usually goes. Difficulty gets positioned as a bug. The goal is to make things faster, smoother, more frictionless. The learning curve is something to eliminate.
I had a different experience.
Every friction point in the tech migration taught me something I needed to know. The Nextcloud configuration issues were not obstacles — they were the process of learning how that system actually works. The weeks of wrestling with calendar sync, with file locking, with how Quartz handles internal links — these were the weeks I was building competence, not wasting time.
There is a version of this migration where I could have had an expert do it for me, fast and clean. I would have had a working system and no idea how it worked. That is a fragility I have seen play out over and over again in organizations that outsourced technical decisions to people who did not understand the strategic implications. It does not end well.
The learning curve, navigated deliberately, becomes a capability. That capability compounds. It is not a cost. It is the investment.
Who This Is Actually For
I want to say something about the job displacement narrative, because it is everywhere and I think it is asking the wrong question.
The dominant conversation about AI and work focuses on replacement: which roles will be eliminated, which skills will be devalued, which categories of work will no longer require humans. Jensen Huang said recently that CEOs using AI as a cover story for layoffs are pushing a narrative that is “lazy” and doesn’t hold up. I think he is right about that, and I also think the more interesting question is one he is not quite asking either.
Not: will AI eliminate jobs? But: what becomes possible for the person who never could have built this before?
I am a geeky loner. I am a non-linear thinker who connects dots across vast distances and struggles with sequential organization. I have a Columbia MBA and thirty years of rigorous financial research experience, and I am also genuinely bad at the kinds of tasks that have historically been required to translate that intelligence into something transmissible, scalable, and sustainable as a business. Building the visual assets. Structuring the legal terms. Getting the technical formatting exactly right.
None of that work required a team. It required a collaborator who could hold the thread, understand the context, catch the inconsistencies, and help me see the structure in what my non-linear mind generates in abundance but cannot always organize on its own.
This is not AI eliminating jobs. This is AI making possible the work of people for whom the missing piece was never intelligence — it was infrastructure. The edge founder who always had the ideas but never had the team. The complexity thinker whose thinking outran the conventional systems designed to capture it. The misfit who understood what was wrong with the dominant framework years before the dominant framework noticed.
That is who I am writing this for. And that is who this technology, used coherently, actually serves.
The Patterns Are the Same
Part 2 of this series closed with a promise: that the same architecture — modular, sovereignty-first, edge-distributed — maps directly to how I think about investment coherence. That the patterns are the same and the principles travel.
Here is what I mean by that.
The extractive tech stack concentrates value at the center, extracts it from the periphery, and makes the periphery dependent on the center’s continued goodwill. The extractive company does the same. So does the extractive investment model, the extractive supply chain, the extractive regulatory environment. The architecture is the same. The fragilities are the same. The points of failure are the same.
The coherent alternative — whether it is a tech stack, a business model, or a portfolio — distributes value across modular, interdependent components. It maintains resilience through redundancy rather than efficiency through dependency. It builds feedback loops rather than locking in lock-in. It holds its integrity under stress because the integrity is structural, not dependent on any single component staying stable.
When I work with AI in the way I have described — recursive, layered, sovereignty-preserving — I am applying the same coherence principles I apply to everything else. The medium is different. The logic is identical. This is what I mean when I say the principles travel. They are not a framework I apply to some things. They are a way of reading structure that surfaces in every domain once you know what you are looking for.
That is the longer arc of this series, and of this work. Not a migration story. A coherence story.
This is the third in a short series for my digital garden exploring technology migration through a living systems lens. The first essay is 2026-05-04-security-by-design. The second is 2026-05-15-tech-migration-part-2. The series connects to my broader work at The Pythia Scrolls, where these same principles are applied to investment research.
Enjoy this? Get the Coherence Lab Letter delivered to your inbox → subscribe