In my formative years working in audit and risk at EY, mostly within life sciences companies, I never once asked a client whether they trusted their ERP system. It would have been a strange question. Nobody trusts an ERP (people have feelings about their ERP, but trust is rarely one of them). What we asked was whether the system was controlled - or, if it was a GxP system, whether it was validated - it was the standard stuff: who had access to what, whether the audit trail was switched on, and what happened the last time something went wrong.
We've come to equate knowledge with wikis, and most wikis are wrong - or at least they will be eventually. That's not a knock on wikis, it's a knock on the assumption underneath them: that the stuff we wrote down is correct. This post is my attempt to argue for a different starting assumption (that it probably isn't) and to explore what a knowledge system looks like if you build it that way from the beginning.
Real experts mess up sometimes, admit it, and learn faster the next round. If someone never changes their mind, I assume they stopped paying attention.
Many teams still pitch the same idea: "Grab the latest base model, fine-tune it on our data, and boom—we have an expert." The second you do that, the model stops being current, and worse, it freezes your thinking. Knowledge keeps moving; your model should too.
We accept that high‑quality training demands curated, labeled datasets. Inference should follow the same rule. CxUs bring that discipline to runtime: the same clarity, labeling, and provenance we require for training data—applied to the context we feed the model when it answers.
Even though we've seen this train coming, seeing it in action still gives me pause - the moment something becomes "easy," it also becomes dangerous. Over the weekend I started playing around with Sora 2 from OpenAI, and let's just say — this thing is equal parts miracle and nightmare fuel.
The “cameo” feature alone is so good it’s unsettling. What used to take teams of visual-effects artists and GPU clusters now happens on a laptop — or worse, on your teenager’s iPhone. While you have to admit the tech is amazing - if we put a malicious bent on things, we have to consider that when the technical barrier drops, the attack surface explodes. And Sora 2 just dropped that barrier to the floor.
What seems like eons ago, back in March 2025 I wrote a probably-too-long paper about the nature of establishing relationships between humans over digital channels. Digital Identity Verification was an attempt at a framework to categorize how trust is formed between people — and then use that framework to systematically explain how bad actors manipulate that trust to perpetuate identity fraud.
I still think the paper is accurate, though I missed a critical point. I had positioned my argument around people establishing trust with other people via technology. What I overlooked is that the majority of technical approaches today focus on establishing trust between devices and/or software and, by extension, the people using them. These approaches often assume that the identity of the person becomes inherent to the device.
It is this critical gap — between human and technology — that we are looking to bridge with Vero. Building tools to facilitate real-time peer-to-peer authentication inherently creates friction in the process. But our friction is intentional: it’s a demonstration of trust, a clear signal of intent. And that signal is the smoke. In our case, if you see the smoke, you can be confident there is no fire.
The 2025 Context Engineering Survey, which reviewed more than 200 research papers and enterprise pilots, cautions: “Simply enlarging an LLM’s context window does not guarantee reliable attribution or auditability; we still need explanation systems, audit mechanisms, and governance structures.” (Context Engineering Survey 2025 §4). Put differently, the problem isn’t raw memory capacity — it’s the provenance of the information we cram inside. This is exactly the rationale we followed when designing Context Units for our Pyrana platform.
It's funny I've spent years talking about tokens in all of my blockchain work. Whether it be NFTs or stablecoins, it seemed like tokens were everywhere. Now that we're diving in deep and trying to fix some problems with context in AI, it seems the topic of tokens (albeit a very different kind) is still center stage. The phrase "tokenization of the world" has taken on a new meaning to me.
Enterprises plough well over $200 billion every year into data-warehouses, BI dashboards, and analytics suites,- and analysts now forecast the big-data analytics market to top $1.1 trillion by 2033¹. Financial-services firms alone spend $44 billion a year just on market-data feeds². Yet studies keep showing that roughly half of the information companies collect never gets used³.
