Amir Kabir on Infrastructure Layer Concentration & Isaac Arnold on Durable Application Layer Companies

Orchestration systems can aggregate signals and generate interfaces, but they often struggle to ensure that context is complete, stable, and appropriately framed at the moment a decision is made. In high-variance environments, the challenge is less about producing outputs and more about helping users interpret uncertainty in a way that supports confident action. This exploration focuses on that “last mile”—reducing the time it takes for a user to understand what is happening, how reliable the signals are, and what trade-offs exist. Rather than trying to control variability directly, the goal is to make it legible and actionable within the decision itself. Current approaches tend to optimize for aggregation and automation, but can fall short when conditions are dynamic, multi-organizational, and partially observable. In those cases, small shifts in how information is surfaced—especially around confidence, drift, and competing signals—can have outsized impact on outcomes.

What if LLMs are not missing intelligence, but subjectivity? We’ve built systems that can reason, generate, and explain. But they still feel like tools, not partners. I’ve been exploring how to evolve LLMs into a cognitive partner — a system with continuity, memory, and controlled initiative. A rough roadmap: - Persistent identity (not session-based) - Structured memory (decisions, context, history) - Goal layer (clarity, consistency, completion) - Initiative (challenge, not just respond) - Orchestration (context-aware reasoning) - Reflection (learning from interaction) The real gap isn’t capability, it’s architecture. Curious who else is working in this direction. Let’s connect.

I joined Ontra five years ago believing PE operations were ready for technology-driven automation. In today’s fundraising market, that’s truer than ever: firms that adopt the right technology will see massive returns on that investment and outpace the competition. It’s time for PE firms to become operationally elite.

Most market decisions still rely on signals reconstructed after the fact. The problem isn’t technological. It’s structural: fragmented sources, aggregation latency, and reliance on probabilistic models. As long as the signal isn’t captured at the source, in real time, orchestration remains approximate. And at enterprise scale, approximations are costly.

Most market decisions still rely on signals reconstructed after the fact. The problem isn’t technological. It’s structural: fragmented sources, aggregation latency, and reliance on probabilistic models. As long as the signal isn’t captured at the source, in real time, orchestration remains approximate. And at enterprise scale, approximations are costly.

I'm neither frightened nor surprised by technological acceleration. What does frighten me is that, in parallel, consciousness and awareness are neither growing nor accelerating. There is a widening gap between our technical prowess and our non-technical maturity, with tools evolving at an exponential rate while our biological hardware and our social software seem stuck in a linear, or even stagnant, progression. Will we need some sort of "mental plugin" to fill this gap ?

Failure itself isn’t that interesting. What happens after it is. In modern systems, failure isn’t the end. It’s expected. Links drop. Nodes disappear. Data becomes incomplete. And still… the mission continues. So the real question is not: “Does it work?” But: 👉 what happens when it stops working as intended? • How does the system adapt? • What information is still usable? • What decisions are made under uncertainty? • How do operators respond? That’s where things break. Not at the moment of failure. But in what happens next. We don’t need better tests. We need a better understanding of how systems behave under loss.

Technology doesn’t enter a market on its own. What makes the difference isn’t the technology. It’s how it enters. Through who. With what context. At what stage of the relationship. Technology is evaluated after the people behind it are understood. And long before any engineering team formally signs off on it. That sequence is easy to miss. And expensive to get wrong.

A pattern I keep noticing in discussions about automated decision systems is how differently people think about failure. During development, the focus is usually on model performance, monitoring, and edge cases. Those are important. But they assume the system will be evaluated in real time. In practice, many of the most difficult questions appear much later. An outcome is challenged months after it happened. An incident triggers an internal review. A regulator asks how a specific decision was produced. At that point, organizations often discover something unexpected. Their systems can explain behavior in general. But they cannot reconstruct the specific event as it actually occurred. Logs exist. Model versions are recorded. Monitoring dashboards show system health. Yet the decision itself cannot be reproduced exactly as it existed at the moment it was made. The system worked. But the event is gone. That difference rarely matters during experimentation. It becomes critical once automated systems move into operational infrastructure.

It does not take super-human intelligence to leave us in the dark about the technology we depend on. We are the last generation that will be able to completely understand how the technology we use every day works. We are on the brink of a different type of singularity, not necessarily of super-intelligence, but of our inability to understand how critical systems work. We have generated code and designed silicon with software for decades, but we have always been able to understand how that worked because it followed our mental models. All it takes is for architectures and design patterns tailored to LLMs, but alien to humans, to emerge and we will stop being able to parse how hardware and software works. At some point agents will generate code for silicon that we don't know how to code for because it was not designed for humans to program. There is no pun.