It's been a few weeks since my last post, and I'm feeling the weight of the challenge I've undertaken. How do I build an intelligent AI system I can trust implicitly? Against the backdrop of this hobby project, I've found myself increasingly reliant on centralized LLM tools for work at Codex, especially on the coding front.

Re-reading my last post, I realize I was too optimistic and a bit naive. I greatly underestimated the complexity, both technically and practically. Yes, I nailed down the pillars (evaluation, persistent memory, and inference orchestration), but I glossed over the realities of actually getting to a working system. The closer I get to the problem, the more evident it becomes that a singular model won't cut it. Instead, I'm drawn toward an agent swarm approach…effectively multiple specialized models coordinated by an orchestrator that decides which model handles which task.

At Codex, we've spent months deploying a liquidity provisioning engine powered exclusively by LLM-generated code. This effort has helped really shine a light on current SOTA models' inability to keep overall project context in mind. They certainly excel at focused, isolated tasks but fall short at retaining the bigger picture. If you're not vigilant in providing constant context, they drift (confidently and spectacularly) into errors. Having spent over 15 years as a product manager, I feel oddly well-prepared for this reality. Most of my day now involves writing detailed prompts and specifications (effectively PRDs for LLMs). The parallels to managing human engineering teams are striking. So far, the fundamentals remain the same: clear communication, constant reinforcement of context, and relentless validation. My advice to anyone working with LLMs is simple: trust nothing, verify everything.

The unverifiability of closed-source LLMs only adds urgency to my sovereign AI mission. These models evolve behind closed doors, their weights and biases shifting with every update. Just look at Grok 3 this week. When you're relying on them for critical tasks, that kind of uncertainty is a deal-breaker.

One piece of good news: OpenAI does allow you to export your data, including chats and preferences. That’s huge because it means I can fine-tune an open-source model using my own interaction history, aligning future behavior more closely with how I think and work. This gives me a way to escape the data moat I’ve been digging with each prompt using centralized models.

Stepping back, evaluation still remains my biggest hurdle. Having read almost every Issac Asimov story in existence, I keep coming back to the Three Laws of Robotics. These were effectively system/model cards for each robot that kept them aligned with mankind. These laws could not be violated though they were somewhat open to interpretation. Eventually there was a Zeroth Law which put humanity’s survival above individual humans for example. When starting to evaluate an OS model I always focus on the system card to begin with as a way to suss out potential bias or challenges. It’s made me think about what my own version of those laws might look like. I’ve started sketching my own system spec...a kind of values-aligned system prompt or "core directive" that I can embed into whichever models I eventually select.

I’m not there yet. I’ll keep iterating on my evaluation tools while I wait for the next generation of open-source models to arrive.

Hopefully, I’ll be ready for them when they do.