How I rebuilt the blog

I find myself using coding agents a lot. Not only to code, but also to get ideas out of my head. I studied math when I was in college, and a large part of that was studying algorithmic game theory and computational learning theory. I was looking at how AI can optimally play games or react in situations of uncertainty, situations with many decisions, and find ultimately optimal paths and trajectories through the action and decision space of these environments.

When you’re using coding agents, you are struck with the reality that they are attempting to choose an optimal response to your outputs. The goal right now of users using these systems is to optimally respond to their underlying intent behind their prompts. The user might want you to do X, Y, or Z, but really what they want is a system built around X, Y, and Z and an output that is optimal against their intent. The coding agents, however, don’t know this information. They don’t know the user’s priors or bias or taste.This is why ChatGPT and Anthropic, Claude-type platforms want to create a memory system that remembers you across your chats. If the system can understand who you are or what you like and other attributes about you and your interests, tastes, and preferences, then it can serve you an optimal or a better response from these agents.

I think a lot about how we want to find optimality when we use these systems and when we build these systems, and one of them is ultimately finding ways to get the ideas, the thoughts, out of my head and into this blog. I want to write about the things that are fascinating to me about this industry right now. Part of doing that is figuring out how to leverage agents to tell me what concepts I am getting at and to educate and explain these properly back to me. They can potentially even write blog posts about them so that other people can connect and learn or critique my views or techniques.

And so I rebuilt my blog and I built it in a way that leverages this loop, leverages this concept of optimality. I spend most of my days working on agentic infrastructure and auto research infrastructure. What that means is that I build environments for different types of agents to improve themselves by letting coding agents improve those agents through a rigorous feedback loop that measures relevant statistical metrics, optimizes prompts, and learns to improve itself against subjective and objective measures.The part that’s fascinating to me about this experiment of this blog is whether or not I can clearly get the ideas out of my head or out of my work environment, where I’m building this technology and playing with ideas around society, economics, computer science, math, philosophy, and any other subject that dares breach this thought barrier.And then whether I can optimally generate writing that is both educational and potentially useful for the futurethere might come a time, far or soon, where I can clone myself. I can leverage the information that both my agent generated from me and that I wrote myself to create a version or a fleet of agents that understand my taste and understand what I like and the types of abstractions that I enjoy thinking about and building.

So, down to the mechanics of this blog. I built this blog where a majority right now of these posts are generated by AI. The flow is to use a skill or directly tell Claude or Codex to mine my raw session data from talking with these coding agents and extract architectural software or mathematical concepts, philosophical ideas, things that I could be in the mood to explore about my weekly sessions with coding agents.It is then to extract the best ideas out of those that I like and write interesting posts that blend math, design, computer science, and whatever is topical with what I’m building at work or in my free time.

Every edit or pass over a blog after it’s written, I aim to track. I aim to show the actual revision history, both with the real trace data from these agents and eventually to the scores and subjective or objective metrics that I record for an auto research style optimization loop. I want there to be some recursive nature to this blog, that both posts continue to be updated, it’s fluid, and that these posts influence potential future posts in writing style, in UI components that are shared in concepts. Really anything that the agent derives is of interest here to try and understand and extract both the agent’s intent, my intent, and anything relevant for optimization of building a blog and writing your ideas to the world.Not necessarily a very objective thing, except if people engage with it.This is why eventually you might see the ability to engage with these posts, and I can then leverage that data to auto-improve the content, the ideas, the rigor, and anything people share.