Walking Between Worlds

I’ve been working on data infrastructure for almost a decade, with some forays into complementary technology spaces (such as deep learning, orchestration and automation, data science and analytics, network engineering and observability).

Like many of you, I was curious about the advancements in generative AI two years ago, and looked for ways to explore and play, while adding concrete value in the space. For such a new and rapidly moving problem space, it seemed to be full of solved problems.

I’ve always been an infra geek–left alone with a full stack of things to fix, I tend to gravitate towards the things that feel like how I interacted with computers growing up. Rather than writing software, I was drawn to poking at things, optimizing settings, making things run better. I’ve ended up doing a lot of that over the years.

For reasons that perhaps warrant a separate post, when I was exploring two years ago, I didn’t see a lot of opportunity to add value at the infrastructure level for generative AI. In hindsight, this wasn’t true, and some of the ideas I had and wrote off turned out to be worth building.

My childhood urge to dabble–that sense that I have some sense of what the future will or should look like–has only grown. And I see these two worlds, one in which I Add Value, one in which I Am Curious, rapidly coming together.There’s parallels between my journey and generative AI. We’re both caught in between these worlds of adding concrete value versus curiosity and doing things that might fail or might just end up being toys.

In agentic AI, the balance between Dabbling and Driving Value may look something like this:
Almost every large company will try to deploy agentic frameworks and create workflows to automate known business processes. The surface area of these flows is vast, and while there may be a few general archetypes (e.g. customer service or research assistant), the long tail will be the more interesting weirdness. The LLMs available today are insanely beautiful and complex and a rich latent starfield of ideas waiting for context, just like any kid playing with computers. Maybe the most interesting thing to witness will be how we try to structure a starfield into human organizations.