Learning how we learn

In 1995 a friend of mine entered my dorm in College Station, Texas with a twenty-pound grey laptop and the future in his mind. Within an hour, he downloaded a web page - the first web page I had ever seen. I honestly have no idea what the page was, but I doubt it was CNBC.

The next ten years were exciting times to be starting a career while interested in technology. Looking back on the early stages of connectivity, I now understand we were not only creating a vocabulary - but we were also building a paradigm.

I'm sure you see where this is going, but it's worth recognizing how formative and daunting the next few years will be in defining the paradigm of connected artificial intelligence.

Life moves pretty fast. If you don't stop and look around once in a while, you could miss it. -- Ferris Bueller

Over the past few years, I have intentionally observed how my mind attempts to understand AI, including:

• - The What: Foundational Technology

• - The How: Conceptualization and Implementation

• - The Why: Value and Threat

To be clear, I'm referring to how I learn about AI - not what I learn about AI. To me this has been more interesting and intrinsically more valuable than a practitioner's guide to building an agentic workforce.

And here's the interesting confluence of process and practice with AI:

How we think, is the product we're building.

Watching, watching

To better understand "The What: Foundational Technology" I spend a lot of time wandering down trails of thought with destinations unknown. Not surprisingly, I end up somewhere, and most often landing on a concept I was not intending to visit.

This week I paused when I stumbled on an article about a concept known as The Staircase Paradox or Paradox of Approximation.

"Hmm, why does this sound familiar?"

Without diving headfirst into the Pythagorean Theorem, let's just agree that the shortest path between two points is a straight line. Any adjustments to that path from point A to point C will become longer if the line is not perfectly straight.

I've never felt a necessary or appropriate time to write this...but "I'm not a mathematician".

With that said, approximation seems like it would play an important role in probabilistic theory - and here we are back at AI.

This is not about math. This is about finding a different perspective for understanding how AI models provide answers.

The Staircase

A staircase can have two steps or two billion. As the steps shrink, from a set viewpoint, the staircase looks more and more like a smooth line; the space between the steps disappears.

But its length never becomes a "line". It remains a staircase: separate horizontal and vertical moves.

The diagonal is not an infinitely fine staircase. It is a genuinely different geometric object.

That feels relevant to AI.

More context, tokens, agents, retrieval, and verification can make an answer increasingly resemble the truth. But approximation is not determinism. Better inference does not, by itself, create a validated relationship to reality.

I chose to build LocalCode (and thereby Lumberg.ai) because that distinction is consequential: not what will actually get built - politics, discretion, and negotiation are real, but:

- Given these property facts, what does the written code say can be built, and can we prove the path from question to controlling authority?

That is not making the staircase smaller. It is building the diagonal.

LocalCode does not make “what can actually be built” deterministic. It makes the bounded question of what the written code authorizes traceable, testable, and source verifiable.