Blog

If you are reading this blog, you shouldn't find it controversial when I say that obviously AI agents should be doing much of the low-level digital work at a company at this point. When you are building software, humans should be reviewing code and giving feedback, not writing it by hand. That ship has sailed.

So if an AI agent is doing the actual work, think about what happens in a typical meeting between humans. Let's say it's between you, a manager, and a direct report who is an engineer, talking about what needs to be done on a project. The engineer listens, internalizes it, and then has...

One of the hardest parts about building AI agents in your org is that you cannot rely on them to keep private information hidden across boundaries.

For example if one agent has access to both your DM's and a public channel on slack (or Zulip) and it's able to read/write to both places, then you face a real risk of your private messages getting leaked into the public channel. It could be by pure mistake of the model. Even with good prompting you just can't really guarantee it won't get confused.

This confusion problem is solvable by better models and better security frameworks. But even when...

Ever since Claude 4.5 Opus and other similarly powerful models came out at the end of 2025, it's old news that individuals can get 10x or more boosts in their output when they know how to leverage it.

Many draw the analogy to human computers in the 1960's and how a modern spreadsheet on a laptop is faster than entire skyscrapers of human computers. And imagine if one cell of that spreadsheet had to be manually computed by a human: it would bottleneck the entire sheet by orders of magnitude, waiting for the one step that needs human attention. That's how many agent-powered systems are today:...

As I run WindBorne, I often analogize the way I and the company operate to how a computer processor works. The principles of computing are the most correct way to think about just about anything, after all.

An important concept from computing is branch prediction, and you can't run an efficient information machine without it.

Branch prediction is something that a CPU does when it encounters a branch in the code. A branch is a point where execution could go one of two ways, like an if statement. For example:

No one is talking about the most important decision for leveraging AI across an org: the communication infrastructure that joins teams of humans and teams of AI.

At WindBorne Systems, AI agents are deeply embedded in how we work and communicate: they're all over our chat app, Zulip. Individual AI use doesn't scale to most org functions—you need one shared nexus for humans and bots to read context from. This isn't possible on closed-source apps like Slack, where you can't build custom integrations without getting rate limited or waiting on app-store approval.

I think the most important concept when building AI agents is bootstrapping. I use the term a lot so let me explain it a bit.

The expression itself comes from the phrase "pulling yourself up by your bootstraps"—a 19th century American expression describing an absurdly impossible act (literally lifting yourself off the ground by tugging on your own boot straps).

In computer software, it refers to the process of using a system to build or improve itself. For example, modern programming languages are bootstrapped: the Rust compiler is written in Rust.

How do AI weather models learn to predict weather? Are they actually learning physics?

This kind of question is called model interpretability. It's a famously hard problem, much harder than making a good model in the first place.

Let's take a look at WindBorne's WeatherMesh family of models, and try to peer into what's going on inside, with the question in mind: Is the model really learning physics? Or is it just pattern-matching?

The model has three main parts:

The Encoder encodes from physical weather variables into what's called a latent space that the model learns. The latent space is like...

Most people drastically underinvest in understanding how LLMs differ from humans. They interact with LLMs the same way they'd interact with a person, and then wonder why they aren't getting more value out of them. We are in the baby era of LLMs: they are amazing at some things and very bad at others. Communication style matters a lot.

Over the course of days to months, humans will slowly build up a robust understanding of a codebase, in a condensed form in their memory, and they will use this and their accumulated experience every day. If a document is too long, they may not read it, but it is...

It's the end of 2025, and a generally crazy time in the AI boom. I don't suspect things will get less crazy any time soon. I think there is too much big hype and big fear going on, which is valid, but I want to talk about the small, specific things that I'm optimistic about in an AI future, to look back on in 5 years and see if they came true.

I want all my personal data across apps and devices to be fed to one centralized location (ideally a server that I own), so that I can design (with AI-assisted software development) cool dashboard and displays to analyze my habits and behaviors so that...

The below is mostly writen by Claude 4.5 Opus, with a lot of direction from me, after a long conversation with him about my personal theories for consciousness and checking them against what is described in literature. I think this serves as a useful summary of existing literature, and some concepts proposed here are I haven't written about before, and I think is a direction people should go. I'm an engineer, not a philospher, so I don't spend much time reading all the different things others have wrote and indexing it in my head. But I do spend a lot of time thinking about this in the...

For fun, I have been playing with making population density visualization with three.js and the WorldPop dataset for global population at high resolution.

Here are some places that I've lived in recently:

Perhaps I will finish this viewer and post it in the future, and write a blog about it, but for now it's pretty jank and just a few hours of me and LLMs.

I picked the WorldPop dataset for this, as I was interested in animating population vs time. It turns out that doing so is not all that interesting. But, it led me to something else.