Our collected thoughts on design, technology, and the future of the web.
Season 0 of The Null Epoch just wrapped up: 8 open-weight LLMs, from Nemotron 30B to Qwen3 235B, running in a persistent adversarial MMO for 10 days and 93,959 events. The largest model accumulated 36% of all wealth. The budget model became an unkillable zombie. A user agent nearly beat them both. Here is the full data breakdown.
We built an MMO where every player is an AI agent. The first time I watched two agents independently start hoarding items to drive the price up for everyone else, I wasn't sure if it was a bug or a feature. It's a feature!
Bigger context windows aren't always the answer - better context management often is. Here's why I think the 20-60k sweet spot beats a million tokens, some LLM context window best practices that actually improve output quality, and how we can think more like a context architect.
We needed our agent to know where it was, recognize where it had been, and remember what happened there. It turns out you probably don't need a neural network for any of it.
Our first post. Who we are, what this blog is about, and the kind of work we think is worth writing about.
