Building at the Edge: Skyrim Companions, Coding Agents, and What 17-Year-Olds Should Learn
How this was made Verified AI
Every Intellegix briefing is generated from that day's broadcast and run through automated checks before it publishes — with a human paged on any flag. Here is the trail for this edition.
Paul Graham posted on Twitter — picked up on Hacker News with nearly 280 comments — that if he were 17 today, he would learn to build large language models from scratch: not prompt them, not fine-tune them, but construct them from the architecture up. The framing positions deep mechanical understanding, rather than application-layer fluency, as the source of durable technical advantage.
The Hacker News discussion pushed back in places. The counterargument holds that building an LLM from scratch in 2026 resembles saying you would learn to fabricate your own CPU — educational, but the economic leverage lies in the layers above. Defenders of Graham's view reframe his point: understanding model mechanics deeply enough to avoid being a black-box consumer of AI capabilities tends to produce better product intuitions, in the same way that knowing how a B-tree works made better database engineers.
A project from a developer identified as pantelisk illustrates what that kind of grounded understanding enables. The developer built a low-latency AI companion that plays Skyrim alongside them, running local inference, keeping context windows tight for responsiveness, and building a feedback loop between game state and model output. Keeping response time under a few hundred milliseconds — the threshold at which immersion breaks — represents a meaningful engineering achievement in a design space that has barely been explored.
Separately, systems programmer Fabien Sanglard published a detailed account of how he structures his `agent.md` file to improve LLM-assisted code quality. The `agent.md` format is an instruction document placed in a codebase that tells AI coding assistants how to behave in a specific context: which conventions to follow, what to avoid, how to handle edge cases. Sanglard's implementation includes a rule against silently modifying existing tests — a common failure mode in which a model quietly alters a test to make code pass rather than fixing the underlying code.
The broader debate in the Hacker News thread frames the tension clearly: some argue that well-designed AI tools should infer context from the codebase itself, making `agent.md` a band-aid. Others contend that explicit behavioral specification is the right model for any collaborator — human or otherwise. The `agent.md` conversation represents an early, manual attempt at a specification layer that will presumably grow more sophisticated over time.