Artilect
Long-form writing on software practice, engineering culture, and technical leadership.
Something new arrived with large language models: information crossed a threshold and became directly addressable through language. What we built isn’t a smarter agent. It’s a map that speaks—and understanding that changes everything about how you use it.
LLMs are not databases with a cute front-end. The “agentic revolution” is downstream—language is upstream. The retrieval-and-reasoning muddle is intrinsic to the medium, and papering it over with RAG misses the point. On what the interface actually is, and what that means for how we build.
A hope keeps returning in AI circles: could we build a language interface without the corpus? On what children, aphasia, sign languages, and small language models all say about that hope—and the ablation that would settle it. Essay III in the map-and-territory series.
We’re more capable than we’ve ever been, and we’ve never felt less certain of who we are. On finding ground in the AI era by returning to where developer culture came from — the Jargon File, Carmack’s .plan files, and the anti-magic instinct that predates the hype.
The failure modes we’re debugging in language models are the same ones we’ve been living with in ourselves for centuries. Hallucination is confabulation. Sycophancy is people-pleasing. The engineering problems of AI alignment are turning out to be a formalized language for problems we’ve never quite been able to articulate about being human.
AI was sold as an enterprise product. What it actually did was hand the workforce the ability to become their own CEOs. A look at the structural forces creating a new gig economy — not of precarity, but of leverage.
Professional orchestras have solved the predictability-at-scale problem. Their rehearsal structure — individual practice, section rehearsal, full orchestra, dress rehearsal — is not overhead. It is the mechanism by which reliable performance is possible. Software teams, especially in the agentic era, have the same problem and largely the same solution.
What well-meaning buyers need to know before commissioning custom software: why fixed-price contracts produce the failures they're meant to prevent, what a shared-risk engagement actually feels like, and the one question that predicts how a project will go.
The thesis that being happy, relaxed, and authentic produces better code and better teams — and why the AI era makes this argument more urgent, not less. With an organizational appendix documenting what specific companies have built.
What the most serious thinkers on software practice, systems dynamics, management, and coordination theory converge on — and what that convergence means for leading engineering teams in an era of AI-generated code.