Knowledge Infrastructure for LLMs
A language model can only work from what someone wrote down and put in front of it. Most companies that bought AI maintain their code and little else.
A language model can only work from what someone wrote down and put in front of it. Most companies that bought AI maintain their code and little else.
Red-teaming a model's behaviour is chaos engineering, and operations already wrote the rules for it.
A hook can block a word on every reply. Nothing I run can yet catch a model before it folds, only after.
Advice about language models depends on which model, on whose hardware, inside how much harness
A reading of the Goodhart Labs chess honeypot, three additions to the design, and a dated prediction.
Where Claude's writing style came from: text written to be heard, read in silence
Most agent harnesses ship without swap.
Everyone got a race car, but no track yet.
The modern world was built on explosions: contained, timed and measured ones. That is what an engine is. Language model output is the fire; the harness, the hook and the far-end gauge are the engine.
Error budgets, reconciliation loops, separation of duties and recovery over prevention: four operations practices that fit agent work almost line for line.
Nobody reads your documentation. Both ends are a model now, and the file between them is a wire format.
Seven months of running an agent-built pipeline unattended, 308 lines of recorded failures next to 208 lines of instructions.
A language model is a golem made of English. It behaves well only where it can see what its act will cost, and infrastructure is the craft of bringing that cost into view.
What happens when you say yes to the whole batch and figure it out as you go.
Everyone says "AI." Nobody means the same thing. Most of them are wrong.
Seven posts from 2025 on Substack, before this blog existed.