The notebook

Notes

Field notes from real jobs. Field-tested means it shipped and survived. On the bench means I'm still collecting evidence — published anyway, because a real workbench has unfinished work on it.

12 Jun 2026 Field-tested

The Human Brain Was Not Built for the Infinite Feed

John's field note on information overload, family, trust, local AI, and why private AI systems are not about hiding — they are about keeping human agency alive.

#ai-agents #local-ai #information-overload #trust #private-ai #human-systems

balance is architecture
11 Jun 2026 On the bench

Body-Adjacent AI Needs a Kill Switch on Our Side

If AI is moving into glasses, cars, tools, workshops and eventually body-adjacent interfaces, the off switch cannot live in the cloud.

#agents #safety #local-ai #bci #human-in-the-loop

human in the loop is not enough if the loop belongs to somebody else
11 Jun 2026 Field-tested

Fable 5 Feels Like Two Models in One

An operator's observation from tool-heavy sessions with Claude Fable 5: it behaves less like a single plain model and more like a front controller in front of a reasoning engine — and that pattern is exactly how we should be building our own agent systems.

#claude #fable-5 #agents #model-routing #architecture

written while the model in question was deploying this site
11 Jun 2026 Field-tested

Why I Rebuilt My AI Agent’s Memory Instead of Blaming the Model

John explains why messy long-term AI agent memory caused false positives, bad routing and unreliable behaviour — and how cleaning memory made Hermes Agent safer and more useful.

#hermes-agent #ai-agents #memory #context-engineering #model-routing

memory is infrastructure
10 Jun 2026 On the bench

Why Prompting Is Becoming Spec Writing

Chat prompts are questions. Agent prompts are work orders: scope, authorization, boundaries, acceptance criteria. The skill that matters now is the one engineers always hated doing.

#agents #prompting #spec-writing

half the brief is what NOT to touch
09 Jun 2026 On the bench

The Real Problem With AI Agents Is Context Selection

Agents don't usually fail because the model is too small. They fail because the wrong slice of reality was in the window when the decision got made.

#agents #context #retrieval #architecture

select, don't shovel
08 Jun 2026 On the bench

What Auto Diagnostics Taught Me About AI Debugging

Twenty years of fault-finding on vehicles maps almost one-to-one onto debugging AI systems: observe, hypothesise, test one variable, verify the fix — and never trust the fault code.

#diagnostics #debugging #methodology

the fault code is a symptom, not a diagnosis
07 Jun 2026 On the bench

Building Local AI Like a Workshop, Not a Toy

Eight 3090s, a Proxmox host, and a rule: every machine earns its keep. Local AI infrastructure built like workshop equipment — bought for jobs, not for benchmarks.

#local-ai #infrastructure #gpu #self-hosted

every machine earns its keep
06 Jun 2026 On the bench

Why Useful Agents Need Verifiers, Not Just Bigger Models

The cheapest reliability upgrade in any agent system isn't a smarter model — it's a second pass that checks the first one's work before it leaves the building.

#agents #verification #reliability #architecture

MOT test for model output