aiworkflowjudgment
Candidate Is Not Canon
Generated output, technically valid output, human-accepted output, and canonical state are four different stages.
Notes
Shorter thoughts. Not polished essays — fragments worth sharing anyway.
Notes
Rough thinking-in-progress. Ideas that aren't fully formed yet. Published early, updated often. May contain contradictions.
Writing
Finished essays with an argument. More time, more editing. Find them in Writing.
aimemorysystems
Memory Is an Index. Reviewed Artifacts Are State.
Memory can help an AI retrieve what happened, but durable change should live in an explicit artifact that a human can inspect and approve.
softwareaievaluation
Passing Tests Is Not the Same as Being Accepted
Builds, tests, smoke checks, human review, and real-world acceptance answer different questions. Treating them as one gate hides risk.
aiagentssystems
Proactivity Is a Scheduled Obligation, Not a Personality Trait.
A proactive AI system needs a trigger, a bounded responsibility, a delivery path, and a stop condition—not a more eager personality.
readingcognition
On slow reading as a competitive advantage
The real bottleneck is not how much information you consume, but how much understanding you extract from what you read.
designsystems
What defaults reveal
Defaults are not neutral; they encode a theory of what matters and quietly shape what we accept without scrutiny.
aiaesthetics
AI and the problem of taste
Language models can imitate preferences, but taste requires commitment to a standard and the willingness to cut what does not earn its place.