The Journal
Essays
Long-form writing on AI, philosophy, psychology, and systems thinking.
-
Responsible AI Is a Cross-Functional Process
Implementing responsible AI as a cross-functional process reduced bias incidents by 48% and shifted detection from 6 weeks post-deployment to 3 days pre-deployment.
-
Message Queue Selection Is a Personality Test for Your Architecture
Message queue selection for 13 organizations revealed that the choice between Kafka, RabbitMQ, SQS, and NATS exposes deeper assumptions about consistency, throughput, and operational philosophy.
-
Data Retention Policies Are Architecture Decisions
Automated data retention reduced cloud storage costs by $18,000 per month and eliminated 4.2TB of unjustified data. Retention policies are architecture decisions, not compliance paperwork.
-
The Demarcation Problem in Data Science: When Is It Science
Popper's demarcation criterion asks if claims are falsifiable. When 68% of data science models are never evaluated against predictions, the field faces the question: is this science or storytelling with numbers?
-
Bayesian Reasoning as Engineering Philosophy of Judgment
Bayesian reasoning is the discipline of holding beliefs as probabilities and updating as evidence arrives. Google teams using explicit probabilistic reasoning experienced 34% fewer costly reversals.
-
Quietism vs. engagement: Two philosophical responses to technological acceleration
-
The Automation of Judgment
-
Your Data Catalog Is Lying to You
An audit of 3 enterprise data catalogs found that 38% of table descriptions were inaccurate and 22% of documented columns no longer existed. Catalogs create false confidence.
-
Runbooks Are the Most Undervalued Documentation
Incidents with current runbooks resolved 62% faster, with median resolution of 23 minutes versus 61 minutes without. Runbooks matter most when cognition is most impaired.
-
Complexity Budgets: Treating System Complexity as a Finite Resource
Teams that allocated explicit complexity budgets reduced accidental complexity by 38% over 12 months, measured by glue code, integration adapters, and operational runbook length.