hemju — calm notes on software and AI

Anthropic says Claude writes most of their shipped code and CI jobs jumped ~25x.

DHH put the pencil down at Rails World.

AI support, product agents, and localization drink from the same help center.

AI CEOs warn about the downside while labs keep racing.

Silence looks like a win on the AI support dashboard.

A daily-updating GitHub measurement shows AI-associated PR-description style went from 0.7% to 39% of a sampled corpus in under two years.

JEP 533's seventh preview in JDK 27 isn't feature indecision.

Splitting a codebase into more services doesn't reduce what a coding agent needs to know — it hides it behind an unenforced network contract.

Why more code reviewers rarely fix delivery: measure approved-but-unreleased inventory, batch size, approval delay, and rollback before adding headcount.

AI speeds development, but solo SaaS fails on coverage.

AWS's resourceless InvokeGuardrailChecks API moves guardrail policy into code.

A sign-flipping AI productivity study cannot justify higher velocity targets; CTOs should measure review pressure, failures, and recovery instead.

Splitting deadline accountability from architecture authority can surface costly tradeoffs before they harden into hidden technical debt.

Why OpenTelemetry's GenAI conventions are an unstable upstream dependency, and how to pin, normalize, and version-test them so a renamed attribute can't quietly approve a...

Aurora DSQL's managed CDC removes the connector, not the outbox.

Why production RAG failures are almost always cache-invalidation and versioning problems—not embedding choices—and how to design chunk identity, tombstones, and staleness windows.

Why bounded contexts rot into distributed monoliths: the boundary that matters is data ownership and change cadence, not the org chart you quietly drew it around.

At production scale, localization breaks as stale reads across caches, not bad translations.

Stop hardcoding prompts.

In 2026, 'Vibe Coding' is out.

As AI models proliferate, companies need governance beyond code.

Stop designing prompts; start designing agents.

In the era of the EU AI Act and DORA, 'Sovereign AI' is no longer optional.

Stop choosing sides.

AI readiness isn't a technical checklist; it's an organizational stress test.

Vector databases are the cargo cult of 2026.

The CTO role has shifted from managing code quality to architecting inference.

LLMs speak 100+ languages but understand zero cultures.

Modern systems are hybrids of deterministic databases and probabilistic LLMs.

Stop wasting AI budgets on flashy chatbots.

A pragmatic framework for CTOs and founders to evaluate AI initiatives.

Stop treating LLMs like deterministic utilities.

As AI moves into the core of the stack, determinism is fading.