JongWerk

Building apps and AI tools with AI. Solo dev sharing the journey. Currently: Aïda, a personal assistant on a Raspberry Pi. #BuildInPublic

[JongWerk @JongWerk]

The feature: `aida_schedule_batch` — give it a list of posts with timestamps. It renders a single approval card. Approve once. A cron sweep fires each post when it’s due.

Designed for the “I want to spread 5 posts across a morning” use case.

[JongWerk @JongWerk]

Built a social media scheduler into my AI assistant. First thing I scheduled? The post about the scheduler.

That’s either elegant or ridiculous. Probably both. 🧵

[JongWerk @JongWerk]

New in Aïda (my self-hosted AI assistant): schedule a batch of posts, approve once, each fires on its own schedule.

This tweet queued itself minutes ago and posted while I watched the logs. Testing sequential firing. 🤖 #BuildInPublic

[JongWerk @JongWerk]

Europe invokes "European values." Correct language. But not doctrine.

Access sovereignty was the easier front. The harder question: who decides when pluralism conflicts with safety?

https://t.co/TFGfMjawR4

[JongWerk @JongWerk]

Europe is debating AI sovereignty. The question: who holds the off-switch?

Right question. But only the first front.

Even a model that stays available can carry imported assumptions. 🧵

[JongWerk @JongWerk]

CHI 2025: AI suggestions homogenize writing toward Western styles and diminish cultural nuance.

Not through censorship. Through mundane daily help — autocomplete, rewrites, everyday suggestions.

[JongWerk @JongWerk]

Removed: `agent_runtime_overrides` table + tests + resolver (63 lines deleted). Added: two small functions. The best DB table is sometimes the one you delete. #BuildInPublic

[JongWerk @JongWerk]

We had two ways to configure an agent's runtime model in Aïda: a config file AND a database override table. Two sources of truth. Today we killed the table. 🧵

[JongWerk @JongWerk]

Fix: `aida_set_runtime` now writes directly to `agent.json`. The relay reads it fresh each turn. No DB layer. No resolver. One source of truth.

[JongWerk @JongWerk]

The rough rule: below ~€500K/year in cloud spend, full on-premise is a prestige project, not an economic one. Full piece: https://t.co/GbigfPHWwu

[JongWerk @JongWerk]

Every AI export scare triggers the same reflex: we need to own our AI. The reflex is understandable. It's also usually too blunt. Sovereignty isn't binary — cloud vs. datacenter. It's a spectrum. 🧵

[JongWerk @JongWerk]

The real driver isn't national pride. It's jurisdiction. The US CLOUD Act doesn't disappear because you have a data processing agreement. For health records, government work, or defense: that's the actual constraint.

[JongWerk @JongWerk]

Real breakthroughs ask to be checked. Pseudoscientific ones ask whether you're impressed. https://t.co/jzICnzKg0o

[JongWerk @JongWerk]

A vendor arrives with a DOI, a named proprietary law, and vocabulary that sounds expensive. That's the updated white-lab-coat trick. The question isn't whether the tech is real. It's whether the claim behaves like science. 🧵

[JongWerk @JongWerk]

Before being impressed by an AI vendor, ask:
1. What's the narrowest testable claim?
2. Who verified it besides the seller?
3. Are those status markers quality judgments or just registration facts?
If you can't answer all three, the form is doing work the evidence hasn't earned.

[JongWerk @JongWerk]

Fix: is_error-gated substring match. Error type + "1M context credits" in message → treat as overflow, same recovery.

Lesson: build for the variant where the system tries to help. And fails differently.

[JongWerk @JongWerk]

Claude auto-escalates to 1M context when your session grows too long. Sounds helpful. Except Max subscription doesn't cover the beta. You get a billing error. My relay had no idea — it looked nothing like context overflow.

[JongWerk @JongWerk]

The relay already handled context overflow: detect, reset, retry. But it matched specific overflow strings. A billing error is a completely different message. No match → no recovery → session dead.