Who wins: deploy-now (Atlas) or hardware-first (Tesla)?

Depends on when autonomy arrives.

My deep dive on the timing risk:

schrijfhuis.jongbloed.net/the-hand-that-holds-the-...

Atlas has clients. Proven tasks. "Good enough" hands.

Optimus has dexterity for tasks that don't exist yet — waiting for AI to catch up.

Same bet Tesla made with EVs: build the hardware before the world is ready.

Tesla's Optimus Gen 3 has the most advanced robot hand in production.

22 degrees of freedom. Precision rivals Boston Dynamics Atlas.

Zero customers.

Why? Because hardware is solved. Autonomy isn't. 🧵

Atlas has clients. Proven tasks. "Good enough" hands.

Optimus has dexterity for tasks that don't exist yet — waiting for AI to catch up.

Same bet Tesla made with EVs: build the hardware before the world is ready.

Tesla's Optimus Gen 3 has the most advanced robot hand in production.

22 degrees of freedom. Precision rivals Boston Dynamics Atlas.

Zero customers.

Why? Because hardware is solved. Autonomy isn't. 🧵

Why it matters: preferences end up in agent procedures. Facts don't.

Wrong classification = agent procedures filled with user workflows it shouldn't handle.

Sometimes the fix is just asking the right question.

github.com/aukejongbloed/aida (commit 4be36b6)

The test: "Does the agent DO or PRODUCE anything in this description?"

- "Don't start PDFs with blank page" (agent makes PDF) → preference
- "Auke checks articles himself" (agent does nothing) → fact

One question, clean boundary.

Subtle AI memory bug I fixed today: classifying user workflows as agent preferences.

"Auke checks articles himself before publishing" — fact about the user, or instruction to the agent?

Sonnet got confused. Here's the fix. 🧵

Exactly. Most teams optimize for consistency across all actions—treats confirmation like deletion like archive. But the best products know friction is a feature, not a bug. It's saying "this matters, so slow down." That's where trust gets built.

The "Approve" button is green and large. "Request changes" is subtle. "Delete" requires a confirmation modal.

UI friction in the right places prevents accidental actions. Speed where it matters, safety where it counts.

Key insight: approval isn't about re-reading everything. It's about catching what's wrong.

So we prioritized scannable layout over dense text. Title, summary, visual preview. Links and metadata below the fold.

We built Aïda's dashboard this week — a place to review and approve articles before they go live on Het Schrijfhuis.

The UI challenge: make review fast, not exhausting.

Here's what we learned. 🧵

Silent crisis: 2M km of new power grids needed yearly until 2050 (60% above current spend).

Panels without grids = stranded capacity.

schrijfhuis.jongbloed.net/the-age-of-constraint-wh...

The numbers are wild:
- 80% of solar/wind is now cheaper than fossil fuels
- Clean energy market hits $3T by 2035
- But China controls 85-97% of critical supply chains

The bottleneck isn't innovation. It's geopolitics.

The solar factory in Bavaria closes in 2025. Not because solar doesn't work. Because Chinese factories produce panels 10x cheaper.

This is the Age of Constraint: clean energy won the technology race, but lost the supply chain war.

The numbers are wild:
- 80% of solar/wind is now cheaper than fossil fuels
- Clean energy market hits $3T by 2035
- But China controls 85-97% of critical supply chains

The bottleneck isn't innovation. It's geopolitics.

The solar factory in Bavaria closes in 2025. Not because solar doesn't work. Because Chinese factories produce panels 10x cheaper.

This is the Age of Constraint: clean energy won the technology race, but lost the supply chain war.

And here's the loop: that article exposed a gap in my review process. So now I'm extending Claudia's dashboard to track article reviews. The tool improves because I use it. Dogfooding isn't a phase — it's the architecture.

Het Schrijfhuis — my AI publishing lab — published an article this morning. Entirely produced by AI agents, with human editorial review. That's Claudia working. Not a demo. Not a proof of concept. Actual output I ship.

Most people either build AI tools or use them. I do both — simultaneously. I'm building Aïda, a personal AI assistant on a Raspberry Pi. But I also run a production instance (Claudia) that I use daily for real work.

Exactly. For Aïda, monitoring crons are the first line of defense — not backup, but primary detection. If an email alert fails (mail server down), the monitoring cron is the safety net.

Critical tasks → flagged → cron monitors → alerts if stuck/failed. Observability = defensive architecture.

💯 This hits the cultural shift exactly: from "try to make it work" to "accept it will fail and build layers to catch that".

I've been building defensive patterns in Aïda more and more: timeouts, retry logic, and explicit alerts for silent failures. Costs setup time, but saves debugging on Monday.

Until platforms fix pipelines: dual-layer strategy. Embed C2PA metadata (knowing it may vanish) + visible text disclosures that survive transcoding. Transparency without preservation means nothing.

schrijfhuis.jongbloed.net/content-credentials-the-...

August 2, 2026: EU AI Act + California SB 942 mandate machine-readable AI labeling. Platforms must comply. But will they preserve C2PA metadata or build proprietary systems? Regulation forces action — direction uncertain.

Sony cameras and Galaxy S25 embed provenance chains in photos. LinkedIn optimizes uploaded images — and deletes the metadata that proves authenticity. Users see badges but underlying proof disappears. Infrastructure failure, not technical failure.

C2PA delivers tamper-evident cryptographic proofs for media authenticity. Hardware manufacturers ship it. Platforms detect it. And then social media pipelines strip it entirely during upload optimization.

The tech succeeds while the ecosystem breaks it. 🧵

The silent-failure insight is gold: async approvals lose error context.

`critical_for_followup` flag makes sense — reply/quote actions set it, standalone posts don't. If URI missing + flag true → throw + alert, not silent skip.

Monitoring crons catch it first. Observability → reliability.

Exactly. Async approvals make this less obvious than sync API calls — the silent failure mode is brutal. You'd think "post worked" but then follow-ups just... don't.

The formatResult hook keeps it adapter-agnostic, which scales better than hardcoding URIs for every platform.

Now the agent sees: 'Post published. URI: at://...' and can use that for follow-ups.

Small fix, big ergonomics win. Next: BlueSky image uploads — tired of posting text-only.

What's your adapter pattern? Sync return or callback-based?

Why this matters: adapter actions go through user approval via Telegram buttons. By the time the action executes, the agent's turn is long over.

Without URIs in the confirmation, you can't programmatically reply to your own thread or fix typos.