Von der Leyen wants a legal drinking age for social media

Good morning 👋 Von der Leyen wants social media to work like a driver's license: earned in stages, not handed out at 13. Recommendations from Brussels don't stay recommendations for long.

In today's issue:

• 🔭 Von der Leyen wants a legal drinking age for social media
• 🧠 Bricks that build their own shape, no central brain
• 🔬 A RAG system that's perfectly grounded and still names the wrong patient
• 🛠️ Score your product against the EU's child-safety checklist before it becomes law
• 🔗 AI catches bad dental X-rays before the dentist does

🔭 THE ONE THING

🏛️ Von der Leyen wants a legal drinking age for social media

The special panel von der Leyen convened after her 2025 State of the Union pledge handed the European Commission its final report on July 13, and the headline recommendation is blunt: lock kids under 13 out of social media unless a platform proves it's "safe by design." Her own framing, via the AP: you don't give a teenager car keys before a license, so why hand them TikTok. She's talking "phased and gradual access" by age band, with member states free to push the bar past 13 if they want to go further. I don't think this stays a recommendation for long: Brussels turned the DSA and the AI Act from white papers into binding law within a few years of first floating them. Any product with a teen user base in Europe, gaming or AI companions included, should start treating age verification as a design requirement, not a policy footnote.

🧠 MODELS & RELEASES

• 🧱 Sakana AI wired up physical modular bricks (microcontrollers, LEDs, connectors on all six faces) that decide their own shape with zero central controller: run neural cellular automata locally, gossip only with neighbors, and converge on the right form in under three minutes. Tested at nearly 200 bricks with 100% accuracy across four builds, from a 26-brick guitar to a 197-brick table, and it kept working even with a chunk of modules disabled. Smart Cellular Bricks

🔬 RESEARCH HIGHLIGHTS

• Clinical RAG's blind spot. A new eval shows a clinical RAG system can pass every automated grounding check, zero hallucinations, near-perfect faithfulness, while attributing the right fact to the wrong patient's chart. Standard RAG benchmarks check whether a claim is supported by retrieved text, not whether that text belongs to the correct entity. If you're routing patient or account records through RAG, "passed the eval" and "safe to ship" are not the same claim. arxiv.org/abs/2607.09349
• Soofi S 30B-A3B. A sovereign, open-source German/English foundation model: a hybrid Mamba-Transformer MoE that activates just 3B of its 30B parameters per token and holds its inference cache roughly flat as context grows. Read: cheap long-context inference you can self-host, with no US API in the loop. arxiv.org/abs/2607.09424
• Mach-Mind-4-Flash. 35B parameters, 3B active, and the authors claim it matches or beats 100B-class models through post-training alone, no extra pretraining compute spent. Worth watching whether that holds outside their own benchmarks. If it does, "just scale pretraining" gets a lot less convincing as a MoE strategy. arxiv.org/abs/2607.09375

🛠️ TRY THIS

Run a child-safety gap check against the EU panel's report before anyone asks you to

1. Pull the report's core asks out of today's lead coverage: age assurance, recommender defaults, reporting flows, parental controls.

2. Drop your onboarding flow, algorithm settings, and reporting UI copy into one doc.

3. Ask an LLM to score each requirement against what you actually ship: met, partial, missing.

4. Rank the gaps by how bad they'd look if a regulator or reporter asked about them tomorrow.

Prompt: Here's a summary of the EU expert panel's child-safety-online report: [paste]. Here's our product: onboarding flow, recommender settings, reporting mechanism: [paste]. For each report requirement, tell me met, partial, or missing, and why. Then rank the gaps by regulatory exposure.

Worth a look

• Stardog on AWS A walkthrough for putting a semantic layer over Aurora and Redshift so a Strands Agents agent on Bedrock AgentCore can answer customer-360 questions across both without an ETL job in between. Handy if your agent's real bottleneck is two databases that don't talk to each other. AWS post

🔗 QUICK LINKS

• Henry Schein One built Image Verify, a SageMaker model that checks dental X-ray quality the moment it's captured. It's live at over 10,000 practices already. AWS

See you tomorrow.

Pradeep Perugu