From CSS math to sovereign AI: lean EU stacks and better product loops
NEWSLETTER | Amplifi Labs
Goodbye nth-child hacks: sibling-index() and sibling-count() power mathematical layouts
Smashing Magazine • May 21, 2026
New CSS Level 5 functions sibling-index() and sibling-count() return integers for an element’s position and sibling total, enabling pure‑CSS staggered animations, equal‑width tabs, hue distributions, circular menus, and z‑index stacking without nth-child boilerplate or JavaScript. Shipping in Chrome/Edge 138 and Safari 26.2 (Firefox in progress), they’re safe to progressively enhance via @supports, but watch Shadow DOM scoping, display:none still counting, custom property evaluation timing, and potential recalculation costs at massive scales. These features are visual-only—don’t use them to change semantic order or accessibility without syncing ARIA in JavaScript.
Shipping infrastructure: EU‑native stacks, sovereign AI, and messaging APIs
Norway’s National Library builds sovereign Norwegian LLM on 2PB Huawei flash
Around the web •May 25, 2026
Norway’s National Library is building a sovereign Norwegian‑language LLM, driving petabyte‑scale cultural data through an in‑house pipeline backed by 2 PB of Huawei OceanStor Dorado flash for low‑latency prep on a DGX H200/CPU cluster, before training on the Sigma2 Olivia HPE Cray EX (448 GPUs, 64,512 CPU cores; 5.3 PB ClusterStor). The main bottlenecks are data quality and throughput from a 60 PB, high‑latency preservation archive, necessitating custom ingestion, cleaning, deduplication, normalization, and validation. Open issues around evaluation (two written forms, dialects), governance, and orchestration across archive/on‑prem/supercomputer make this a blueprint for sovereign LLM efforts—and highlight Huawei storage’s growing footprint in Europe.
Build an all‑EU bootstrapper stack for under €10 per month
Around the web •May 25, 2026
A practical guide outlines a complete EU-native MVP stack that keeps monthly spend near €7–10, with compute as the only fixed cost. Use Hetzner Cloud’s CX33 (~€7; 4 vCPU/8 GB RAM/80 GB SSD) or Netcup for hosting, then layer permanent free tiers: Ahasend (1k transactional emails/mo), Brevo (300/day + marketing), Sender.net (2,500 subs/15k emails), Simple Analytics/TelemetryDeck, UptimeRobot/Healthchecks, Tally/Formbricks, Hanko passkeys, and Mollie or Creem for payments. The stack enables GDPR-friendly apps on EU infrastructure and scales smoothly without early lock-in to US hyperscalers.
Chert launches Twilio-like API for real iMessage at scale
Around the web •May 25, 2026
Chert offers a single API and webhooks to run real iMessage threads—blue bubbles, read receipts, tapbacks, group chats, and attachments—with automatic SMS/RCS fallback and sub‑second delivery telemetry. Positioned as a higher‑response alternative to A2P SMS, it manages sender reputation (identity rotation, warm-up, per-identity caps), and ships integrations for Salesforce, HubSpot, Slack, and more for onboarding, support, outbound, and AI agent workflows.
Product & UX excellence: Close the loop, learn faster, understand users
Ship Better: Close the Loop on Design Critiques
Nielsen Norman Group •May 22, 2026
NNG outlines how to make critiques actually improve shipped product by closing the loop on feedback. Use two followups: a quick post-session recap of what’s in, under investigation, and out, and a deeper update that attributes before/after changes to specific input while explicitly acknowledging what you didn’t implement and why. Bake these triggers into your workflow after major design changes, before dev handoff, and when critique input drives direction to prevent engagement and trust from eroding.
Skip the Polish: Use Design Disposables to Learn Faster
Nielsen Norman Group •May 22, 2026
The article argues for “design disposables”—rough sketches, notes, and quick prototypes made solely to think—distinct from polished deliverables meant for stakeholders. Over‑polishing early artifacts breeds sunk‑cost bias and slows exploration; be explicit about intent, keep disposables fast and discardable, and invest in deliverables only when communicating. AI can speed first drafts but may bypass the learning that comes from making—use it to accelerate, then critique and iterate rigorously yourself.
Stop Validating, Start Diagnosing: Four Levels of User Insight
Smashing Magazine •May 22, 2026
A four-level model reframes customer understanding—what users say, think/feel, do, and why they do it—prioritizing unobtrusive observation and mixed-method research over surveys, NPS, or think‑aloud protocols. It flags the unreliability of verbal probability terms and treats emotion as a signal, not an endpoint, while recommending practical tactics like exposure hours, live UX testing, co-design, and mining helpdesk data. Teams that adopt these practices can reduce bias, surface true user needs, and ship features that actually solve problems.
