SBCs Hit a New Gear, Safer Local Agents, and No‑GIL Reality Checks
NEWSLETTER | Amplifi Labs
2025 SBC Roundup: Qualcomm arrives, CIX P1 dominates benchmarks
Around the web • March 5, 2026
A 15-board 2025 SBC roundup highlights Qualcomm’s strong entry—Radxa’s $70 Dragon Q6A (QCS6490) delivers Raspberry Pi 5–class single‑core and even dabbles in Windows on ARM—while Raspberry Pi 500+ is a polished, iterative update. New CIX P1–based boards (Radxa Orion O6N, Orange Pi 6 Plus) lead multi‑core performance, Rockchip’s RK3576 becomes the mid‑range workhorse, and RISC‑V options improve but still trail ARM. With LPDDR4/5 “RAMageddon” lifting prices and affecting availability, developers should verify current pricing and favor boards with strong mainline/Armbian support.
AI Agents, Guardrails, and Developer Workflows
Lock down local AI agents with macOS-native, deny-by-default sandbox
Around the web •March 8, 2026
Agent Safehouse is a macOS-native, kernel-enforced sandbox that runs local AI/dev agents with deny-by-default permissions. A single Bash script auto-grants read/write to your current repo (git root) and read access to toolchains while blocking home-directory secrets (e.g., /.ssh, /.aws); it’s tested with agents like Claude Code, Cursor, Aider, and Gemini CLI. An optional LLM-guided flow generates least-privilege sandbox-exec profiles and shell wrappers so agents run safely by default.
LLM Agents Revive Literate Programming: Executable Runbooks, Synced Prose
Around the web •March 8, 2026
The author argues that LLM coding agents make literate programming practical by auto-syncing prose and code, generating Org Mode runbooks with executable blocks, and handling tangling/extraction—removing the historical overhead of maintaining parallel narratives. This enables reproducible tests and narrative-friendly codebases that may improve code quality, though the approach is currently proven mainly for testing/ops workflows and is limited by Org’s Emacs tie-in and Markdown’s lack of metadata.
Claude, Google Bring GenUI to Chat with Buttons and Checkboxes
Nielsen Norman Group •March 6, 2026
GenUI is moving from concept to practice as chat systems dynamically generate simple controls—checkboxes, buttons, and form fields—to capture context inline. Examples include Google AI Mode’s hotel-selection checkboxes and Claude’s AskUserQuestion module (capped at four questions), which reduce typing and memory load while eliciting structured inputs for more personalized results. For teams building chat UX, this pattern argues for constrained, contextual UI elements and guardrails to boost response quality and cut conversational back-and-forth.
UX in the AI Era: Measurement, Inputs, Inclusion
AI Recasts UX: From Pixel Production to Strategic Direction
Smashing Magazine •March 6, 2026
Generative AI is automating wireframes, prototypes, and design-system enforcement, shifting UX designers from artifact makers to directors of intent. The article argues AI’s strengths—speed, consistency, and large-scale data analysis—free humans to focus on empathy, ethics, judgment, and product strategy, with practical changes like prompt-driven ideation and curating AI options. Teams should adopt AI for exploration while doubling down on research, accessibility, facilitation, and ethical decision-making to increase impact and accountability.
Stop Chasing P-Values: Optimize for Practical UX Impact
Nielsen Norman Group •March 6, 2026
NN/g reminds teams that p-values indicate reliability, not impact—large datasets can make trivial changes (e.g., +0.2% conversion, -8 ms load) look “significant.” The article details how to judge practical significance via user perception, business value at scale, and effect size, and notes that big, consistent patterns in small samples can still warrant action. For A/B testing and analytics workflows, combine both lenses to prioritize changes that measurably improve UX and business outcomes.
Practical Guide to Inclusive UX Research: Recruit, Facilitate, Analyze
Smashing Magazine •March 3, 2026
Smashing Magazine has released “Accessible UX Research” by Dr. Michele Williams, a hands-on roadmap for integrating disabled participants across all research phases and moving beyond checkbox compliance. The book covers recruiting, accessible study design and facilitation, bias-aware methods, and impactful reporting—helping teams bake accessibility into their product pipeline and build better experiences for more users. Available now in hardcover and as eBook (PDF/ePub/Kindle); the eBook is free for Smashing Members.
Developer Tools & Code: Practical Wins and Performance Tradeoffs
FontCrafter: Free, Client-Side Tool Turns Handwriting into Installable Fonts
Around the web •March 9, 2026
FontCrafter is a free, browser-only utility that converts scanned handwriting into installable fonts entirely on-device—no accounts, uploads, or server processing. It exports OTF, TTF, WOFF2, and Base64 for CSS embedding and supports ligatures and contextual alternates, making it useful for bespoke UI, branding, and web projects. For developers, this offers a privacy-preserving way to generate web-ready fonts without vendor lock-in or subscription costs.
Python's No-GIL: 4x Speedups, Memory Overhead, Serial Energy Penalty
Around the web •March 6, 2026
A study on Python 3.14.2's free-threaded (no-GIL) build benchmarks execution time, CPU utilization, memory, and energy across NumPy, sequential kernels, threaded numerical, and threaded object workloads versus the standard GIL build. Parallel, independent-data workloads achieve up to 4x speedups with proportional energy reductions and strong multicore use; sequential tasks see 13-43% higher energy, and shared-object threading often stalls or regresses due to lock contention. Memory usage rises—especially virtual memory—due to per-object locks, added thread-safety mechanisms, and a new allocator, so teams should validate workload parallelism and memory budgets before adopting no-GIL.
