Seroter's Daily Reading — #750 (March 26, 2026)

By Seroter's Daily Reading March 27, 2026

Audio summary of Richard Seroter reading list #750 covering Gemini 3.1 Flash Live, git branching strategies, Kubernetes beginner guide, trusting instincts at work, Ironwood TPU training, deployment strategies, Forrester AIQ on employee AI readiness, AI maturity assessment, Google neutral atom quantum computing, context engineering in 2026, open source licensing trends, and vibe coding XR.

Seroter’s Daily Reading — #750 (March 26, 2026)

🎧 Listen to the audio summary

750 of these daily notes. Richard is happy about that, and appreciates each reader who takes time to skim through it.

Gemini 3.1 Flash Live is Google’s latest real-time audio and voice model, designed for natural dialogue. It leads benchmarks in complex multi-step function calling, scoring over 90% on ComplexFuncBench Audio. The model has improved tonal understanding — detecting frustration or confusion and adjusting dynamically. For developers, it’s available via the Gemini Live API in Google AI Studio. Companies like Verizon and Home Depot are already testing it. All audio output is watermarked with SynthID to flag AI-generated content.

A practical post on choosing a git branching strategy for continuous delivery walks through three approaches: Gitflow, GitHub Flow, and trunk-based development. Gitflow’s multiple long-lived branches create friction for teams deploying web apps multiple times a day. GitHub Flow strips it down to one long-lived branch with feature branches and pull requests. The author recommends GitHub Flow as the sweet spot for most teams and raises an interesting question about whether AI-generated code volume will push new branching patterns.

A comprehensive beginner’s guide to Kubernetes walks from VMs and containers through pods, deployments, services, and ingress. The explanation of the control plane — the API server as central hub, etcd as cluster memory, the scheduler deciding where pods run, and the controller manager reconciling desired vs. actual state — is particularly well done. Written for someone who finds Kubernetes intimidating, and it succeeds at making the architecture approachable.

From HBR, advice on how to convince others to trust your instincts. The scenario: you’re in a meeting, the strategy sounds logical, everyone is nodding — but something feels off. Speaking up risks looking like a blocker. Staying silent risks a flawed plan. The article offers actionable advice for navigating that tension, especially for people who haven’t yet built the political capital to override a room’s consensus.

Google published a developer’s guide to training on Ironwood TPUs, their seventh-generation tensor processing units. Key innovations include native FP8 support (theoretically doubling throughput vs. BF16), the Tokamax kernel library with splash attention for long contexts and grouped matrix multiplication for mixture-of-experts models, and the ability to offload collective operations to dedicated SparseCores. Deep infrastructure content for teams training frontier models.

CircleCI’s guide on deployment strategies covers the full spectrum: big bang (simple but causes downtime), rolling (Kubernetes default, requires backward compatibility), blue-green (instant rollback but doubles infrastructure), canary (metric-driven with small traffic percentages), and A/B testing (business metric evaluation). The key nuance: deploying means putting code in an environment; releasing means exposing it to users. Some strategies treat these as the same event, others deliberately separate them.

Forrester’s AIQ assessment reveals employees aren’t ready for AI. While most organizations have rolled out tools like Copilot, Gemini in Workspace, or ChatGPT Enterprise, only about half offer AI training for non-technical staff. Knowledge of prompt engineering went from 22% in 2024 to just 26% in 2025. Employee fears about job loss persist, sometimes fueled by CEOs who blame layoffs on AI when it isn’t the cause. The core message: AI tools don’t make employees productive by default — you have to invest in organizational capability.

An engineering leadership newsletter explores how to assess AI maturity. Quotient developed a five-stage maturity model across six capability areas: enablement, governance, validation, workflow embedding, automation, and data context. Most organizations sit between stages 1–3. Risk follows a U-shaped curve — early experimentation and full autonomy are both high-risk, while stages 3–4 offer the best return. The article argues measuring AI’s impact should go beyond tool usage to delivery metrics like cycle time and change failure rate.

Google Quantum AI is expanding to neutral atom quantum computing alongside their superconducting qubits. Superconducting qubits excel at circuit depth with microsecond cycles; neutral atoms scale to arrays of ~10,000 qubits with flexible any-to-any connectivity but slower millisecond cycles. Dr. Adam Kaufman from CU Boulder and JILA will lead the neutral atoms team. The bet: investing in both modalities accelerates progress toward commercially relevant quantum computers by decade’s end.

The state of context engineering in 2026 from SwirlAI traces the field from posts by Manus and Anthropic in mid-2025. Matured patterns include progressive disclosure (agent skills loading information in tiers), compression for shrinking action history, and tool management for controlling capability surface. The most interesting pattern: agents that write their own skills, observing successful behavior and generalizing it. The unsolved problem: knowing when to deactivate loaded skills before accumulated context destroys the token advantage.

RedMonk’s analysis of open source licensing charts the continued dominance of permissive licenses. The crossover from copyleft majority happened between 2014–2017, driven by Apache and MIT. There’s a slight hint the pendulum might swing back — permissive ticked down from 82% in 2022 to 73% in 2025 on GitHub Archive data, possibly due to projects returning to AGPLv3. The unresolved question: how open source licenses apply in the AI era, where code is consumed by models in ways nobody anticipated.

Vibe Coding XR lets you describe an extended reality experience in natural language and have Gemini generate a working Android XR app in under 60 seconds. Built on the XR Blocks framework (WebXR, three.js, LiteRT), the demos include a math tutor visualizing Euler’s theorem, a physics lab with interactive balance scales, and a Schrödinger’s cat quantum state demonstration. Try it at XR Blocks Gem.

Source: Seroter’s Daily Reading List — March 26, 2026 (#750)

Articles covered: