Seroter's Daily Reading — #738 (March 10, 2026)
Seroter's Daily Reading· Seroter's Daily Reading — #738 (March 10, 2026)
Welcome to Seroter's Daily Reading, an audio summary of Richard Seroter's daily reading lists. This is number 738, from March 10th, 2026. Eleven articles today, heavy on AI tooling and developer experience.
1. Gemini Embedding 2
Google released Gemini Embedding 2, their first natively multimodal embedding model. Previous embedding models handled text or images separately, using paired encoders stitched together. Gemini Embedding 2 is built on the Gemini architecture itself and maps text, images, video, audio, and PDF documents into a single unified embedding space. You can send interleaved modalities in one request — like an image with its caption — and get a single embedding that captures the joint meaning. It supports Matryoshka Representation Learning, meaning you can truncate vectors to smaller dimensions for fast retrieval and use the full 3072 dimensions when you need precision. Supports over 100 languages, up to 8192 input tokens for text, 120 seconds of video, and native audio processing without transcription.
2. Gemini in Google Sheets
Gemini in Google Sheets hit state-of-the-art performance on the SpreadsheetBench benchmark. The new beta features let you describe what you need and Gemini will create, organize, and edit entire sheets, handling complex data analysis. Their 70.48% success rate on SpreadsheetBench surpasses competitors and approaches human expert ability.
3. Dynamic UI for Dynamic AI: A2UI
A2UI (Agent to User Interface) addresses a real bottleneck: AI agents are increasingly dynamic, but the UX layer is still static. A2UI is a specification where agents communicate with a renderer that dynamically builds screens based on JSON content that agents produce on the fly. Instead of hard-coding every form, the ontology defines the business concepts and A2UI defines how components render. Companies like CopilotKit are actively building A2UI renderers.
4. Your Data Is Made Powerful by Context
Charity Majors on why separating observability data into pillars destroys its value. Her core argument is mathematical: the power of a dataset grows combinatorially with the number of attributes. Four fields give you 15 possible combinations. Fifty fields give you over a quadrillion. When you split telemetry into separate "pillars" — metrics, logs, traces — you destroy the relational seams between attributes. She cites an AI SRE startup whose agents went searching for richer signal and ended up going back to raw, pre-digested telemetry. If you care about AI-driven observability, this is essential reading.
5. Extend Your Coding Agent with .NET Skills
Microsoft released a public repository called dotnet/skills that hosts agent skills following the Agent Skills specification. These are the same workflows Microsoft uses internally while shipping the platform. The skills work with GitHub Copilot CLI, Visual Studio, VS Code, Claude Code, and other agents. Each skill is just a SKILL.md file with supporting artifacts. They run lightweight evaluators to score whether a skill actually improves results — like unit testing for agent context.
6. Genkit Dart
Google's existing Genkit framework, already available for TypeScript, Go, and Python, now comes to Dart. It's model-agnostic, supporting Google, Anthropic, OpenAI, and compatible models. The compelling part: you can write AI logic once and run it as a backend service or directly inside a Flutter mobile app. The "remote actions" pattern lets you move complex logic to the backend while your Flutter app calls it as a typed remote action, sharing schemas for end-to-end type safety.
7. Cracking the Code on Corporate Visibility
Tom Johnson on how he achieved corporate visibility by leading biweekly AI lunch-and-learn sessions for tech writers. Fifty to ninety people attend regularly. The formula mirrors his blog strategy over twenty years: produce content for large numbers of people on a regular basis. He makes a meta-point: his site is now so saturated that when he runs Gemini Deep Research queries on AI and tech comm, his own writing frequently appears in the sources.
8. A2A vs MCP
A comprehensive DigitalOcean tutorial comparing the two agent protocols. A2A handles agent-to-agent communication: capability discovery, task delegation, modality negotiation. MCP handles agent-to-tool integration: standardizing how AI applications connect to tools, resources, and APIs. Key insight: they're complementary, not competing. A planner agent might use A2A to coordinate with other agents, and those agents use MCP to access their tools.
9. Fixing AI Slop with a Gemini CLI Skill
A developer built a "deslopify" skill that cross-references AI-generated text against patterns tracked by tropes.fyi — overused adverbs, the word "delve," structural crutches like "It's not X, it's Y." The skill is just a SKILL.md file and a style guide. No complex tooling, no API calls — just a markdown file that teaches an agent to recognize and fix AI writing patterns.
10. AI Coding Assistants May Influence Language Choice
GitHub's Octoverse data shows TypeScript became the most-used language on GitHub in August 2025, with a 66% year-over-year increase. The article highlights a "convenience loop": when AI tools perform well with a language, developers find it easier to work there, which generates more training data, which makes AI tools perform even better. A study found 94% of compilation errors in AI-generated code were type-check failures — languages with strong type systems have an inherent advantage.
11. The Last Mile Problem Slowing AI Transformation
From the Harvard Business Review: executives have invested heavily in AI, initiated hundreds of pilots, and given employees access to tools like Copilot and ChatGPT. But even enthusiastic adopters are hitting barriers in the last mile — the gap between having the tools and actually transforming workflows.
The thread running through this list is the infrastructure layer maturing fast. Embeddings going multimodal, agent protocols getting standardized, programming languages competing for AI-friendliness, UI layers becoming dynamic. The tools are increasingly capable. The hard part, as the HBR piece suggests, is the last mile — actually integrating all of this into how people work.