Today at a glance

No single platform announcement dominated today. The stronger thread is more useful: developer tools are getting smaller, faster, and easier to inspect. The best reads cover compact speech, real-time document compilation, SQLite query-plan explanations, semantic metadata for AI and BI, and practical ways to make AI coding tools review and migrate code with more discipline.

Items

  1. Speech Recognition and TTS in less than 500kb · Hacker News

    Moonshine’s micro example puts speech recognition and text-to-speech under 500KB, which is a useful counterweight to the usual cloud-scale AI story. The interesting part is not just model size, but deployability: offline behavior, latency, power use, and privacy all improve when speech can run close to the user. This is the kind of constraint that matters for embedded devices, classroom tools, field apps, and edge interfaces.

  2. Real-Time LuaTeX: Recompiling Large Documents in 1ms · Hacker News

    This paper looks at recompiling large LuaTeX documents in roughly a millisecond. Even if you do not work with TeX, the underlying problems are familiar: dependency tracking, caching, incremental recomputation, and keeping the feedback loop tight. The same ideas show up in docs platforms, notebooks, preview panes, and design tools.

  3. Gleam Is Now on Tangled · Hacker News

    Gleam is now present on Tangled, giving the language community another collaboration home beyond the default GitHub path. That matters because language adoption is not only syntax and compiler quality; it is also issues, patches, identity, sponsorship, and community norms. Watching a smaller language experiment with those defaults is more interesting than the headline first suggests.

  4. SQLite Query Explainer · Simon Willison

    Simon Willison built a browser-based tool that runs SQLite through Pyodide and explains EXPLAIN and EXPLAIN QUERY PLAN output. SQLite is everywhere now, from desktop apps to edge services, but query plans still intimidate plenty of developers. Treat the explanations as a learning aid, then verify performance against real schemas, data volumes, and runtime conditions.

  5. apache/ossie · GitHub Trending

    Apache Ossie is an effort to standardize semantic metadata exchange across analytics, AI, and BI platforms. It is less flashy than another agent demo, but more foundational for enterprise AI. If models are going to answer questions about business metrics, teams need shared definitions for dimensions, measures, lineage, and ownership rather than a pile of vendor-specific semantic layers.

  6. tirth8205/code-review-graph · GitHub Trending

    code-review-graph builds a persistent local intelligence graph of a codebase so MCP and CLI tools can read less irrelevant context during reviews and large-repo workflows. That is a better direction than simply hoping for larger context windows. For big repositories, the durable asset is the map: what depends on what, which files move together, and where a change is likely to matter.

  7. Can AI recreate an app if you copy every page into it? · V2EX

    A V2EX thread asks whether AI can recreate an app if every screen is copied into the prompt. It is a sharp product question disguised as a community discussion. Screens can reveal layout and some interaction hints, but they do not expose permissions, state machines, billing rules, error handling, data contracts, or operational constraints. That is the gap between AI prototyping and shipping software.

  8. GLM 5.2 providers are undercutting each other on OpenRouter · V2EX

    Another V2EX thread tracks price competition between GLM 5.2 providers on OpenRouter. Model access is moving from launch-day excitement to commodity routing economics: price, latency, throughput, stability, and fallback behavior matter as much as benchmark charts. Teams using model routers should resist picking solely by the cheapest line item.

  9. Claude Codeと取り組んだ大規模レガシー移行の記録 · Zenn

    Luup published a detailed write-up about using Claude Code on a large legacy migration involving 38,000 added lines and 13,000 deleted lines. The strongest lesson is not that AI wrote a lot of code; it is that the team framed the work around behavior preservation, observation, review boundaries, and tests. That is the shape serious AI-assisted migration work needs.

  10. AIに「レビューして」はもう古い？「敵対的検証」のすすめ · Zenn

    Loglass argues for asking AI tools to perform adversarial verification instead of a generic review. The prompt shift is simple but meaningful: assume there may be a flaw, search for counterexamples, give evidence, and make a judgment. This is easy to turn into a pull-request checklist or design-review habit, and it tends to produce more actionable output than a vague review request.

Editor’s note

Today’s 10 picks break down as 3 English-source items, 2 Chinese community items, 2 Japanese items, and 3 wildcard picks. All seven configured sources were reachable; Anthropic News was reachable too, but its latest items were less implementation-heavy than today’s selected candidates. Dev Digest editor would start with SQLite Query Explainer, Apache Ossie, and Luup’s Claude Code migration write-up.