Saturday is quiet, with seven useful items centered on local inference, agent permissions, and operational boundaries. DwarfStar 4 leads because its small MIT-licensed C runtime makes current open-weight MoE models usable across Apple, Nvidia, and AMD hardware, including SSD-backed caching for memory-constrained systems. Around it, Apple is tightening Full Disk Access as agents gain autonomy; gVisor is moving to CNCF governance; Meta opened Muse device firmware and SDKs; ChatGPT Sites turns prompts and Codex projects into hosted apps; Wagtail published a revealing two-billion-token GLM routing postmortem; and Nvidia’s new 64 GB DGX Spark makes local AI hardware economics look worse, not better.

Agent frameworks & tooling

  • DwarfStar 4 — an MIT C runtime serves current open-weight MoE models locally across Apple, Nvidia, and AMD hardware.

    • Supports DeepSeek V4/V4.1, GLM 5.x, and Qwen3.8 through CLI, agent, OpenAI, Anthropic, and Responses APIs.
    • Asymmetric 2-bit quantization and SSD-backed KV/prefix caching target 32–512 GB systems.
    • Measured: 27.6 generation tokens/second · 65K context · M5 Max with 128 GB.
    • Flag: Hardware results are project-run; model and quantization support is deliberately narrow.
    • (HN 249 · 68c)

  • Apple tightens macOS Full Disk Access — forthcoming controls will require more explicit user action before apps receive system-wide file access.

    • Current grants can expose files, mail, messages, and browser history without sufficient user understanding.
    • Apple explicitly identifies increasingly autonomous AI agents as raising the risk.
    • Flag: Apple gives no macOS version, API details, migration path, or release date yet.
    • (HN 208 · 141c · lobste.rs 13 · Techmeme)

  • gVisor moves to CNCF — the userspace Linux sandbox is leaving unilateral Google governance as demand grows for cheaper isolated execution.

    • CNCF accepted the donation; the project will enter Sandbox status and later pursue Incubation.
    • Non-Google maintainers gain merge access; the repository and testing infrastructure will move.
    • OpenAI, Anthropic, Modal, Nvidia, and others use or contribute to gVisor for isolated workloads.
    • Flag: This changes governance, not runtime behavior or performance.
    • (lobste.rs 15 · 2c)

  • Muse Gadgets — Meta released Apache-2.0 firmware and SDKs for connecting Muse to ESP32 boards, Raspberry Pis, and Linux devices.

    • Interfaces cover displays, buttons, audio, sensors, actuators, Home Assistant, and custom Linux commands.
    • Source includes board-specific examples plus ESP32 and Linux SDKs.
    • Flag: Access requires a Muse SDK token; community-grade hardware integrations carry no warranty.
    • (HN 202 · 87c · Techmeme)

  • Sites in ChatGPT — OpenAI’s public beta turns prompts or Codex projects into hosted, collaborative websites and apps.

    • Sites can store visitor files and progress, use custom domains, and restrict viewing or editing.
    • Business and Enterprise users can approve read-only access to their connected tools.
    • Scheduled work cannot access a visitor’s connected-tool data after the session.
    • Flag: Availability, limits, external sharing, and workspace controls vary by paid plan and region.
    • (HN 279 · 260c)

Models & research

  • One month coding with GLM 5.3 Flash — Wagtail reports a measured two-billion-token trial exposing cheap routine work and expensive agent mistakes.
    • GLM handled the first half for $68; provider degradation forced later routing to DeepSeek and Qwen.
    • One poorly chosen model consumed 450M tokens and $150 overnight on an MCP prototype.
    • GLM ultimately handled 1B tokens; Wagtail now targets flash models at routine work, not R&D.
    • Flag: This is one organization’s workload and accounting, not a controlled model benchmark.
    • (HN 165 · 125c)

Industry

  • Nvidia’s 64 GB DGX Spark costs $4,999 — the lower-memory local-AI box launches above the original 128 GB model’s price.
    • October 23 launch; memory falls from 128 GB to 64 GB and supported models from 200B to 100B parameters.
    • The current 128 GB line starts at $5,999; Nvidia attributes increases to memory constraints.
    • Multiple units can cluster, but the smaller configuration does not improve local-inference economics.
    • (Techmeme · PCMag)
All gathered items - what was cut and why (8)