Radar Daily Briefings
A clearer signal for WordPress and engineering
/\_/\
(=^.^=)
(")_(")
/\_/\
(=^.^=)
(")_(")
WASTE runs open-weights Kimi K3 on a 64 GB laptop
WASTE reports running the open-weights Kimi K3 model, with 2.78 trillion parameters, on a 64 GB MacBook Pro. Its converted container is 982 GiB, and measured decode speed is 0.45–0.62 tokens per second.
The C11 engine keeps the model trunk resident, streams selected mixture-of-experts weights from disk, and uses remaining RAM for a bounded expert cache. Read-ahead overlaps I/O with computation, while lookahead fetches likely experts for the next layer.
The result is local inference without a network connection, but storage and memory requirements are substantial. K3 support is model-specific, the API is not frozen, and several platform paths have not been validated on real target hardware.
WordPress 7.1 roundup covers iframe editor and Abilities API
Gutenberg Times’ Weekend Edition 371 surveys WordPress 7.1 Beta 4, Gutenberg 23.6, the forced iframed post editor, the Abilities API, Playground, and developer tooling. WordPress 7.1 remains in progress rather than final release.
The roundup highlights iframe migration work for blocks that access global browser objects or rely on editor styling. It also describes Abilities API contracts, WP-CLI inspection, REST execution, block bindings, and Playground site-management workflows.
For engineers, the key task is compatibility testing before the iframe editor arrives and evaluating the documented API boundaries. The source also flags experimental tooling and unresolved or unsupported capabilities, including theme.json overwrites without backups and limited Playground API availability.
QM releases multiplayer agent harness for collaborative work
QM is an open-source multiplayer agent harness for startup work, supporting isolated employee and room workspaces across Slack and the web. It provides collaboration through channels, group messages, projects, and shared organizational controls.
Each scope has its own memory, files, permissions, keychain view, scheduled jobs, web apps, and durable sandbox. A TypeScript core on Node uses Fastify, Postgres persistence, and interchangeable harness interfaces, including Pi, OpenCode, Codex, and Claude Code.
QM offers Strict, Auto, and Dangerous security postures, with command policies applying in all modes. Deployments run in the operator’s cloud account, while adoption, scale, and performance are not evidenced.
Study finds LLM financial advice improves with structured prompts
An MIT Sloan study examined long-term financial outcomes from advice generated by large language models. Researchers simulated people aged 22 to 89 following recommendations from GPT-5.2, GPT-5.6, and Gemini 3 Flash.
The models generally promoted saving, diversified stock funds, and lower stock exposure later in life. However, they handled unemployment poorly and often allowed portfolios to drift without active rebalancing. More detailed prompts containing financial and economic assumptions improved results.
The study also reports wealth differences linked to gender, financial literacy, and prior AI experience. Because the results are simulated and demographic benchmarks remain unclear, engineers should treat prompt design and bias evaluation as central requirements for financial AI systems.
NetBSD 11.0 released with three pending security fixes
The NetBSD Project has released NetBSD 11.0, with architecture-specific installation guidance and downloadable images. ARM users can obtain images preconfigured with U-Boot.
The announcement transparently documents three open security-related pull requests: an hdaudio ioctl access-check issue, a remotely triggerable IPFilter null-pointer dereference, and a PF fragment-reassembly use-after-free. It lists a local workaround for hdaudio and notes that IPFilter and PF are not enabled in released kernels by default.
The project says the fixes will enter the stable branch after release and target NetBSD 11.1, which it aims to publish within two months. Operators should review the release caveats before deployment.
A week of Matrix federation and administration failures
A Matrix operator recounts a week of recurring problems in a community deployment, including space-room permissions, moderation bans, missing rooms, federation failures, and desynchronized state. The account does not describe a formal release or confirmed fix.
Reported mechanisms include policy-room bans affecting an entire homeserver domain, Dendrite dropping outbound IPv6 traffic, DNS-driven federation failures, and room permissions that diverged across homeservers. Workarounds included changing permissions from another homeserver, re-inviting a bot, and unsubscribing from a policy list.
The incidents illustrate operational risks in federated Matrix deployments, including notification floods, hidden messages, weakened moderation, and uncertain membership state. The evidence is a single operator’s experience and leaves several issues unresolved.
datasette-apps 0.2a0 adds agent browser testing tools
datasette-apps 0.2a0 adds tools intended to improve Datasette Apps when they are created and edited using Datasette Agent. The release introduces app_debug() for invisible testing and app_list() for listing apps the user can edit.
According to the release notes, app_debug() displays an app in an opaque iframe with pointer events disabled, then executes agent-provided JavaScript inside the iframe. This supports smoke testing and measuring rendered element dimensions. The mechanism uses context.browser_task() from datasette-agent 0.4a0.
The release is relevant to engineers exploring agent-assisted Datasette workflows. The supplied evidence does not include broader compatibility or performance results.
MkLinux restoration completes on an Apple Workgroup Server 9150
A Floodgap lab restoration rebuilds an Apple Workgroup Server 9150 and configures it to dual-boot classic Mac OS and MkLinux. The completed system combines expanded memory, a faster CPU, and NuBus video hardware.
The A/V card is detected by MkLinux but causes the system to halt before the pager and Linux tasks start. A 1MB HPV card instead boots the installer kernel, although the Mach kernel identifies it generically and supports only 8-bit color at 1024×768.
The project provides hands-on evidence about preserving and adapting rare PowerPC-era Macintosh server hardware. Further kernel experimentation is planned, so the restoration remains an exploratory preservation effort rather than a contemporary production platform.
WordPress contributor requests PTE access for two plugins
The author of Maps & Route Planner and Cookie Consent Tool has requested Project Translation Editor access for the de_DE locale. The request concerns reviewing and approving translations for both plugins.
The author says complete German translations, including plugin strings and readmes, have been prepared. WordPress Project Translation Editors can approve translations for specific projects within a locale, with appointments handled by general translation editors after a request.
This is an access request rather than a completed editorial change. The supplied document does not indicate whether the user was appointed or whether the prepared translations were approved.
Contributor requests Gujarati translation editor access for Hueman
A WordPress contributor has requested Project Translation Editor access for the Hueman theme in the Gujarati (gu) language. The document describes the request as part of ongoing Gujarati translation work.
A Project Translation Editor can validate and approve strings for a specific project and locale. The contributor says Gujarati editor availability is limited, which may slow localization progress for the theme.
The document does not report that access has been granted or that an editor has been appointed. It is therefore a proposed community request, relevant mainly to contributors working on Gujarati WordPress translations.