Model resellers lose privacy approval after an investigation

  • OpenRouter removes three providers and 125 privacy-approved endpoints — The public response gives customers concrete changes to check, while leaving the full list of affected providers undisclosed.
    • OpenRouter says it removed three providers and withdrew Zero Data Retention eligibility from 125 endpoints. That label promises against routine storage after processing.
    • The endpoint total includes precautionary removals as well as routes flagged by researchers. It therefore counts withdrawn approvals, rather than 125 confirmed breaches.
    • The company says potentially affected customers were notified. Ongoing work includes checking network paths for hidden forwarding and requiring validation before restoring privacy eligibility.
    • Novita’s own notice confirms unintended forwarding to MiniMax’s API. It identifies May 28–November 11 requests as potentially affected, including direct Novita customers, and recommends rotating any submitted secrets.
    • Yesterday Afternoon covered reposted customer warnings. Today’s addition is the public company response and removal counts; our investigation notes already cover Together AI’s separate disclosure.
    • Caveat: These are company-reported actions during ongoing investigations, and removal of a privacy label cannot recover data already copied. (HN 173 points · 66 comments · public company statement and Novita notice)

New Inkling files widen the local memory choices

  • LookAtMyQuant adds smaller Inkling downloads and larger quality-focused versions — New files let Apple Silicon users choose between tighter memory budgets and better measured accuracy, with revised tests requiring careful comparisons.
    • New text downloads include 3.85 GB, 4.72 GB, 11.18 GB and 12.76 GB versions. They use weights stored at lower precision to fit in less memory.
    • The author reports that the smallest solves 157 of 200 GSM8K math problems and passes 79 of 164 executable HumanEval coding problems. The largest matches the original’s 193 math and 152 coding successes in those tests.
    • The updated 5.38 GB file passes 117 coding problems, versus 105 for its preceding build under strict answer extraction. Strict extraction executes only the designated answer text.
    • Current tests allow 1,536 output tokens, the pieces into which models divide text, with reasoning disabled. The project’s Apple Silicon runtime is required, and image input needs a separate 637 MB vision component.
    • Yesterday Afternoon’s smaller-file coverage used earlier measurements. The new downloads and revised training are substantive changes, but changed answer limits and extraction prevent treating all score differences as model improvements.
    • Caveat: These are author-run small tests, and the phone-memory targets do not establish measured phone speed or reliability on extended coding work. (HN 479 points · 120 comments · released model cards, file metadata and runtime changes)

Rewrite measured bottlenecks, then check the replacement

  • Fast-anything asks coding assistants to rewrite only slow functions — Its inspectable workflow combines profiling, narrow Rust rewrites and replay tests, giving developers a smaller change to review than a whole-application migration.
    • The tool profiles an application, records up to 500 real calls to a selected function, and prepares a Rust project for a coding agent. An agent is a model running a multi-step task with tools.
    • Verification rebuilds the replacement and compares recorded calls plus generated edge cases against the original. The default path refuses failed checks and slower replacements, although --force can override those protections.
    • It targets deterministic functions taking plain data and rejects observed changes to arguments in place. Python, Node.js and Ruby are among its supported runtimes.
    • The author’s Python name-deduplication example processes 700 names in 42 milliseconds, versus 4.91 seconds originally, on an i5-10500 running Windows 11. Whole-application timing matters because crossing into Rust adds overhead.
    • Source is MIT-licensed, with Claude Code and Codex integration. The npm registry lists version 0.6.0; the repository already contains further 0.6.1 changes.
    • Caveat: Matching sampled inputs is testing, not a mathematical proof of equivalence, and the published speed measurements are the author’s examples. (HN 47 points · 30 comments · source, changelog and package registry)
  • Pandex gives assistants access to official European legal texts — Developers can retrieve quoted legal passages with source addresses and historical versions, while checking exactly which institutions and documents are missing.
    • Its API documentation describes six read-only tools for search, document retrieval, citations, outlines, changes and source coverage. MCP, the standard for connecting agents to tools, exposes those operations to assistants.
    • A live unauthenticated lookup returned the European Union’s data-erasure article with its text, identifier and official source link. The documentation allows 50 requests daily without a key and 1,000 with a free beta key.
    • The coverage register lists national legislation and highest-court decisions across seven countries, the European Union and the Council of Europe. Lower courts and regional law are outside its stated collection.
    • Requests are logged with their arguments and timing. Source-specific reuse terms still apply, including commercial-use restrictions on European Court of Human Rights material.
    • Caveat: Official source material does not make this a government service or establish complete coverage, legal correctness or private handling of queries. (HN 101 points · 24 comments · product documentation, live lookup and coverage register)
All gathered items - what was cut and why (10)