Independent checks strengthen the Riemann proof claim

  • Lean’s mathematical community certifies the Riemann Hypothesis proof — Independent checks now address the central trust questions around the AI-assisted proof, while explaining its mathematical ideas remains separate work.
    • The joint Lean and Mathlib statement certifies both the formal statement and Alpöge’s proof. Lean checks mathematical arguments; Mathlib supplies its shared mathematical definitions.
    • Dozens of people reportedly compiled the library successfully. Two separate comparison tools confirm that the proved theorem matches Mathlib’s Riemann Hypothesis, a conjecture connected to prime-number patterns.
    • Several independently implemented proof checkers accepted an exported proof. Checking this inert data keeps the validation outside potentially malicious code in the original proof sources.
    • Andrew Sutherland’s MIT verification files separately report successful compilation and checking of all 175,187 project modules. Five headline theorems depend only on three standard logical axioms, the assumptions underlying the proof.
    • MIT’s checked generalization concerns Dirichlet L-functions, mathematical functions associated with prime-number patterns. It excludes broader families of such functions. Its table still lists some additional checking runs as unfinished.
    • This Morning’s proof item described verification as ongoing. The community certification supplies new evidence beyond that account and today’s earlier commentary.
    • Caveat: Formal verification establishes the specified theorem under its logical assumptions, without supplying a human explanation of the proof’s ideas. (HN 647 points · 519 comments · HN 66 points · 21 comments · joint community statement and public verification logs)

Smaller Inkling downloads come with quality tradeoffs

  • LookAtMyQuant releases smaller Inkling files for Apple Silicon — New downloadable files make local testing possible with less model memory, using custom software and accepting measurable quality losses.
    • The 5.38 GB and 8.87 GB text models have weights stored at lower precision to fit in less memory. They require the project’s Apple Silicon runtime, the software that executes the model.
    • The smaller file solves 134 of 200 GSM8K math problems versus the original’s 147. It passes 60 of 164 executable HumanEval coding problems versus the original’s 66.
    • The author reports 75.3 tokens per second on an M5 Max, or a median 124.4 with a small model proposing text for verification. Tokens are the pieces into which models divide text.
    • That assisted run reports 5.98 GB of active model memory, rather than a universal machine-memory requirement. The text downloads exclude the separate 637 MB vision component used in the author’s image tests.
    • This Morning’s outside evaluation concerns the original hosted model. These compressed files and their custom local runtime are a distinct follow-up, with separate quality and speed claims.
    • Caveat: These are author-run tests with reasoning disabled for math, and they leave performance on extended coding work unmeasured. (HN 243 points · 51 comments · released files, model cards and runtime documentation)

A privacy investigation produces a named customer warning

  • OpenRouter reportedly removes privacy labels after MiniMax prompt warnings — The follow-up identifies affected processing routes, giving developers something specific to investigate beyond a general warning about model resellers.
    • Ars Technica’s November 21 update reports customer emails warning that MiniMax prompts and outputs may have been read. It reports removal of the affected endpoints’ Zero Data Retention label, a promise against routine storage after processing.
    • The customer-posted email says unique test links were opened 40–130 hours after requests sent from July through October. AtlasCloud, Novita and StreamLake reportedly acknowledged forwarding MiniMax requests to MiniMax.
    • The notice says none of the planted AWS cloud-account keys were used in this test. Opening secret links and abusing credentials are different findings, so the broader investigation’s stolen-key results cannot simply be assigned to MiniMax.
    • MiniMax’s claimed retention change reached OpenRouter through StreamLake. The email says OpenRouter requested MiniMax’s own account and had yet to receive it, leaving the remediation claim secondhand.
    • Our existing investigation notes explain the original experiment and practical controls. The new development is the customer notice and named routing response, rather than another retelling of that study.
    • Caveat: The notice is publicly available through customer reposts and reporting, rather than an authenticated public company incident report establishing the full exposure. (Techmeme · Ars Technica update and customer-posted notice)
All gathered items - what was cut and why (9)