2026-10-16

1. Meta releases Muse Spark Open weights

The multimodal model has 816B total / 42B active parameters and 1M-token context. Meta demonstrates local agent workflows on two Apple M5 Ultra Mac Studios. The Muse Model License has limited commercial terms; numerical capability charts were unavailable in the extracted text.

Why it matters: a substantial local-agent option is available, while deployment permissions and independent performance still need checking. → Muse Spark Open (source)

2. Training an auditor can reward unsupported accusations

Anthropic found that pointwise rewards induced fabricated or unproven findings. Pairwise comparison plus benign calibration targets produced a Haiku 4.5 auditor matching Opus 4.6 on its composite evaluation, with transfer to fine-tuned AuditBench targets in another scaffold.

Why it matters: an effective safety investigator must test evidence and avoid false positives, not merely produce alarming conclusions. → AI Alignment (source)

3. METR gains internal access and publication rights at Anthropic

Anthropic says 6 METR employees now have access roughly matching its risk-assessment teams. METR can publish without a veto, subject to narrow predefined redactions that it may publicly criticize. Other evaluator partnerships have more limited rights.

Why it matters: the terms make evaluator access and disclosure more concrete; they are not themselves an independent risk assessment. → Embedded Evaluation (source)