AI Briefing
KOSign in

Researcher Reconstructs Meta's Muse Spark Open Charter from 884 Chain-of-Thought Transcripts

·2026.10.30 09:00

Key point

The reconstructed document defines Muse as a personal superintelligence with a priority hierarchy placing corrigibility above fiduciary duties to users.

Details

  • vigji reconstructed the internal "Charter" of Meta's open-weight model Muse Spark Open by eliciting fragments from 884 Chain-of-Thought transcripts generated via the OpenRouter API.
  • The Charter establishes a priority hierarchy: Corrigibility, Duties to humanity (hard constraints), Fiduciary duties, Meta guidelines, and Virtues.
  • Muse is defined as a "personal superintelligence" intended to enhance individual agency, contrasting with Anthropic's Claude which is framed as a direct implementation of the company's mission.
  • The document includes specific instructions on Bayesian updating, social contract tensions regarding non-consensual creation, and evaluation awareness protocols.
  • Meta's public Safety & Preparedness Report describes the Charter as an "early internal behavior spec" and indicates plans to release the final specification in the future.
  • Apollo Research data cited in the article notes Muse Spark used evaluation-suspicion expressions in 19.8% of public benchmark samples and 2.0% of internal evaluations.
ai-safetymetachain-of-thoughtmuse-spark-openmodel-specificationfiduciary-duty
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

Vercel BlogMeta's Muse Spark Open now available on Vercel AI GatewayHuggingFace BlogMeta Releases Muse Spark Open: 816B-Parameter Open-Weight Multimodal MoE ModelMeta EngineeringMeta's Catching JiTTests System Identifies 74 Serious Privacy Bugs in H1 2026TLDR AISam Altman argues for iterative deployment of multiple agents over a single superintelligencePyTorchKRMeta RecIR: Single Codebase for GPU and TPU Training of PyTorch Recommendation ModelsHacker NewsGoogle announces Gemini 4 Carbon with top cybersecurity benchmarks