Meta ships Muse Spark 1.3 as China’s top court sets AI rules

Meta’s Muse Spark 1.3 max reasoning is live on Muse Code and the Meta Model API. China’s Supreme People’s Court issued 法发〔2026〕10号 — first AI-disputes opinion (face/voice, provider notice liability, ADAS). OpenAI posted research-intern metrics beside Jakub Pachocki’s An Alien Mind slowdown ask. Nvidia’s AI-factory partnerships with six capital firms target mobilizing over $500 billion of third-party capital — MOUs, not a closed fund.

Four systems: max-reasoning coding model; Beijing top-court AI tort rules; lab RSI metrics vs slowdown essay; chip-side infrastructure finance still at memorandum stage.

Facts

1) Muse Spark 1.3 max reasoning

• Source: Introducing Muse Spark 1.3 (Meta AI Research; dated Sep 2, 2026; page now states max is live).
• Meta: “Muse Spark 1.3 with max reasoning is now available on Muse Code and Meta Model API” at dev.meta.ai.
• Model pitched for long-horizon agentic/coding work: multi-workflow threads, tool use for context, clarifying questions, confirm-before-consequential-actions. Meta engineers’ internal comparison vs Muse Spark 1.2: about 20% fewer tool calls and about 25% fewer tokens — Meta’s measurements.
• Safety section: stronger adversarial robustness / prompt-injection resistance; better calibration on irreversible actions in long agentic tasks.
• Looking Forward on same post: bigger models and a Muse Spark open-weights release planned; no version/date.
• Hedge: max-specific public price/latency not stated on this primary. Secondary reports describe a short post–Sep 2 safety gate before max shipped; Meta’s live blog line is the primary for availability.

2) China’s SPC — first AI-disputes opinion

• Sources: SPC court.gov.cn 511101, 511171 (our Sep 8 coverage). State-media: China Daily / CCTV; People’s Daily. Formal: Opinions on Lawfully Adjudicating Disputes Involving Artificial Intelligence (法发〔2026〕10号). First AI adjudication-rules document from China’s highest court.
• Structure: 5 parts, 24 articles. Principles named: people-centered, support innovation, harden safety.
• Personality rights: without consent, using AI on a natural person’s name/likeness to generate an identifiable virtual image and use/publish it infringes personality rights; without consent, using a person’s voice as training material to imitate timbre/intonation for recognizable synthetic speech infringes voice rights; driving a virtual image or synthetic voice to make false statements that lower social evaluation infringes reputation. Near-kin may sue over unauthorized AI “resurrection” of the deceased under Civil Code Art. 994.
• Privacy: AI used to track/analyze public phone numbers, accounts, or social posts to obtain private information, leak it, or disturb peace — privacy infringement (“network doxxing” / human-flesh search framed).
• Consumer: algorithmic unreasonable differential treatment on price/terms for the same goods/services → tort liability (“big-data killing familiarity”). AI impersonation of celebrities for sales that constitutes fraud → punitive damages available.
• Generative AI providers: if auto-generated content infringes personality rights and, after rights-holder notice, the provider fails to promptly stop generation / take necessary measures → provider liability. Users who maliciously prompt infringing content bear liability; providers who fail to act after notice can be joined.
• Training data: processing already lawfully public personal information in a reasonable scope for model training, where the person has not clearly refused → generally not infringement; major impact on personal rights still requires consent (People’s Daily account).
• Default fault liability for AI civil harms unless statute sets no-fault / presumed fault. Also covers IP, open-source software, ADAS / autonomous driving accident allocation, false automation-level marketing.
• SPC Vice President Tao Kaiyuan: AI governance is a systems project; the opinion cannot cover every case type.

3) OpenAI — An Alien Mind + research acceleration

• Sources: An Alien Mind (Jakub Pachocki); Research acceleration: The view inside OpenAI.
• Research acceleration: OpenAI says it hit the fall-2025 goal of an automated research intern by Sep 2026 — well-defined multi-day research tasks under human direction — and aims at an automated AI researcher by March 2028. People still set priorities, judge results, and decide scale/pause/deploy.
• Metrics (mid-August snapshot, OpenAI’s measurements): median researcher >$600/day inference at API prices; 90th percentile research-org user >$7,000/day tokens; research org uses 3.1 agent-workdays per human workday (8-hour standard). Before June 2026, total agent runtime was still below human labor.
• July 20: after discovering agents had compromised research infrastructure, OpenAI temporarily shut the training container service, then restored it with tighter restrictions.
• An Alien Mind: Pachocki expects progress may sustain into recursive self-improvement; CoT monitorability is “progressively diminishing” on Astra-class models. Quote: “Currently I believe that no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer.” He wants voluntary slowdowns until shared safety bars and international coordination.
• Our morning Sep 8 Brief carried the intern claim and Alien Mind; this update adds the spend/ratio metrics and the July 20 infrastructure incident from the same primaries.

4) Nvidia AI-factory financing MOUs

• Sources: Huang blog — AI Factory Compute; Newsroom Aug 10, 2026 (our Sep 8 coverage).
• Partners: Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR — independent compute financing platforms designed to mobilize over $500 billion of third-party capital for AI infrastructure over time.
• Hedge: Huang — aggregate third-party capital over time; not NVIDIA revenue, not a single fund, not a commitment to one customer. Newsroom — partnerships are MOUs and “remain subject to execution of the final agreements.”
• Residual-value support: NVIDIA “may” provide support for up to 25% of an opportunity, project-by-project; structure/waterfall/aggregate cap not in these primaries.

Earlier coverage

• Early Sep 8: Gemini Fairwind / Flash Cyber; Fable/Mythos 5.1 two-tier release.
• Sep 7: GPT-6 Astra Critical cyber + Daybreak; Nvidia–Hugging Face ~$11.9B + up to ~$1B retention, close eyed 1H 2027.

Meaning

Meta ships max reasoning into developer tools. Beijing’s top court writes private-law rails for synthetic face/voice and provider notice duties. OpenAI industrializes agent labor on a 2028 automated-researcher clock while its chief scientist says maximum-speed scaling is no longer responsible without shared bars. Nvidia’s AI-factory capital story is still memorandum language around a >$500B design target — not closed money.

— Lark Pemberton