Superintelligence,
for everyone?

Research date: 2026-10-05 (UTC) Terminology: in this report, "Muse" refers to the personal AI agent product and model family Meta released in September 2026 (Muse Spark), unrelated to Anthropic's Claude. Confidence labels: [Officially confirmed] = published via Meta's official channels or by Zuckerberg himself, or verified live in-browser; [Authoritative media] = reported by Reuters/WSJ/FT/Bloomberg/TechCrunch/Fast Company with multi-source cross-verification; [Single-source] = single or weak source only, not cross-verified.

Executive Summary

Meta's "Personal Superintelligence" is the company-level AI vision Zuckerberg established in July 2025: unlike the industry path of using superintelligence to "centrally automate all valuable work," Meta advocates distributing superintelligence into everyone's hands, as personal agents that understand individual goals and handle to-dos 24/7[Officially confirmed]. Its strategic pillars are: ① Meta Superintelligence Labs (MSL, founded 2025-06-30, Alexandr Wang as Chief AI Officer)[Officially confirmed]; ② hundreds-of-billions-scale compute investment (2026 capex guidance $130–145B; GW-scale clusters like Prometheus/Hyperion) [Officially confirmed + authoritative media]; ③ product vehicles: the Muse personal agent (launched 2026-09-08) + AI glasses/wearables + the WhatsApp/Instagram/Facebook ecosystem[Authoritative media]; ④ open-source strategy shifting to a "hybrid" model: largest models closed, some models open-weight [Authoritative media / single-source].

Its distinctiveness lies in "decentralized personal empowerment vs. centralized AGI": OpenAI (Gentle Singularity), Anthropic (powerful AI + strong safety-ism), Google DeepMind (AGI for science), and xAI (maximum truth-seeking) all lean toward centralized R&D and institutional deployment, while Meta bets on putting AI into the pockets and onto the faces of 3.6 billion users [Analysis, multi-source basis].

The three biggest controversies: ① the inherent conflict between the advertising business model and "AI doing errands for users, bypassing ads"[Opinion]; ② a privacy trust deficit (Muse demands deep permissions; Reuters exposed the "human concierge" manual-call incident, later rolled back)[Authoritative media]; ③ the forced unwinding by China's NDRC of the $2B Manus acquisition — the first revoked closed cross-border AI deal[Authoritative media].

For users in China: Muse launched US-only and Meta services are unavailable in mainland China — out of reach for now; alternatives include the independent Manus, domestic in-ecosystem API routes (Qwen) and system-level GUI routes (Doubao), plus self-hosting Chinese and foreign open models [Officially confirmed + authoritative media].

1. Origins and Official Statements

1.1 Three Core Primary Texts

1.2 Core Arguments of the 2025-07-30 Open Letter (Key Points)

1. Timing: "Developing superintelligence is now in sight" — signs of AI systems self-improving have appeared in recent months; "The rest of this decade seems likely to be the decisive period"[Officially confirmed].
2. Core definition: everyone owns a personal superintelligence that "helps you achieve your goals, create what you want to see in the world, experience any adventure, be a better friend to those you care about, and grow to become the person you aspire to be"; "Meta's vision is to bring personal superintelligence to everyone"[Officially confirmed].
3. Demarcation from the industry: explicitly distinguished from "others in the industry who believe superintelligence should be directed centrally towards automating all valuable work, and then humanity will live on a dole of its output"[Officially confirmed].
4. Product form preview: "Personal devices like glasses that understand our context because they can see what we see, hear what we hear, and interact with us throughout the day will become our primary computing devices"[Officially confirmed].
5. Openness and safety: will strictly manage risk — "careful about what we choose to open source" — while advocating benefits be shared as widely as possible[Officially confirmed].

1.3 Evolution of Official Statements (2025→2026)

• 2025-07-14 (Threads): Zuckerberg announced "hundreds of billions of dollars into compute" for superintelligence; the first multi-GW data center Prometheus goes live in 2026, Hyperion to scale to 5GW in the future [Officially confirmed, Reuters reporting].
• 2026-04 (The Information TITV interview "Inside Zuckerberg's AI Playbook"): reiterated that personal superintelligence differs from the industry's centralized path; discussed capex strategy; responded to high-salary poaching rumors, saying the $100–200M figures were exaggerated [Authoritative media relay, single-source chain, flagged single-source].
• New content in the 2026-08-10 manifesto (relative to the 2025 open letter) [Officially confirmed via TechCrunch/multiple summary cross-checks]:
  • Distilled into the slogan: "Rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it";
  • Made concrete as an "exceptionally capable personal agent": understanding the user, their goals, and everything they care about, handling to-dos 24/7 (health/relationships/career/finance/housework/hobbies); Zuckerberg says he uses it to filter information, prototype designs, monitor sleep and training, and plan baking recipes;
  • Economic view: "invention, not automation, will be the greatest contribution of superintelligence," predicting more rather than fewer jobs in the future;
  • Freemium commitment: free tier for billions; compute priced via a "dynamic auction mechanism" to guarantee the lowest price;
  • Privacy goal: analogous to WhatsApp end-to-end encryption, possibly achieving a model where even Meta cannot access user information; Confidential VM launching within the year (users hold their own keys);
  • Policy proposal: frontier labs should share mid-training model checkpoints with the US government before public release [Single-source, mentioned only by business20channel].
  • Companion move: released the open-weight model Muse Glimmer (29.6B parameters, Apache 2.0, distilled from Muse Spark) [Single-source, mentioned only by bit.com/businessmodelanalyst].

2. Strategic Substance

2.1 Technology Path

Model trajectory: Llama 2/3 (open weights, 2023–2024, developer goodwill) → Llama 4 (2025-04, poor market reception; Yann LeCun admitted to the FT that benchmark scores were "fudged a little bit," and Meta promoted scores from an unreleased version) → the rumored "Behemoth" flagship delayed (WSJ reporting; Meta spokesperson dismissed it as "speculation and hearsay") → next generation codenamed "Avocado" → Muse Spark (released 2026-04-08, MSL's first product, first fully closed proprietary model; Wang said "rebuilt our AI stack from scratch"); three versions iterated within 4 months of launch (Muse Spark series, Fast Company 2026-09)[Authoritative media]. Benchmark scores (Artificial Analysis 52 points, GPQA Diamond 89.5%, etc.) appear only in weak sources[Single-source].

Open-source strategy shifts to "hybrid": the largest, strongest models stay proprietary, served via cloud APIs; some future versions are planned to be open-sourced (Wang's words: "plans to open-source future versions"; a Gartner analyst called this a "major shift"). Background: DeepSeek built a competitive Chinese model on the Llama architecture, causing internal displeasure; Llama 4's reputation slid; training costs climbed [Authoritative media / multi-source]. Meanwhile Chinese open models (GLM-5, Qwen 3.5, DeepSeek V3.2) have surpassed Llama 4 on benchmarks[Single-source].

Compute investment (the strategy's second pillar):

• 2026 capex guidance of $130–145B, mostly for AI infrastructure (more than double the 2024–2025 combined total)[Authoritative media];
• Prometheus (New Albany, Ohio, 1GW, online 2026), Hyperion (Louisiana, 5GW at full build; budget from $10B in 2024 → $27B JV with Blue Owl in 2025 → announced expansion to $50B+ in 2026-07; Bloomberg says including chips it could reach $250B), El Paso 1GW campus (JV with BlackRock in 2026-07, online 2028) [Officially confirmed + authoritative media];
• In-house inference chip MTI 500 (30 petaflops MX4)[Single-source]; 33 data centers worldwide[Authoritative media].

Organization and talent (MSL):

• In 2025-06, acquired 49% of Scale AI for $14.3B (Meta's second-largest deal ever), Wang became Chief AI Officer[Authoritative media];
• Large-scale poaching from OpenAI/DeepMind/Anthropic (including ChatGPT/GPT-4 co-creator Shengjia Zhao, o1 researcher Trapit Bansal, the Zurich vision team)[Single-source];
• Yann LeCun left at end-2025 to start his own venture (world-model direction); FAIR marginalized, marking the end of the "academic era"[Authoritative media];
• MSL restructured and cut ~600 people in 2025-10[Single-source]; Meta cut 8,000 jobs in 2026-05 to fund AI [Weak dual-source, questionable].

2.2 Product Vehicles

1. The Muse personal agent (the core productization of the vision): launched 2026-09-08, which Meta called "the world's first personal AI agent built for everyone"; standalone iOS/Android apps, the muse.ai website, available inside WhatsApp, with a Mac client added 09-17 [Authoritative media / multi-source]. Each user gets a private cloud PC — Secure VM (tasks keep running in the cloud after the app closes) — plus the Sentinel network-control layer; it can send emails, book travel, fill forms, shop, make phone calls (calling launched 09-16), and operate browsers; sensitive actions require user confirmation; Stripe Link / Shopify Shop Pay / PayPal × Mastercard Agent Pay form the payments loop [Multi-source cross-check]. Pricing: free tier (Zuckerberg says 100M tokens per week) + Power $20/month + Maximum $100/month[Single-source]. Launched US-only, 18+, with a linked payment card required[Multi-source]; Sensor Tower data claims 2.5M+ US downloads in two weeks, topping the US App Store [Single-source chain, questionable].
2. AI glasses / wearables (Zuckerberg-defined "primary computing device"): Connect 2026 (09-23/24) launched Ray-Ban Meta Gen 3 ($449), Ray-Ban Meta Audio ($349, first camera-free audio-only), Meta Adventurer ($249), Ray-Ban Meta Display ($799, with display), Muse Charm (keychain AI companion), Meta VR Glasses ($1,299.99, spring 2027); Muse officially entered glasses (voice-dispatching schedule tasks); FDA-certified hearing enhancement ($149.99); 100+ combinations across three brands by year-end [Multiple secondary media cross-checks; meta.com navigation already shows Meta VR Glasses / Muse Charm entries — verified live in-browser]. Market: IDC figures put Meta-family smart glasses at ~76% of shipments (2025) / 69% combined with EssilorLuxottica (Fast Company citing IDC); Ray-Ban Display sold only ~15,000 units in its first quarter (IDC via Reuters) [Weak source chain, questionable].
3. Social ecosystem: Meta AI is built into Facebook/Instagram/WhatsApp/Messenger; Muse can take over Instagram/Facebook merchant tools (including ad buying); 3.6 billion daily-active users form the distribution moat[Authoritative media].

2.3 Timeline and Milestones (see "Timeline" at the end)

Key milestones: 2025-06-30 MSL founded → 2025-07-30 open letter → 2026-04-08 Muse Spark (closed-source turn) → 2026-08-10 manifesto → 2026-09-08 Muse launch → 2026-09-23 Connect 2026 (glasses + Muse convergence). Prometheus online 2026, Hyperion at 1.5GW in 2027, El Paso online 2028 form the compute timeline [Officially confirmed + authoritative media].

3. Comparison with Competitors' AGI/Superintelligence Visions

Meta's distinctiveness (analysis): Meta is the only one of the five to make "personal" the qualifier of its vision, and to argue that distribution breadth itself is a safety strategy ("centralization is the risk"), directly countering the centralized, cautious path of OpenAI/Anthropic — TechCrunch readers called its manifesto "anti-Dario." Its confidence comes from a 3.6-billion-DAU distribution network and advertising cash flow; its weakness is that model capability is still seen as lagging (SemiAnalysis in 2026-07 saw Meta as having the best chance to catch OpenAI/Anthropic, but Fast Company noted it trails by years in code and cloud services) [Analysis, based on authoritative media].

4. Execution Checklist (as of 2026-10-05)

• MSL founded, AI organization consolidated (2025-06-30)[Officially confirmed]
• Scale AI $14.3B/49% closed, Wang became Chief AI Officer[Authoritative media]
• Open letter + manifesto established the narrative (2025-07, 2026-08)[Officially confirmed]
• Muse Spark closed model released and iterated rapidly (from 2026-04)[Authoritative media]
• Muse personal agent launched (2026-09-08, US 18+; App/Web/WhatsApp/Mac)[Authoritative media]; official help center confirms "not yet available in all regions" [Officially confirmed, verified live]
• Payments loop: Stripe Link, Shopify Shop Pay, PayPal Agent Pay (2026-09) [Single-source chain, questionable]
• Connect 2026: glasses lineup refresh + Muse in glasses + new hardware (Muse Charm, VR Glasses) [Multi-source; official site navigation corroborates]
• Hyperion expanded to $50B+/5GW; El Paso BlackRock JV; 2026 capex $130–145B[Authoritative media]
• Manus acquisition: FAILED — blocked and forcibly unwound by China's NDRC in 2026-04; internal sunset in 06; Manus restored to independence 08–11[Authoritative media]
• Hybrid open-source strategy: open-weight releases like Glimmer[Single-source]; flagship stays closed[Authoritative media]
• International expansion: Muse still US-only; glasses in Japan/UK/Canada/Europe (Display lands in France/Italy/Germany 10-13), no mainland China plans [Officially confirmed + authoritative media]

5. Controversies and Doubts

Factual

1. Manus acquisition forcibly unwound: Meta announced the ~$2B acquisition on 2025-12-29; in 2026-04 the NDRC's foreign-investment security review ruled to block it and demand unwinding (a completed deal unwound — a first); in 03 two co-founders were barred from leaving the country; on 06-12 Bloomberg obtained an internal memo showing Meta cutting Manus's system access and "sunsetting"; in 08–11 Manus announced restored independence, with user data deleted after 12-29; Caixin reported Tencent/ZhenFund/HSG bought it back for $2B (unconfirmed by either side) [Authoritative media; buyback figure single-source, questionable].
2. Llama 4 benchmark controversy: LeCun admitted to the FT that scores were "fudged a little bit"; "Behemoth" delayed (WSJ), Meta dismissed it as hearsay[Authoritative media].
3. Muse "human concierge" incident: Reuters exclusive on 2026-09-22 — Meta had human contractors make some of Muse's phone calls; employees protested the privacy risk, one saying a contractor made racist remarks on a call; an MSL VP admitted inadequate disclosure was a mistake and rolled it back[Authoritative media].
4. Financial structure: Q2 2026 advertising was 98% of revenue, net profit $15.8B; Reality Labs lost $84B cumulatively 2020–2025; eMarketer expects Meta to surpass Google as the world's largest digital advertiser by end-2026 [Authoritative media via Fast Company].
5. Polling: Bentley-Gallup 2026-07: Americans who think AI does more harm than good rose from 31% to 39% within a year; only 9% think benefits outweigh harms [Fact — poll].

Opinion (criticism)

1. Inherent conflict of the ad model: if Muse does errands for users and bypasses recommendations and ads, Meta loses traffic entry points and ad reach — "whether an app hungry for personal data will eventually backfire on Meta" (Bernstein analyst); Chinese observers call it "a fight over the toll booth charging for customers' attention"[Opinion].
2. Privacy trust deficit: Muse demands deep permissions (email/payments/location), while Meta has the Cambridge Analytica record; whether the technical promises of Secure VM/Confidential VM can offset distrust is the biggest challenge (livemint, NPR, TechCrunch)[Opinion].
3. Hollow vision: a Fast Company long read says Zuckerberg is vague on "what to do once you have a transformative model"; Mollick says Meta "fell out of the race"; TechCrunch criticizes the manifesto for dodging risk and calls "dynamic auction" compute pricing detached from user experience[Opinion].
4. Compute monopoly and environment: hundreds of billions in compute, the water/power and community impact of GW-scale clusters; the Wyoming data-center wastewater incident [Single-source]; supporters cite Entergy's projection of $2B in residential electricity savings over 20 years (projection) [Mostly opinion].
5. The manifesto's political intent: read as a lobbying document — arguing "Meta should be allowed to build data centers anywhere," with the checkpoint-sharing proposal seen as regulatory arbitrage[Opinion].
6. Organizational turmoil: LeCun's departure, FAIR's marginalization, the cultural shock of high-salary poaching[Opinion].

6. Implications for Users in China

1. Current availability [Officially confirmed + fact]: Muse launched US-only (18+), requiring US IP and payment-card verification; Meta's official help center (zh-cn, verified live in-browser) states "the Muse app and website are not yet available in all regions"; Facebook/Instagram/WhatsApp were never available in mainland China; Ray-Ban Meta glasses launched in Japan in 2026-05, with no mainland China release plans seen. V2EX already hosts workarounds from mainland users (cloud-browser US IP + Visa/Mastercard $1 verification) — a gray area.
2. Structural judgment (marsbit 2026-10-04, opinion): China "can't wait for Muse" not because of a technology gap, but an ecosystem-structure difference — America's open web plus antitrust wall-breaking lets browser agents reach services across the whole web; China's services are enclosed in super-apps with a barren web layer: "the same key opens a vast expanse in America, and hits a wall in China."
3. Alternative paths:
   • The independent Manus (Chinese founding team): resumed operations; in 2026-09 negotiated a $500M raise at a $4B valuation (Bloomberg, single-source) [Fact + single-source];
   • Domestic in-ecosystem API route: the Qwen app connects to Taobao/Alipay/Fliggy/Amap, with 400+ AI errand features (2026-01) [Authoritative media via The Paper];
   • Domestic system-level GUI route: the Doubao phone-assistant consumer edition (2026-09) + SAEP protocol [Authoritative media via The Paper];
   • Open-source self-hosting: Chinese open models like Qwen/GLM/DeepSeek plus Muse Glimmer (if Apache 2.0 holds — single-source) for developers to build their own agents.
4. Policy risk note (analysis): the Manus case shows China treats autonomous-agent technology as a sensitive asset with restricted cross-border M&A; under data-localization requirements, Meta's personal agent has no short-term path to the mainland market; Chinese users of Muse must also consider data-export compliance and account-risk exposure.

Timeline of Key Events

Unconfirmed Items

1. Muse pricing tiers (free / $20 / $100 and token caps): only weak sources like tech-insider; Meta's help-center "subscriptions" page was not verified item by item — defer to Meta's official help center.
2. Muse downloads (2.5M in two weeks, topping the App Store): only a second-hand citation of Sensor Tower via the LA Times; no official Meta figures.
3. Muse Glimmer (29.6B parameters, Apache 2.0): mentioned only by bit.com/businessmodelanalyst; no Meta official press release seen.
4. Tencent et al. buying back Manus for $2B: the Caixin report explicitly states "unconfirmed by either side."
5. MSL's 2025-10 layoff of 600, Meta's 2026-05 layoff of 8,000: weak sources, no Meta official confirmation seen.
6. Manus's $500M raise at a $4B valuation: Bloomberg's 2026-09-17 report said it was "not yet completed."
7. "Jony Ive involved in designing Meta's new AI assistant": mentioned only by techdefused, questionable, not used.
8. Meta recruiting Signal's founder for Muse's confidential VM: mentioned only in one podcast summary, not cross-verified.
9. $567M child-safety court fine: mentioned only by TechCrunch, no judgment document or second source seen.
10. Quantified goals like "billions of personal agents within years": seen only in video-description relays, no official original text.
11. This research did not obtain Meta's official "Introducing Muse" press release (the specific about.fb.com URL did not appear complete in search results; no guessed visits were made); Muse feature descriptions are based on the official help center (verified live) cross-checked with multiple media outlets.

Source List

Official primary sources (verified live / read in full)

• Zuckerberg's open letter, full text: https://www.meta.com/superintelligence/ (2025-07-30; index read in full)
• Meta Muse help center (zh-cn): https://www.meta.com/zh-cn/help/artificial-intelligence/1303670544995562/ (verified live, 2026-10-05; confirmed regional restrictions and agent definition; meta.com navigation includes Meta VR Glasses / Muse Charm)
• MSL memo (2025-06-30): via full or extensive quotes from Fortune (Sharon Goldman) / Bloomberg / CNBC
• Compute Threads statement (2025-07-14): via Reuters https://www.reuters.com/business/zuckerberg-says-meta-will-invest-hundreds-billions-superintelligence-2025-07-14/ reporting
• "The Future is for Everyone" (2026-08-10): published on the Meta newsroom, confirmed via TechCrunch https://techcrunch.com/2026/08/10/mark-zuckerbergs-ai-manifesto-is-exactly-why-people-dont-like-ai/
• Ray-Ban Meta Japan launch (with personal superintelligence wording): https://about.fb.com/news/2026/05/ray-ban-meta-ai-glasses-launch-in-japan/

Authoritative media

• Fast Company long-form investigation (2026-09): https://www.fastcompany.com/91587899/mark-zuckerberg-meta-ai-personal-superintelligence
• Reuters Muse "human concierge" exclusive (2026-09-22): via https://www.lapost.com/content/meta-testing-a-human-concierge-for-its-new-personal-ai-agent-muse reprint; https://www.techspot.com/news/113957-meta-muse-ai-used-human-contractors-calls-including.html
• Manus independence: https://technode.global/2026/08/12/singapore-based-manus-ai-to-leave-meta-as-china-blocks-metas-2b-acquisition/ ; deal-unwinding background https://www.indexbox.io/blog/meta-cuts-ties-with-manus-after-china-orders-2b-acquisition-reversal/ ; buyback rumor https://www.ebc.com/forex/manus-independent-meta-2-billion-deal-unwind (noting Caixin unconfirmed)
• Compute: https://www.datacenterdynamics.com/en/news/meta-to-invest-hundreds-of-billions-of-dollars-into-compute-to-build-superintelligence-with-several-multi-gw-data-center-clusters/ ; https://siliconangle.com/2026/07/13/meta-boosts-investment-hyperion-data-center-campus-50b/ ; https://datacentremagazine.com/news/inside-metas-data-centre-strategy-for-ai-infrastructure
• Connect 2026: https://virtual.reality.news/news/meta-connect-2026-live-blog-muse-ray-ban-vr-glasses/ ; https://pulse2.com/meta-unveils-1299-99-vr-glasses-expands-muse-ai-and-introduces-more-than-100-ai-glasses-styles/amp/
• Glasses market: IDC data via https://glassalmanac.com series citing Reuters (2026-01-06 reporting)
• Competitors: Altman's "Gentle Singularity" via https://www.startuphub.ai/ai-news/ai-figures/2026/figure-sam-altman-public-position-evolution-2026-06-13 etc.; Amodei's "Machines of Loving Grace" via https://github.com/martinkuba/ai-notes/blob/HEAD/summaries/dario-amodei-machines-of-loving-grace.md ; Hassabis via https://en.sedaily.com/international/2026/05/27/alphagos-father-closes-google-io-signaling-ai-power-shift , https://www.thehindubusinessline.com/info-tech/deepminds-demis-hassabis-predicts-agi-within-five-years/article70651735.ece ; Musk via https://www.duck9.com/blog/elon-musk-how-to-start-up/ etc.
• China perspective: https://news.marsbit.co/20261004113915252511.html (2026-10-04); The Paper via https://k.sina.cn/article_5044281310_12ca99fde02002l2lo.html ; V2EX signup tutorial https://www.v2ex.com/t/1244732
• Muse privacy/security: https://www.livemint.com/ai/meta-launches-muse-ai-agent-can-users-trust-it-with-email-payments-and-personal-data-11788929192329.html ; NPR https://www.knkx.org/2026-09-30/metas-muse-killer-app-security-nightmare-both
• Manifesto reception: https://www.fool.com/investing/2026/08/23/mark-zuckerberg-published-a-6500-word-ai-manifesto/ ; https://www.bit.com/knowledge-hub/zuckerberg-ai-manifesto

Single-source / weak sources (flagged as questionable in the report): tech-insider.org (pricing), insearchofyourpassions.com (Muse review), mexc.com/webpronews (Muse Spark scores), implicator.ai (hybrid open-source), business20channel.tv (checkpoint proposal), financefeeds.com (8,000 layoffs), currato.com (NYT interview relay), sources.news (podcast summary), nerdheadz.com (Manus funding), businessmodelanalyst.com (poll and manifesto criticism), misryoum.com (data-center criticism).