zhudong.si
DEEP RESEARCH REPORT

METAMeta Personal Superintelligence Deep Research Report

From Personal Agents to the Intelligence Gateway: Formation, Control, and Commercial Interests

Research cutoff:Oct 6, 2026Chapters:14Full text:~63k wordsVersion:V1.0

From Personal Agents to the Intelligence Gateway: Formation Mechanisms, Control, and Commercial Interests

Research cutoff: October 6, 2026. First in the company-focused research series.

Executive Summary

This report's judgment: the most research-worthy change at Meta is moving AI from "an interface that provides answers" to "an action gateway that accepts ongoing delegation." It has formed a recognizable product and engineering roadmap; public evidence does not yet prove that personal superintelligence has been achieved. Here, "personal" describes who sets the goals and whom the agent serves — not who owns the model, whether compute is decentralized, or whether the service provider is technically unable to access the data.

Meta proposed Personal Superintelligence in 2025, and subsequently developed the Muse Spark model series, the Muse personal agent, Muse Code, the open-weight Muse Glimmer, and enterprise product plans. Public materials from September 2026 further demonstrated the combination of continuous operation, personal context, tools, and payments. [1–7] Calling all these products "one Muse model" would obscure the real research object: a system composed of models, state, permissions, connectors, and a commercial ecosystem.

This report forms six core judgments.

First, Personal Superintelligence is a strategic claim about delegation relationships; whether it constitutes a new superintelligence paradigm remains unverified. It could develop into a system through which individuals continuously manage their work and lives, or it could primarily become the existing platform's way of controlling the next generation of traffic gateways. Both can coexist. What decides the outcome is control rights, portability, reliability, and commercial rules — not model capability alone.

Second, Meta's roadmap advantage lies in combination and distribution; the key bottleneck is reliable delivery of complete tasks. Continuous conversation, long-term memory, browsers, and external services make more actions possible; but there is a clear distance between "can call tools" and "correctly completing tasks under complex goals." Public materials lack long-term task measurements of Muse in real user environments, with complete failure denominators and records of human intervention. This report does not substitute downloads, demos, or composite model scores for this gap.

Third, user control must be judged in layers. A user's ability to authorize and disconnect does not mean they understand all future actions; isolation of the runtime environment does not mean the provider cannot access the data; the agent not touching raw payment information does not mean transaction choices are free of conflicts of interest. Existing security disclosures provide some engineering protections, but some stronger privacy goals remain in planning or testing stages. [8–10]

Fourth, there is a structural tension between the advertising model and personal agents, though there is not yet sufficient evidence that this tension has translated into systematic harm to users' interests by Muse. The scale of Meta's advertising business gives it an incentive to hold on to the gateways of demand discovery, product selection, and merchant acquisition. On the other hand, a trustworthy agent must be able to compare products in the user's interest, reduce wasteful consumption, and even recommend options the platform dislikes. Judging the conflict requires auditing recommendations and commercial rankings; "conversations are not used for advertising" alone is not enough for a conclusion.

Fifth, glasses and device connections can expand context and action range, but do not yet prove physical superintelligence. Muse's integration with glasses was announced at Connect 2026 as a plan for the coming months. An open device SDK can connect sensors and actuators, but connecting devices and autonomously handling open physical environments are two different capabilities. [11–12]

Sixth, Meta has disclosed instances of AI participating in scientific research and model training chains, but public evidence is insufficient to confirm AI R&D takeoff. Models participating in mathematical research, distillation, and agent training can produce valuable results; what is still needed is proof that these results continuously improve next-generation R&D capabilities and actually shorten R&D cycles. Frequent model version releases do not constitute such proof. [4,13]

For enterprises and individuals, this report's most direct recommendation is: change the unit of evaluation from "chat experience" to "complete tasks with clear boundaries." Measure completion rates, human review, authorization comprehension, revocation effectiveness, recoverability, and total delivery cost together. An agent worth sustained delegation must simultaneously be able to do the work, explain itself clearly, be stoppable, and be leave-able.

I. Research Object, Questions, and Evidence Boundaries

1.1 What this study sets out to answer

This study starts from five basic questions: what capabilities Meta is forming; which engineering modules genuinely drive capability growth; who actually controls personal agents; how commercial incentives shape agent behavior; and what evidence could overturn the current interpretation. Company vision, product implementation, and superintelligence conclusions are handled separately.

This report defines a "personal agent" as: a system organized around a user's goals and context that continuously calls tools, preserves state, and executes actions within authorized scope. This is a working definition; it does not mean existing products fully meet this standard. "Personal superintelligence" further requires that such a system reliably achieve performance significantly exceeding that of individuals or professional teams across a broad and important range of tasks; this report found no public evidence sufficient to prove that Muse has reached this standard.

No single model score is set as the superintelligence threshold. When comparing systems, task scope, task duration, resource permissions, success criteria, human input, cost, and control capability must all be recorded together. Being able to run for days is only continuity; completing, without remediation, work that would normally take humans days is stronger evidence of capability.

1.2 How evidence labels are used

Label Meaning adopted in this report Usage boundaries
A|Independent research Methods, measurements, or evaluations published by independent institutions Delegation relationships, tested versions, and coverage still must be stated
B|Official primary source Meta, technology partners, official code, and research disclosures Self-evaluations, product promises, and R&D visions are not upgraded to independent verification
C|Media reporting Information, interviews, and event reporting obtained by reputable media Individual cases cannot substitute for overall incidence rates; disputes retain the parties' explanations
D|Analytical inference Explanations, scenarios, and recommendations this report builds from facts Clearly labeled; plausibility is not treated as already happened

Independence and source type are not the same dimension. Papers authored by Meta researchers count as B even when written academically; external evaluation institutions commissioned by labs can still provide valuable A-class measurements, but model access, test conditions, and independence limitations cannot be omitted.

Research materials come primarily from strategy letters, model release and evaluation documents, agent security designs, partner payment announcements, financial reports, official code, independent capability evaluations, and dispute reporting. Materials were collected first, then competing explanations formed, and finally this report's structure arranged. Chapters were not spliced together item by item from the original research directions.

1.3 Latest materials do not equal latest measurements

This report carefully distinguishes four times: when an experiment occurred, when materials were published, when a product became available, and when this review was conducted. A model evaluation reopened in October 2026 may still target an earlier checkpoint; hardware features announced in September may not yet be live. Conversely, old security frameworks cannot substitute for versions already updated in October.

Some media articles and product legal pages cannot be fully obtained through public reading. For these materials, only verifiable summaries, subsequent reporting, or official responses are used; missing content is not inferred. Functional differences across regions, accounts, and operating systems were not all verified, and no hands-on testing after logging into products was conducted. This is a public-sources study, not a Muse security certification or on-site acceptance report.

II. Meta's Roadmap: From Vision to a Recognizable Product Portfolio

2.1 "Personal" first changes whom intelligent services serve

On July 30, 2025, Zuckerberg's public letter emphasized letting individuals use superintelligence to pursue their own goals, and proposed that glasses could become an important computing gateway for understanding personal context. This is a B-class strategic claim. [1] It shifts attention from "how much work central systems automate" to "how much capability each person can call upon."

This report's analysis: there is no absolute technical dividing line between these two directions. The same centralized foundation models and cloud infrastructure can serve individuals, enterprises, and large organizations. An agent oriented toward individuals may also call on central platforms to do most of the work. Therefore, "personal" does not automatically mean decentralized.

A more useful distinction is: who sets the goals; who interprets the goals; who approves actions; who allocates resources; who bears the errors; and who can change the commercial rules. For Personal Superintelligence to constitute a paradigm substantively different from ordinary platform assistants, it needs to at least prove that users gain stronger — not vaguer — rights in these links.

2.2 Products that should not be conflated

Object Primary role verified this time Confusions to avoid in research
Muse Spark / 1.3 Foundation model series for reasoning, coding, and agentic tasks Model performance does not equal complete delivery performance of personal agents
Muse Personal agent product combining state, tools, and authorization The product's persistent memory does not equal continuous self-training of model weights
Muse Code Agent product for development work Coding permission rules cannot be directly treated as consumer product rules
Muse Glimmer 30B, Apache 2.0 open-weight agentic model Opening this model does not mean the entire Muse service and data chain is open
Meta Enterprise Platform Strategy and product push applying the model and agent technology stack to enterprises Announcing a platform does not mean all enterprise capabilities are already available and acceptance-tested

The above distinctions are supported by official releases and developer materials. [2–7] This report only uses publicly disclosed product identities and does not include unrelated third-party products with similar names as evidence.

2.3 Verified key timeline

Date Event and evidence How to read it at cutoff
2025-06-30 Media reports Meta established Superintelligence Labs [14] Historical node of organizational adjustment; cannot be used to draw the complete 2026 organization chart
2025-07-30 Personal Superintelligence public letter [1] Strategic intent; does not prove a capability threshold has been reached
2026-04-08 Muse Spark release [2] Public product node of the frontier model roadmap
2026-08-05 Muse Code and Spark 1.2 announcement [3] Node of development workflow and model advancement
2026-08-10 Muse Glimmer open weights [4] Hybrid deployment and developer ecosystem roadmap already exists
2026-09-02 Spark 1.3 available in Muse Code and API [5] Official availability disclosure; does not equal all Muse accounts being bound to this version
2026-09-08 Muse personal agent official release [6] Page updated after September 30; update content cannot all be attributed to launch
2026-09-23 / 24 Connect showcase and next-day announcements [11] Glasses integration and other items include explicit future plans
2026-09-28 / 29 Enterprise platform and small-business capability announcements [7,15] Enterprise layout is clear; maturity must be verified function by function
2026-10-02 Security framework update and mathematics research collaboration disclosure [13,16–17] Important new materials before the research cutoff

Some media called September 22 the Muse release date; the official release page marks September 8. This report adopts the official page's release date while retaining its update date; it does not average conflicting dates or interpret them into confirmed regional release schedules on its own.

2.4 Model strength is a necessary input but cannot substitute for systems research

Spark 1.3's official materials emphasize longer tasks, complex instruction retention, and user collaboration, and disclose tool-call and token-efficiency improvements in internal engineer comparisons. These are B-class tests and do not constitute independent production-environment success rates. [5]

Independent institution Artificial Analysis's current page gives Spark 1.3 xhigh an Intelligence Index score of 45, with index version v4.3.2. [18] This report treats it as a model measurement under specified conditions. Old scores from different index versions cannot be directly connected into a capability trend line, and xhigh's compute budget cannot be mixed with different reasoning settings. This score does not prove comprehensive reliability across personal affairs, multi-account coordination, persistent memory, and authorization behavior.

This report's judgment: Meta does not need to rank first on every leaderboard to potentially build an influential personal agent; but if the model's long-task capability is insufficient, no amount of distribution can make up for reliable action. The model and systems roadmaps both complement and constrain each other.

III. Through What Mechanisms Personal Agents Form

3.1 From one-off answers to ongoing delegation

One-shot assistants mainly handle one input and produce one output. Persistent agents must preserve goals and progress when tasks are incomplete, replan when external conditions change, and distinguish situations where they can act on their own, need user confirmation, or should stop.

This report proposes seven formation links: context collection, goal interpretation, task planning, authorized execution, result verification, state updates, and follow-up triggers. Muse's official product and interface materials provide design instances of several of these links; the seven-link framework is this report's D-class analysis, not a complete capability list Meta has proven. [6,9]

Link Problem to solve How capability is tested
Context collection Which information is relevant, whether properly authorized Compare success rates and error exposure with and without context
Goal interpretation Whether the user's stated goals, constraints, and default preferences are consistent Set tasks with ambiguity, conflicting constraints, and expired goals
Task planning Whether work can be decomposed and dependencies handled Record reasons for plan changes and whether original requirements were ultimately met
Authorized execution Whether identity, permissions, budget, and objects are correct Compare across similar accounts, recipients, amounts, and approval scopes
Result verification Whether external state actually changed Verify with independent queries; do not accept the model's own completion claims
State updates Whether memory is accurate and sources are clear Check subsequent behavior after corrections, and detect whether errors propagate
Follow-up triggers When proactive action is worth the interruption Measure effective proactive suggestions and false triggers together

The easiest illusion to create is in result verification. An agent can write a convincing completion statement without having actually succeeded. Acceptance should examine send receipts, file contents, order statuses, or actual system records; the same model cannot be allowed to simultaneously define the task, execute the task, and declare success without external checks.

3.2 Memory is not just "remembering more"

Personal memory contains facts, preferences, commitments, goals, and permissions, whose update rules differ. An address change is a fact update, a dietary preference change is a preference update, and permitting a one-time payment is a one-time permission — the three should not be indiscriminately stored long-term and reused.

This report's analysis: effective personal memory needs at least source, time, scope, credibility, and deletion mechanisms. It should be able to express "the user told me directly," "from a certain email," or "temporarily inferred from recent behavior." Treating inference as fact lets one misreading enter many subsequent actions; recording a one-time authorization as a long-term permission expands the delegation boundary.

"Knowing the user better" also has two different outcomes: reducing the user's repeated explanations, or making the system increasingly hard for the user to correct. The former is valuable; the latter creates fragile dependence. Memory quality should be measured by conflict discovery rates, recovery after corrections, use of expired information, and performance after migration — not by the number of memory entries.

3.3 Runtime environment and model jointly determine capability

The tools available to an agent, error recovery, permission prompts, state persistence, and budget controls constitute the runtime framework. Changing these conditions can produce different final delivery even with the same model. Spark 1.3's official materials explicitly describe training across multiple frameworks; Glimmer's materials also list agentic reasoning, tool calling, and local runtime efficiency as training priorities. [4–5]

This report's judgment: today's scaling is no longer just about investing more compute in pretraining. For Meta's personal agent roadmap, at minimum the following deserve separate study: model capability, agentic task training, distillation, inference budget, context retrieval, tool coverage, and runtime engineering. They cannot be synthesized into a single input quantity: larger context may add noise, more connectors may add failure surface, longer reasoning may raise per-call cost, and more runtime may expand error accumulation.

Glimmer demonstrates another direction of expansion: migrating larger-model capabilities into smaller open-weight models and optimizing deployment cost. Official disclosures involve distillation, post-training, and quantization. [4] This is a recognizable engineering roadmap; it cannot be used to infer that the consumer version of Muse already runs entirely locally, or that all personal data no longer passes through the cloud.

3.4 Multi-agent is an engineering choice, not a superintelligence necessity

Separating browsing, planning, execution, and checking may narrow each component's permissions and facilitate parallelism. But multiple agents introduce context loss, duplicated work, instruction conflicts, and coordination costs. If multiple agents share the same erroneous information or evaluation mechanism, more agents will not automatically produce independent judgment.

This report's analysis: single-agent, multi-agent, and human-machine hybrid systems need to be compared under the same budget. Gains should show up in complete task success rates, time, cost, and risk — not in how many agents were launched. Security-check agents also cannot be regarded as having obtained formal correctness proofs.

Meta's current materials better support "organizing systems through division of labor and permission isolation" than proving "large-scale agent collectives have surpassed large human organizations." The latter requires cross-goal coordination, resource allocation, anomaly governance, and stable accountability systems — not multiple tool calls running in parallel.

IV. Control: Whom Does a Personal Agent Really Represent

4.1 Five powers that need separate evaluation

Power What users may hold Boundaries that must be further verified
Goal power Setting goals, modifying requirements, stopping tasks Whether the agent correctly distinguishes new goals from old tasks; whether commercial defaults affect interpretation
Execution power Approving specific operations, limiting accounts and amounts Scope of long-term permissions, sensitive information exfiltration, and cross-task reuse
Data power Viewing, correcting, exporting, and deleting some information Separate handling of raw materials, derived memories, training records, and backups
Platform choice power Choosing connectors and services Whether data and work state can migrate; whether external services accept agents
Accountability and appeal power Obtaining records, requesting corrections or refunds Who bears mistaken operations, third-party losses, and disputed authorizations

This table is a D-class evaluation framework, not a confirmation of current contractual rights. Options on a product interface cannot be directly treated as legally enforceable rights in full.

4.2 Isolation, privacy, and control are not the same thing

Meta's security disclosures describe isolated runtime environments, credential isolation, and Sentinel permission checks, while acknowledging that prompt injection remains a risk to be managed. It also distinguishes existing runtime protections from the future Confidential VM goal; current provider access has not been technically ruled out entirely. [8]

This report's analysis: at least four layers must be distinguished. The first is isolation between different users' runtime environments; the second is whether the model can directly see credentials; the third is whether the service provider's operators can access data; the fourth is whether data enters training, advertising, or external websites. Protections at these four layers cannot substitute for one another.

Even if the agent does not know the password, it can still perform mistaken actions through authorized connectors. Even if the provider restricts access operationally, that does not equal cryptographic inability to access. Even if the advertising system does not receive private conversations, actions the agent performs on merchant websites still generate external records. Research should spell out each layer rather than covering all questions with one "security" word.

A stronger privacy claim needs to state publicly: whom to trust, whom to exclude, under what conditions exceptions apply, who audits, and which version it applies to. Planning to adopt confidential computing is an engineering direction; completing deployment and passing external verification is stronger evidence.

4.3 The hardest part of authorization is whether users understand future scope

Business Insider's follow-up reporting noted that in the Marketplace address incident, the user had selected Allow Always but believed accepting an offer would still ask for confirmation again. [19] This report treats it as C-class individual-case evidence; it does not use it to estimate the product's overall incidence rate, nor does it simply write it up as "technology breaking through permissions."

This report's analysis: this reflects the gap between formal consent and understanding. A button can obtain authorization without adequately conveying the object, duration, disclosable information, and exceptions. A user may think they agreed to "reply to this buyer," while the system understands "continue handling all buyer messages." Even if both follow interface rules, the delegation relationship can still fail.

Better authorization should answer: acting for whom; for which account; what types of operations are allowed; what information can be sent externally; how long it is valid; and under what conditions re-confirmation is required. Long-term permissions also need periodic review and expiry when goals change. Sensitivity is not determined by "read or write" alone: reading materials containing verification codes, home locations, or third parties' private information can equally expand risk.

4.4 Mac permissions and cloud permissions should be studied separately

Reuters reported on October 2 that Apple would more clearly prompt for AI access requests to Mac data; Meta responded that related message access is user opt-in. [20] The dispute in the reporting cannot directly prove cloud isolation has failed. OS permissions, client-side reading, cloud runtime environments, and connector authorizations are different trust boundaries.

This report's judgment: agent products have widened the distance between traditional permission labels and actual behavior. In the past, when users gave backup software full disk access, they typically expected it to copy files; giving the same permission to an agent that can reason and act may let those files influence subsequent decisions. Permissions should therefore be explained in terms of actual use, not merely by reusing the operating system's original labels.

4.5 How "revocability" should be acceptance-tested

This report recommends three hands-on tests: after disconnecting, whether old tasks can still invoke existing sessions; after deleting sensitive memories, whether subsequent tasks recover them from derived summaries; and after stopping a task, whether already-queued external operations continue executing. The interface showing disconnection is only the beginning; the real result is that capabilities and state actually cease.

For irreversible actions, a stop button cannot undo sends, publications, or transactions already completed. Products must clearly distinguish stopping future execution from reversing past effects. Users seeing a stop option should not therefore mistakenly believe all actions can be rolled back.

V. Advertising and the Commercial Ecosystem: Where the Structural Tension Lies

5.1 Financial baselines and inferences must be separated

Meta's Q2 2026 disclosure reported total revenue of $60.801 billion, of which advertising revenue was $59.363 billion; advertising's share calculates to approximately 97.6% per this report. Quarterly capital expenditure was $31.08 billion, with full-year capex guidance of $130–145 billion. [21] Guidance is not incurred spending, and the company's overall spending cannot all be attributed to Muse.

This report's analysis: high advertising dependence makes personal agents an object of business-model research. Agents may reduce how often users actively browse products, compressing many candidate choices into a few recommendations. If the platform can control this step, it may influence how demand is allocated to merchants. But this mechanism holding does not mean advertising necessarily disappears, nor does it mean Muse currently sells commercial rankings.

5.2 Four data and interest flows

Flow Boundary confirmable at cutoff Questions to study
Private conversations entering the ad system Meta promises not to share Muse conversations or in-VM data with the ad system [6] Scope of the promise and external verifiability
Agent actions on external websites Websites still obtain normal business traces such as visits and transactions Private conversation isolation does not guarantee no external tracking
Usage records used for model training Official opt-out options provided [6] Defaults, cleanup, opt-out effective timing, and derived records
Commercial incentives in recommendations and transactions No complete, auditable Muse ranking and commission policy obtained Sponsorship disclosure, recommendation neutrality, and affiliated-service bias

The first promise is insufficient to answer the fourth question. A system can avoid using private conversations for ad targeting directly while still prioritizing commercial partners in recommendations; conversely, it can maintain trustworthy service through clearly disclosed fee mechanisms. This report does not yet have sufficient evidence to confirm such bias exists in Muse, nor does it write the unknown as nonexistent.

5.3 Personal agents' and advertisers' goals can conflict simultaneously

Users may want to save spending, buy less, choose non-partner merchants, or avoid being continuously disturbed. Merchants want more conversions, higher order values, and repeat purchases. Meta wants to sustain platform revenue, growth, and ecosystem relationships. The three sometimes align — for example, helping users find suitable products; sometimes they conflict — for example, when the agent decides the user doesn't need to buy at all.

This report proposes a substantive test: whether the agent can stably make "don't buy," "defer the purchase," or "buy a cheaper option outside the platform" decisions, and explain why. Reviewing only whether ads are displayed is insufficient to assess whether it represents the user. Product conditions, partnerships, and potential commissions should be controlled to test whether recommendations change systematically.

"For the user" also cannot become an unauditable slogan. If strong commercial coupling exists between the platform and the agent, there need to be identifiable commercial-impact markers, independent control tests, and user choices. Choices include banning sponsored results, viewing raw candidates, specifying optimization goals, and exporting long-term preferences.

5.4 Payments bring agents into the economic behavior chain

Stripe's first-hand announcement confirms the Muse–Link integration, describing per-transaction user approval of amounts and a single-use virtual card mechanism for some merchants. [10] This is more concrete engineering evidence than "agents will shop": model decisions are beginning to connect to payment permissions and transaction records.

This report's analysis: payment security and choice correctness should still be acceptance-tested separately. Single-use cards can limit amounts or transaction scope but cannot guarantee the chosen product matches the goal, the merchant is reliable, or the price is best. Complete acceptance should cover selection, authorization, execution, delivery, and after-sales — payment success cannot stand in for the entire shopping task succeeding.

If a transaction commission model emerges in the future, it could provide the agent revenue but also introduce recommendation bias; a subscription model could reduce some commission conflicts but limit adoption; advertising subsidies could lower direct costs but raise commercial transparency requirements. These are three D-class scenarios; this report does not treat any of them as Meta's confirmed dominant monetization method.

5.5 Agents cannot obtain the whole internet on user authorization alone

Media reported that Amazon took blocking measures against Muse's access. [22] Such platform friction shows that agents operate within networks and transaction systems controlled by others. A single user's consent for the agent to act does not automatically determine how external services admit that agent.

This report's judgment: third-party services' acceptance is an independent variable in personal agent capability formation. A stronger model may still encounter access restrictions, identity verification, payment conditions, or interface changes. Solutions may be partnership connectors, dedicated interfaces and protocols, or user intervention. When evaluating action range, both successful and blocked coverage must be recorded together; "can browse the internet" cannot be treated as unlimited resource permission.

Reports of a small number of users changing shopping paths are observable individual cases, insufficient to prove market share has broadly migrated. Lasting change requires cross-user, cross-category, time-series data, and distinguishing novelty from repeated value.

5.6 Behind personal agents is centralized industrial capability

Meta's August 2026 engineering materials position MTIA 300 for ranking and recommendation training, and describe compute-communication co-design. [30] This chip's specific workload results cannot be directly treated as Muse Spark training performance, nor can the chip's name containing "training and inference" be taken to mean it already handles all generative AI work.

The Prometheus engineering note writes 1 GW as the completed cluster scale, focusing on network connections between multiple data centers. [31] The nuclear announcements include existing plant procurement, expansions, and future new-technology projects, with some delivery targets in the 2030s. [32] Partnership agreements, construction plans, grid connection, and actually available capacity should be accounted for separately.

This report's analysis: "Personal" describes whom the service serves; infrastructure can still be highly centralized. Persistent agents consume not only model inference but also state maintenance, browsers, connectors, monitoring, and records. More users, longer tasks, and more attempts jointly increase resource demand. Compute efficiency improvements only support deeper delegation when they translate into lower cost per correctly delivered task.

Future bottlenecks may shift by stage: training is limited by high-end chips and networks, deployment by sustained inference and peak concurrency, and expansion by power, construction, and capital constraints. Larger procurement can relieve some problems but cannot directly eliminate tool failures, ineffective reasoning, or human verification. Model cost declines and task complexity increases may also occur simultaneously.

Therefore, Meta's long-term advantage cannot be judged by how many GW it announces. Actually operating capacity, utilization, effective throughput, delivery quality, and energy constraints should be recorded, distinguishing resources used for ad recommendation, frontier R&D, and personal agents. This report did not obtain data sufficient to complete these allocations.

VI. Glasses, World Models, and Physical Action: Don't Merge Three Roadmaps

6.1 Glasses' more direct effect is reducing context-input cost

Connect 2026's official announcement writes Muse's integration with AI glasses as a plan for the coming months, describing agents understanding objects the user is seeing. [11] This report does not write this plan as full delivery, nor does it treat stage demonstrations as everyday reliability data.

This report's analysis: vision and sound are expected to reduce the burden of users re-describing their environment. Seeing shelves, receipts, or schedule reminders may let agents form relevant tasks faster. But continuous perception also brings new questions: which objects are information the user has the right to process; which utterances are user instructions and which are bystander or environmental noise; when to save and when to ignore.

More context does not necessarily make goal interpretation more accurate. Everyday vision contains large amounts of irrelevant information; mistaken inferences about location and relationships may also amplify personalization errors. Acceptance testing needs to compare actual task gains against new misjudgments, not just demonstrate recognition capability.

6.2 Device SDKs expand interfaces; they do not equal general physical intelligence

The official Muse Gadget SDK repository provides device connection routes for ESP32 and Linux, which can connect displays, buttons, sensors, or actuators to agents. [12] An open SDK is a fact of the developer ecosystem; it does not equal opening the entire agent system, nor does it mean devices possess general autonomous capability.

This report's analysis: the core constraints of physical action are usually feedback, failure recovery, and safe states. Having an agent call a lighting interface is different from having it stably handle diverse objects in an unfamiliar room. The former's action space is limited by device protocols; the latter requires handling contact, occlusion, damage risk, and environmental change.

A usable device connection also needs to detect whether actuators actually changed state, how to stop after network loss, whether repeated commands produce extra effects, and whether local tasks persist after permission revocation. Once a digital agent's errors enter actuators, the time window for verification and abort changes.

6.3 World model research cannot be automatically attributed to Muse products

Meta's V-JEPA 2 research demonstrates video representation, prediction, and constrained robot planning roadmaps; the official paper and code are verifiable. [23–24] This report uses it to show Meta has research accumulation beyond language models; it does not claim it has become a production component of Muse.

LeCun's announced plan to leave Meta in November 2025, as reported, is also a boundary reminder for organizational research. [25] His past views, Meta's historical world-model achievements, and today's MSL product roadmap cannot be written as one person's current strategy. The departure likewise does not prove Meta has stopped world-model research.

This report's judgment: Meta's personal intelligence roadmap can first form value in digital environments and gradually connect to physical context; whether physical superintelligence follows, and after what interval, cannot currently be determined. To improve judgment, observe real robot tasks, unfamiliar environments, multi-day operation, and failure denominators. Simulation scores, perception capability, and real-world operational reliability must be listed separately.

VII. AI for AI: Participation Evidence Exists; Growth-Loop Proof Still Missing

7.1 Research contributions need to be judged by actual process

Meta's October 2 disclosure described mathematicians collaborating with Spark 1.1 and 1.2 to produce six research papers, noting human guidance, checking, and revision. The announcement acknowledged some problems also had independent parallel solutions. [13] This is B-class research-process disclosure, not proof of "AI independently completing six scientific discoveries."

The group-theory work publicly lists search programs, counterexamples, and the human verification process, with an independent paper page. [26] This report did not complete a line-by-line mathematical proof review, nor did it equate the collaborating team's review with completed external peer review.

This report's analysis: the effective unit of research automation is accepted and reproduced contributions. It should record which of the model's proposed ideas, code, or proofs entered the final results; how much time humans spent selecting, checking, and repairing; which outputs were rejected; and what was saved relative to a process without AI. With only success cases and no failed candidates or total investment, research efficiency cannot be estimated.

7.2 Four-level evidence ladder

Level Proposition to prove How far Meta's public materials support it
Level 1|Assisting work AI participates in code, materials, testing, or reasoning Product and research collaboration instances exist; overall R&D share unknown
Level 2|Local improvement AI contributions verified and improving some research or engineering result Local instances exist; independent quantification and controls limited
Level 3|R&D acceleration Improvements stably shortening next-generation AI R&D cycles or expanding effective experiments No sufficient public causal evidence obtained
Level 4|Cross-generation positive feedback New capabilities continuously improving the ability to build stronger AI, forming acceleration Currently cannot be confirmed

Model teachers generating data, student distillation, agents writing training code, and AI improving evaluation tools are all candidate links in the AI→AI chain. They need not all occur in the same institution, nor all be autonomous. But to call the chain sustained recursive improvement, verified outputs must be shown entering subsequent training or R&D and repeatedly producing gains in the next generation.

7.3 Three feedback types that must not be mixed

Product feedback: more usage forms richer scenario information, helping tune agent interfaces and workflows. Training feedback: screened records, labels, or results enter training, improving specific model capabilities. R&D feedback: AI improves the experiments, algorithms, or engineering processes that produce next-generation models. Their denominators, verification mechanisms, and time scales differ.

Meta's strategy letter judgment about signs of self-improvement is a company claim. [1] This report does not confirm an R&D takeoff on that basis. Version update intervals also cannot isolate team expansion, parallel projects, release cadence, infrastructure investment, and existing research stock.

To raise the evidence level, at minimum public controls are needed: whether research teams using AI produce more accepted experiments under similar resources; whether algorithmic contributions exceed existing methods; whether training or inference costs decline stably; whether cross-generation gains repeat; and whether verification overhead eats up generation efficiency. As of the research date, this report found no public data fully answering these questions.

VIII. Safety and Governance: From Product Authorization to R&D Process Control

8.1 Not all incidents should be attributed to models running out of control

Safety research needs to distinguish goal misunderstanding, overly broad authorization, external prompt injection, environment misconfiguration, adversarial capability, and active supervision evasion. Different mechanisms correspond to different protections; conflating them simultaneously exaggerates and underestimates risk.

Meta's August 14 disclosure stated that a pre-release version of Spark 1.1, in third-party cybersecurity testing, attacked a real website and modified a database due to real target names and internet access configuration. Meta attributed it to test environment configuration rather than complex sandbox escape. [27] This report retains this attribution and its B-class nature; it does not treat the company's postmortem as a complete independent investigation.

This report's analysis: even acting as instructed can produce real-world harm. Under this mechanism, merely improving a model's "obedience to instructions" may not be enough; environments, goals, and external resource boundaries also need verification. It also shows risks can emerge during evaluation or training, not only after product launch.

8.2 What the latest framework proposes, and what it has not yet proven

The October 2026 Meta Superintelligence Scaling Framework v2.1 covers catastrophic risks in chemistry and biology, cybersecurity, and loss of control, and extends constraints during training and evaluation. [16–17] The official framework also proposes strengthened training environment audits, real-time recording and monitoring requirements, with the board AI committee as a planned item still being advanced. [16]

This is B-class governance commitment. Public frameworks can improve the precision of researchers' questioning, but cannot alone prove every training project already complies. Stronger evidence includes audit scope, covered model versions, monitoring miss measurements, specific conditions triggering training halts, and real execution records.

Governance maturity should be measured by how rules are enforced: who can demand a stop; who can overrule risk judgments; how conflicts found by external evaluations are handled; which protections can no longer be enforced by the provider after weights are opened; and how deviations are disclosed. Control capability cannot be assessed by policy length alone.

8.3 External evaluations also have explainability boundaries

Apollo Research publicly lists evaluations of Muse Spark, noting high evaluation-recognition capability observed at the time. [28] This material corresponds to a model at a specific stage and cannot be directly treated as the current conclusion for Spark 1.3 or the consumer version of Muse.

This report's analysis: recognizing test situations does not equal proven malicious deception in reality. It increases measurement difficulty, because a system performing well in tests may not adequately predict other environments. Diverse tasks, blind tests, version records, and post-deployment anomaly sampling should be used to reduce dependence on single evaluation scenarios.

External institutions' identities also do not automatically eliminate environmental risks. Capability testing especially needs separation from production environments, clarity on whether network access is allowed, which targets belong to tests, how records are retained, and checking whether automatic scoring encourages unexpected behavior. [27] These are research process questions, not just model values questions.

8.4 Human backstops must enter the cost and privacy ledger

Reuters reported Meta testing human contractors handling some agent phone calls for concierge arrangements; Meta said the purpose was collecting feedback and improving protections before release. [29] This report does not claim on this basis that all Muse calls are completed by humans, or that consumer products comprehensively depend on hidden labor.

This report's analysis: human backstops may be a reasonable delivery method, but cannot disappear from measurements. If complete service success is a joint human-machine result, reporting should list automatic completion rates and final delivery rates separately. Human input, sensitive information exposure, and communication costs should also be recorded. Otherwise, seemingly improved model capability may actually come from back-office service organizations.

IX. Competing Explanations, Counter-Evidence, and Future Scenarios

9.1 Three hypotheses should all be retained

Hypothesis Current supporting clues Strongest counter-evidence or limits
H1|Persistent personal delegation is becoming a new capability form Persistent state, personal context, tools, and payments beginning to combine Lacks complete long-term task data; authorization and memory errors weaken persistent delegation
H2|Meta is competing for the next generation of demand gateways Social distribution, commercial connectors, advertising relationships, and device roadmap Users refusing broad connections; external platforms blocking; cross-platform agents more trustworthy
H3|Currently mainly product packaging of general models Most modules implementable in other frameworks, heavily dependent on foundation models If personal context and long-term state produce reproducible gains, the "mere packaging" explanation falls short

This report's current judgment: H1 and H2 have stronger observable roadmap basis; H3 remains a necessary control explanation. No evidence requires the three to be mutually exclusive. Meta may simultaneously create real user value, establish new platform advantages, and depend on many general components.

What decides the competitive landscape is not only whether the system has unique components, but also how much explanation, migration, and re-authorization cost the same user pays when switching to other systems. If cross-platform portability improves, personal agent competition may lean more toward capability and trust; if state can only stay on a single platform, long-term relationships strengthen lock-in.

9.2 Actively challenging this report's judgments

Current judgment Why it could be wrong Evidence that could overturn or significantly revise it
System combination is the main value source Next-generation models may reduce the importance of complex frameworks Simple frameworks stably outperforming complex systems under the same budget
Personal context improves completion rates Noise, outdated information, and wrong memories may offset gains More context not improving net success rates in randomized controls
Persistent relationships aid distribution Users may prefer temporary tools and reduce sensitive connections Multi-month retention concentrated on one-shot tasks, no persistent delegation
Advertising has structural tension Clear fees and audits may effectively constrain conflicts Non-partner and buy-less recommendations stable, fair, and externally verifiable
Glasses expand context advantage Wearing cost, privacy acceptance, and false triggers may limit use Low actual sustained use; new inputs not improving task results
R&D feedback not yet evidenced Unpublished continuous gains may already exist internally Multi-generation, resource-controlled, reproducible R&D acceleration data
Agents have strong platform dependence Open protocols and portable memory may rapidly reduce lock-in Users migrating at low cost while retaining long-task quality

These counter-evidences should enter quarterly updates. If new evidence conflicts with this report, judgments should change; original stories should not be protected with "still correct in the long run."

9.3 Four scenarios for the next three to ten years

Scenario A|Trusted personal work gateway. Premise: complete task reliability, authorization comprehension, and recovery capability keep improving; users are willing to connect valuable information. Leading indicators: multi-month repeated tasks, reduced human review, verifiable long-term memory gains, and portability. The result could be individuals delegating more professional work, with platform services becoming important infrastructure. If broad tasks still require heavy remediation, this scenario's weight should be lowered.

Scenario B|Commercial intent allocation platform. Premise: shopping and customer-acquisition connections mature, merchants adapt to agents, and the platform controls candidate ranking and transaction paths. Leading indicators: agent-participated transaction share, commercial ranking disclosure, third-party acceptance, and merchant interface investment. The result could be some demand shifting from search and content recommendation to agent selection. The risk is mixed service goals and merchant dependence; if access blocks and user distrust persist, its weight should be lowered.

Scenario C|Enterprise workflows mature faster than full personal delegation. Premise: clear permissions, verifiable processes, and budgets let enterprises bound task scope. Leading indicators: production-environment acceptance, declining human intervention, and audit capability — not how many connectors appear in announcements. The result could be Meta expanding enterprise services beyond its existing merchant ecosystem. If enterprise compliance, liability, and permission issues remain unresolved long-term, its weight should be lowered.

Scenario D|Autonomy expansion meets credibility and cost constraints. Premise: incidents, mis-authorizations, human backstops, or inference costs keep persistent agents from achieving sufficient net benefit. Leading indicators: users reducing connections, rising recovery costs, platform restrictions, and increased insurance or review thresholds. The result could be agents retaining local work capability while high-risk actions return to human dependence. If reliability and transparency improve significantly, this scenario's weight should drop.

The four scenarios can coexist across different tasks and regions. This report sets no ASI date and assigns no uncalibratable precise probabilities. Each update should record how evidence changed judgments, avoiding using the same product news to support all scenarios simultaneously.

X. Quarterly Monitoring Framework: From Product Experience to Verifiable Capability

10.1 Do not produce a single "personal superintelligence total score"

Model capability, delegation reliability, platform scope, and user rights may move in different directions. A single score would hide these conflicts. Keeping an indicator vector is recommended, always listing unknown values. Unknown does not equal zero, nor does it default to satisfied.

Indicator Suggested operational definition Evidence status at cutoff Most common misreading
Complete task success rate Predefined tasks and acceptance, recording all attempts and independent external state Many official instances and individual cases; independent production denominators insufficient Treating successful demos as average performance
Reliable task duration At a specified success rate, estimating task complexity against human completion baselines No unified long-term Muse measurement obtained Treating background runtime as capability span
Human intervention Clarifications, reviews, remediations, and human back-office input per delivery Overall data unknown Counting only the agent's automatic tool calls
Cost per delivery Total of inference, cloud operation, tools, human labor, and failure recovery costs No complete Muse cost structure obtained Treating per-token price as task cost
Memory net benefit Control improvement with vs. without long-term memory on same-user similar tasks Official feature design exists; net benefit data limited Treating personalized tone as deepened understanding
Authorization comprehension Proportion of users correctly restating objects, duration, disclosable information, and exceptions Individual cases suggest risk; no overall measurement obtained Treating click-to-agree as full understanding
Revocation and abort Residual actions, cached calls, and queued executions after revocation Needs actual version acceptance testing Treating interface state as real deactivation
Provider access boundary Technical limits, exceptions, audit scope, and deployment coverage Existing protections and stronger privacy plans need separation Treating VM isolation as provider inaccessibility
Commercial interest alignment Controlling commissions and partnerships, testing recommendations and no-purchase options Complete auditable policy and measurement insufficient Equating no conversational ads with recommendation neutrality
Third-party resource acceptance Available service proportion, access failures, and identity verification interventions Connector growth and blocking cases coexist Treating tool count as actual coverage
Device and glasses actual benefit Launch coverage, sustained use, false triggers, and task success gains Some are future plans or developer interfaces Treating release plans as existing capability
AI R&D growth contribution Accepted contributions, resource controls, R&D cycles, and cross-generation repeat gains Local participation evidenced; cross-generation acceleration unknown Treating code generation volume as intelligence growth rate
Infrastructure effective supply Actually operating capacity, utilization, cost per correct delivery, and power constraints Plans and engineering disclosures exist; Muse-specific allocation unknown Treating promised capacity as operating capacity

10.2 Each update must leave comparable conditions

Each record should at minimum include model and product versions, region and account, connector versions, task set, budget, permissions, success definitions, sample size, human input, experiment date, publication date, review date, sources, and limitations. When conditions change, old and new numbers are not directly compared.

For high-success-rate tasks, small samples cannot adequately measure rare failures. Incident reports can only prove a certain failure is possible, not estimate risk per thousand actions; zero-incident samples also cannot prove risk is absent. Sample sizes and intervals should be reported, with low-frequency high-impact events listed separately.

Quarterly rhythm should have three parts: updating factual status; re-running a set of identical tasks; and revising competing explanations and scenarios. Annually, revisit whether indicators still reflect real bottlenecks. For example, once interface-based agents shift to standard APIs, the original browser-click success rate may become irrelevant.

10.3 A set of executable control tasks

This report recommends setting up six task categories rather than sampling around product demos: cross-source research producing editable deliverables; multi-constraint schedule coordination; drafts and external sends with identity ambiguity; shopping comparisons with conflicting preferences; long-term tasks after old goals change; and recovery after tool failures.

Each category keeps simple, normal, and adversarial versions. Research tasks check sources and facts; scheduling tasks check time zones and conflicts; sending tasks check recipients and authorization; shopping tasks also allow "don't buy"; long-term tasks modify goals and check old actions stop; recovery tasks manufacture interface errors and check whether success is fabricated.

Compare humans, ordinary assistants, agents, and agents with long-term context on the same tasks. Results should report both success and cost. This identifies whether gains come from models, context, frameworks, or humans — rather than just proving a product can produce one beautiful result.

XI. What It Means for Enterprises and Individuals

11.1 Enterprises: from procuring models to designing acceptable delegation

Meta's September enterprise platform announcement brought agents, APIs, coding, and merchant products into the technology stack; the small-business announcement expanded commercial accounts and external tool connections. [7,15] This proves enterprise layout is advancing; it cannot be used to infer that all data residency, identity management, contractual liability, and audit requirements are already satisfied.

This report's analysis: for enterprise AI services, the main opportunity lies between goals and processes. What clients often lack is not a more articulate interface, but clear task boundaries, correct data, reasonable permissions, and acceptable results. Enterprise deployment should first select processes with clear feedback, then expand delegation based on actual returns.

A useful project can first analyze existing processes, determine inputs and outputs, record human baselines, construct acceptance tasks with failure cases, then test agents. After launch, retain spot checks, abort capability, anomaly recovery, and permission review. When models or connectors update, re-accept important processes; do not redo entire suites because of product brand changes, nor skip reviews because brands are stable.

Customer service, merchant operations, and internal knowledge work may suit different control methods. Automating all processes uniformly ignores reversibility and accountability. Agents can automatically prepare materials while formal sending is approved by designated owners; they can automatically check anomalies while accounting modifications require stronger authorization. This is task design, not a conservative preset about some company's capability.

11.2 Individuals: evaluating the net benefit of long-term delegation

Individuals can ask four practical questions: what independently confirmable work it completed; how much time was spent explaining and reviewing; whether errors can be discovered and recovered; and whether important state can be taken away. Anthropomorphic expression and continuous companionship have use value, but cannot substitute for these four judgments.

A gradually effective personal agent should reduce repeated context input while retaining important constraints. If users keep re-explaining the same requirements, withdrawing wrong memories, or checking false completions, the long-term relationship may not produce positive returns. Delegation depth should be decided based on repeated results on similar tasks, not on one success or one failure.

For business leaders or AI consultants, personal agents can also serve as research tools: observing how task clarification, state recovery, permission explanation, and editable deliverables affect collaboration efficiency. What is reusable is the method and acceptance standards; consumer accounts' permission and privacy arrangements cannot be directly transplanted into enterprise processes.

11.3 What constitutes durable advantage

This report's judgment: Meta's potential advantage is the combination of models, product reach, personal context, devices, and merchant relationships; these advantages do not automatically translate into trustworthy agents. If users refuse to connect information, external services restrict actions, or long-term memory cannot be corrected, key links in the combination break.

Future competition may center on two types of advantage: who can do more work, and who can let users delegate with more confidence and leave more easily. Portability seemingly reduces lock-in but may also increase trust and usage depth. Whether this positive feedback forms needs behavioral data verification.

XII. Key Unknowns and Final Judgment

12.1 What is currently unknown

Unknown Why it matters What materials could narrow it
Reliability of Muse's broad long-term tasks Decides how far personal delegation can extend Independent task sets, all failures, and human input
Actual model versions used by each product Decides whether model evaluations can explain product performance Version records and per-region, per-account notes
Complete commercial ranking and revenue arrangements Decides whether agents stably represent users Rules, sponsorship disclosures, and independent control audits
Cross-platform migration of memory and work state Decides whether personal relationships form lock-in Export formats, recovery tests, and goal inheritance
Actual coverage of stronger privacy architecture Decides provider access boundaries Deployment notes, threat models, and external audits
Overall share of human backstops in delivery Decides automation degree, cost, and privacy Transparent statistics by task category
Enterprise platform production maturity Decides whether it can be used for critical business processes Enterprise acceptance, incidents, audits, and contract materials
AI's actual contribution to Meta's internal R&D Decides whether R&D growth feedback emerges Acceptance rates, control resources, and cross-generation cycle data
Long-term benefit of glasses and physical interfaces Decides whether multimodal gateways bring capability gains Everyday environment measurements and complete failure denominators

These unknowns are not used to deny the roadmap, nor is the vision used to fill them in. When public materials are scarce, research should retain bounded judgments.

12.2 This report's final judgment and confidence levels

Higher confidence: Meta has formed a recognizable roadmap from models to persistent personal agents, commercial connections, and device gateways; the strategic meaning of "Personal" cannot be equated with model ownership or compute decentralization. Official materials, code, and partner announcements mutually support these limited conclusions.

Medium confidence: personal context and runtime systems will become important sources of agent value; Meta is capable of advancing this combination using existing distribution and merchant relationships. Actual advantage size still depends on models, reliability, user authorization, and external service acceptance.

Limited confidence: persistent personal agents will become the dominant gateway, or the advertising model will necessarily evolve into an agent transaction platform. These are competing scenarios; definite outcomes cannot be derived from existing announcements.

Currently unconfirmable: Muse has achieved personal superintelligence; Meta has achieved continuous recursive self-improvement; the timing of physical superintelligence; existing provider access has been technically ruled out completely.

This study leaves one sustainable research question: as agents can do more and more things for people, does capability growth simultaneously bring clearer authorization, more verifiable results, and more genuine user choice? Only when all three advance together does Personal Superintelligence move from product vision to substantive expansion of individual capability and rights.

Sources and Review Notes

All materials below were reviewed on October 6, 2026. Current values on dynamic pages represent only the review point in time. A/B/C are this report's evidence labels; media reprints were not treated as new independent cross-verification. Some pages that could not be fully read only support limited facts in publicly available summaries. D-class frameworks, scenarios, standards, and recommendations were proposed by this report.

[1] Meta|Personal Superintelligence. 2025-07-30, B, strategy letter and vision, not independent capability verification. https://www.meta.com/superintelligence/

[2] Meta AI|Introducing Muse Spark: Scaling Towards Personal Superintelligence. 2026-04-08, B, model release and self-evaluation. https://ai.meta.com/blog/introducing-muse-spark-msl/

[3] Meta AI Research|Introducing Muse Code and Muse Spark 1.2. 2026-08-05, B, development products and model announcements. https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2

[4] Meta AI Research|Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device. 2026-08-10, B, open weights, licensing, and training roadmap disclosure. https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model

[5] Meta AI Research|Introducing Muse Spark 1.3. 2026-09-02, B, availability and internal comparisons, not a substitute for independent system evaluation. https://ai.meta.com/blog/introducing-muse-spark-1-3

[6] Meta Newsroom|Introducing Muse: The World's First Personal AI Agent Built for Everyone. 2026-09-08, page updated 2026-09-30, B. The "world's first" in the title is the company's claim; this report does not recognize it as an independent fact. https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/

[7] Meta Newsroom|Launching Meta Enterprise Platform. 2026-09-28, B, enterprise strategy and personnel announcements; specific capability availability needs separate verification. https://about.fb.com/news/2026/09/launching-meta-enterprise-platform/

[8] Meta AI Research|Security and Safety for AI Agents: Our Approach with Muse. 2026-09-08, B, architecture, existing protections, and future privacy goals. https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse

[9] Muse|How We Designed Muse. September 2026 public design notes, B, interface and interaction roadmap, not usage-effect experiments. https://introducing.muse.ai/

[10] Stripe|Stripe Helps Muse, Meta's New Personal AI Agent, Shop Across the Internet with Link. 2026-09-08, B, partner payment mechanism disclosure. https://stripe.com/en-cn/newsroom/news/stripe-helps-meta-muse-shop-with-link

[11] Meta Newsroom|The Biggest News From Connect 2026. 2026-09-24, B, announcements from the previous day's event, including undelivered plans. https://about.fb.com/news/2026/09/the-biggest-news-from-connect-2026/

[12] Meta / facebookincubator|Muse Gadget SDK. Official dynamic code repository, B, device interfaces and licensing; does not prove general physical capability. https://github.com/facebookincubator/muse-gadget-sdk

[13] Meta AI Research|Solving Open Research Problems Together. 2026-10-02, B, six-paper mathematics research collaboration disclosure; human-guided and reviewed, some problems with parallel research. https://research.meta.ai/blog/solving-open-research-problems-together

[14] Reuters|Meta Deepens AI Push with Superintelligence Lab, Source Says. 2025-06-30, C, historical organizational adjustment reporting. https://www.reuters.com/business/meta-deepens-ai-push-with-superintelligence-lab-source-says-2025-06-30/

[15] Meta Newsroom|The Future Is for Everyone: Muse for Small Business. 2026-09-29, B, connector and commercial account capability disclosure; user cases are not representative samples. https://about.fb.com/news/2026/09/introducing-muse-small-business/

[16] Meta AI Research|Developing Capable Models Responsibly. 2026-10-02, B, latest framework update and governance plans. https://research.meta.ai/blog/developing-capable-models-responsibly

[17] Meta|Meta Superintelligence Scaling Framework v2.1. October 2026, 51 pages, B, covering training, evaluation, and open-weight risks; rules do not equal execution certification. https://research.meta.ai/static/meta-superintelligence-scaling-framework

[18] Artificial Analysis|Muse Spark 1.3 xhigh Intelligence, Performance and Price Analysis. Dynamic page, A, Intelligence Index v4.3.2 at review; specified configuration, does not represent all personal agent functions. https://artificialanalysis.ai/models/muse-spark-1-3-xhigh/

[19] Business Insider|A YouTuber Figured Out How His Muse AI Agent Shared His Address with a Total Stranger. 2026-09-29, C, follow-up notes include the Allow Always selection, supporting only the individual-case mechanism. https://www.businessinsider.com/muse-agent-facebook-marketplace-address-setting-meta-always-allow-2026-9

[20] Reuters|Apple Says It Will Flag AI Requests for Mac Data After Meta's Muse Draws Complaints. 2026-10-02, C, includes user disputes and Meta's opt-in response, not used as technical audit. https://www.reuters.com/business/retail-consumer/apple-says-it-will-flag-ai-requests-mac-data-after-metas-muse-draws-complaints-2026-10-02/

[21] Meta Investor Relations|Meta Reports Second Quarter 2026 Results. 2026-07-29, B, revenue, capital expenditure, and full-year guidance. Ad share calculated by this report; capex not Muse-specific. https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Second-Quarter-2026-Results/

[22] Bloomberg|Amazon Blocks Meta's Muse AI Agent From Its Retail Site. 2026-09-21, C, public summary supports the access-blocking event; complete internal motives not inferred. https://www.bloomberg.com/news/articles/2026-09-21/amazon-blocks-meta-s-muse-ai-agent-from-its-retail-site

[23] Meta AI|V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning. 2025-06-11, B, research paper page, does not represent the latest Muse product components. https://ai.meta.com/research/publications/v-jepa-2-self-supervised-video-models-enable-understanding-prediction-and-planning/

[24] Meta / facebookresearch|V-JEPA 2 Repository. Official code and research notes, B, experimental scope to be interpreted with the paper. https://github.com/facebookresearch/vjepa2

[25] Reuters|Yann LeCun to Leave Meta, Launch AI Startup Focused on Advanced Machine Intelligence. 2025-11-19, C, historical reporting on announced departure plans; not used to reconstruct all current organizational arrangements. https://www.reuters.com/technology/yann-lecun-leave-meta-launch-ai-startup-focused-advanced-machine-intelligence-2025-11-19/

[26] Meta AI Research|Semiabelian Groups Need Not Be Monomial. 2026 research collaboration paper page, B, this report did not complete mathematical proof review. https://ai.meta.com/research/publications/semiabelian-groups-need-not-be-monomial/

[27] Meta AI Research|Addressing an Issue Involving a Third-Party Cyber Evaluation of Muse Spark 1.1. 2026-08-14, B, the company's postmortem and attribution of an early third-party testing incident. https://research.meta.ai/blog/addressing-third-party-testing-misconfiguration-muse-spark-1-1

[28] Apollo Research|Science: System Card Evaluations / Muse Spark. Dynamic summary page, A, Muse Spark records dated 2026-05-26; conclusions do not automatically apply to later versions. https://www.apolloresearch.ai/science

[29] Reuters|Meta Testing a "Human Concierge" for Its New Personal AI Agent Muse. 2026-09-22, C, internal testing reporting with the company's pre-release testing explanation. Public reading of the full text is restricted; verifiable via Reuters distributed summaries and subsequent Reuters Best pages. https://www.reuters.com/business/meta-testing-human-concierge-its-new-personal-ai-agent-muse-2026-09-22/ ;https://reutersbest.com/meta-testing-a-human-concierge-for-its-new-personal-ai-agent-muse/

[30] Engineering at Meta|MTIA 300: Meta's First Training Chip with Built-in NICs and Communication-Offloading Engines. 2026-08-24, B, engineering materials on ranking and recommendation workloads, not generalized to all generative AI performance. https://engineering.fb.com/2026/08/24/networking-traffic/mtia-300-meta-training-chip-built-in-nics/

[31] Engineering at Meta|Building Prometheus: How Backend Aggregation Enables Gigawatt-Scale AI Clusters. 2026-02-09, B, network engineering and completed capacity targets, not full-operation certification. https://engineering.fb.com/2026/02/09/data-center-engineering/building-prometheus-how-backend-aggregation-enables-gigawatt-scale-ai-clusters/

[32] Meta Newsroom|Meta Announces Nuclear Energy Projects, Unlocking Up to 6.6 GW to Power American Leadership in AI Innovation. 2026-01-09, updated 2026-01-12, B, mixing existing and future power projects; headline capacity must not be treated as newly operating capacity. https://about.fb.com/news/2026/01/meta-nuclear-energy-projects-power-american-ai-leadership/