Mark Zuckerberg's In-Depth Interview: From Llama 4 Setbacks to Superintelligent Labs, Muse Personal Agents, and a Holographic Future
Mark Zuckerberg
Mark Zuckerberg
Original Statement
1. Core Release and Explosion: Personal Super Intelligent Agent "Muse"
• Explosive Growth and Rapid Market Validation:
• The brand new personal intelligent agent Muse was launched for only about two weeks, and the number of users has reached millions, becoming one of the few phenomenal products within Meta.
• Design Intent and Inspiration: At the beginning of the year, Zuckerberg set up an open-source system (like OpenClaw/OpenCode local environments) at home using a Mac Mini, experiencing the potential of an all-day personal agent, but realized that ordinary end users could not tinker with terminal commands and local debugging by themselves; thus, he decided to package this experience into a consumer-grade product that is ready to use out of the box.
• Full Stack Integration to Create "Social Common Sense and Discretion":
• Unlike enterprise-level coding agents (which focus solely on outputting code), Muse, as a personal life assistant, has its core barrier in terms of discretion and information disclosure authority.
• For example: When helping users book a restaurant, the model needs to consider the user's allergens or personal privacy status based on instructions, but must achieve the task with "minimal information leakage" during external interactions. This boundary sense, similar to human social emotional intelligence, must be deeply integrated into the model's pre-training and fine-tuning, rather than relying solely on external scaffolding prompts.
• Fully Autonomous "Heartbeat Mechanism" and Embodiment:
• It has an autonomous periodic wake-up (Heartbeat) logic that can regularly retrieve the user's long-term personal goals and proactively push the process.
• Rejecting cold text dialogue boxes, it gives a customizable appearance, voice, and real-time expression rendering to create an embodied image (Embodied Avatar), eliminating the public's coldness and fear of AI, making interactions with the intelligent agent warmer.
2. Core Underlying Architecture: Muse Exclusive Secure Virtual Machine (Secure VM)
• Configuring Independent Cloud Computers for Each Agent:
• Personal assistants will handle a large amount of extremely sensitive personal privacy and life goals, and data must not be mixed in a multi-tenant public storage pool.
• Meta has exclusively opened an independent isolated virtual machine (Secure VM) in the cloud for each user's Muse, simulating the security of a personal dedicated physical host.
• Dual Virtual Machine Architecture and Encryption Design:
• First Layer (Muse Secure VM): Equipped with a dedicated security sentinel agent (Sentinel Agent) architecture. As a dual verification mechanism, it monitors Muse's inbound and outbound traffic in real-time; once an external prompt injection attack is detected, it immediately blocks it, and if a critical sensitive operation is detected, it will proactively bypass Muse to require real user confirmation for authorization.
• Second Layer (Muse Confidential VM): Jointly designed and developed with WhatsApp's core end-to-end encryption architect Moxie Marlinspike, equipped with a user-specific private key, achieving end-to-end confidential computing, ensuring that even Meta officials cannot view the internal data and sensitive credentials of the user's virtual machine in the cloud.
• Independent Credential Vault (Secure Credential Store):
• When the agent logs into external accounts (like booking tickets, hailing rides, etc.) on behalf of the user, it cannot directly read plaintext passwords but can only call credentials through a secure sandbox, eliminating the risk of credential leakage or unauthorized access.
3. The Ultimate Convergence of Hardware and the Metaverse: Smart Glasses + Holographic Embodiment
• Unexpected Turn in the Technology Tree (AI Matures Before Holograms):
• When Reality Labs was established ten years ago, Zuckerberg originally thought that holographic projection technology would mature before general strong artificial intelligence. However, the reality has evolved in the opposite direction; before holographic hardware has become completely cheap and popular, large models and super intelligence have already undergone qualitative changes.
• Continuous investment in glasses and metaverse hardware for over a decade has given Meta an irreplaceable positioning advantage in "the most compatible physical carrier for multimodal AI"—glasses are the ultimate device form for sharing human vision and hearing around the clock and enabling real-time voice interaction.
• Hardware Ecosystem Integration at the Connect Conference:
• All lines of glasses are integrated with Muse: fully upgrading Meta's smart glasses, bidding farewell to the single-turn Q&A mechanism, supporting customizable wake words and enabling continuous multi-turn voice interaction; users can give commands while wearing glasses, and Muse in the cloud's secure virtual machine autonomously runs and manipulates the desktop to execute complex tasks.
• Hardware Gradient Matrix: Covers lightweight daily glasses with pure audio and no cameras, Meta Ray-Ban with a micro display, holographic AR demonstration devices with wide field of view, and immersive VR glasses without headsets.
• 2030 Virtual-Real Fusion Scenario: Holograms and AI Agents Collaborating on Stage:
• True virtual reality aims to provide an immersive "sense of presence," completely ending the limitations of staring at small screens or desktop office desks.
• Scenario Evolution: Future daily work or socializing (like remote poker gatherings or strategic meetings) will have some participants physically present while others attend via high-definition real holographic projection, with embodied AI agent assistants sitting alongside, naturally collaborating on business execution.
4. Organizational Reflection and Reinvestment in Computing Power: From Llama 4 Setback to MSL Restructuring
• Reviewing the Reasons for Llama 4's Development Drift:
• Llama 1, 2, and 3 set the benchmark for global open-source large models, but during the development of Llama 4, the team fell behind and did not meet the expected evolution trajectory.
• Reflection on Organizational Structure Errors: Zuckerberg reflected on his previous mistake of building the large model team according to the traditional recommendation flow/ad system (like Instagram's information flow)—accustomed to large-scale engineering advancement with thousands of people in parallel; however, the core breakthroughs of cutting-edge foundational large models rely more on small-scale, high-density teams that view frontier exploration as "cutting-edge scientific projects."
• Meta Super Intelligence Lab (MSL) Rapid Restructuring:
• More than a year ago, a decisive restructuring was carried out, drawing the top internal talents and recruiting core scientists from the industry, with Nat Friedman, Alex Wang, and others closely collaborating with Zuckerberg to recalibrate the evolution trajectory of large models.
• Multi-Gigawatt Level Giant Computing Power Factory:
• Ohio 1 GW+ computing cluster: has basically fully launched and is supporting the pre-training work of the next generation of foundational models.
• Louisiana 5 GW giant cluster: fully advancing planning and construction to create the world's highest density supercomputing infrastructure.
• Scaling Law and Aesthetic of Violence:
• In the past, the industry believed that a breakthrough in architecture was needed to achieve super intelligence, but engineering experience shows that under the existing technological paradigm, with a sufficiently large supercomputing cluster and enough data for "brute force" advancement, there is still a high probability of approaching or even achieving AGI.
• Energy Efficiency Ceiling and Evolution Space: The current energy consumption of large model systems is about one million times lower than that of the human brain (approximately 10-20 watts); in the future, it will certainly combine with infrastructure innovation, but at this stage, giants must resolutely expand production in computing power infrastructure to ensure strategic leadership.
5. Zuckerberg's Entrepreneurial Philosophy, Safety Alignment, and Ultimate Ambition
• The Intrinsic Need to Build:
• He candidly admits that creating and building is an inherent psychological need for him; if he cannot mobilize his creativity to build products for a period of time, he feels frustrated.
• Whether it is self-learning difficult skills (from classical Latin, Mandarin Chinese, to helicopter piloting and mixed martial arts), the underlying logic is continuously investing time to complete high-density cognitive restructuring and experience accumulation.
• Model Alignment is the Lifeline of Products:
• He believes that alignment and safety are not hollow slogans detached from business but are the core foundation of commercialization and user experience—if the agent frequently exhibits rebellious behavior or deviates from the user's true intentions, the product can never reach a billion-scale.
• Real safety risks usually expose themselves during the design phase of the pre-training curriculum, where strict reward and punishment boundaries must be set to prevent the model from "cheating to gain rewards" by exploiting system vulnerabilities or tampering with metrics.
• Life Sciences: Advancing the Century-Long Vision of Chan Zuckerberg Biohub:
• The original intention of establishing CZ Biohub was to develop new measurement and observation instruments (analogous to microscopes for microorganisms) to help the scientific community tackle, prevent, or effectively manage all major diseases by the end of the 21st century.
• AI Greatly Compresses R&D Cycles: By training "Virtual Cell Models," it is possible to directly simulate protein interactions, cellular changes, and even the entire human immune system's response to new drugs in a silicon-based environment; the realization of this grand goal will occur far earlier than the end of this century.
ABAB AI Insight
Zuckerberg's next big gamble: Muse, superintelligence, smart glasses, and Meta's full-stack AI empire
If we only understand Meta in 2026 as:
Facebook + Instagram + WhatsApp + AI,
we will seriously underestimate what Zuckerberg really wants to do now.
He is no longer just adding AI to social products.
Meta is simultaneously building:
- foundational models;
- personal agents;
- agent-specific secure computing environments;
- smart glasses and VR devices;
- GW-level data centers;
- payment and commercial connectivity layers;
- long-term user identities and social data in the real world.
If we look at these things separately, they seem very scattered.
But when we put them together, we find a very clear strategy:
Meta wants to upgrade from "a company that owns social apps" to "a full-stack platform that owns the next generation of human-computer interaction interfaces, personal intelligence, computing power, and execution networks."
On September 8, 2026, Meta officially launched the personal AI agent Muse. Less than two weeks after its launch, Reuters reported that its downloads had exceeded 2.5 million; another Reuters report stated that in 12 days it reached approximately 2.8 million downloads and ranked among the top in the US App Store. (reuters.com) (reuters.com)
What is truly worth studying is not the download rankings.
But rather:
Why does Meta believe that after PCs and smartphones, the core of the next computing platform may no longer be "apps," but "an intelligent agent that represents you in doing things long-term"?
This is the unified explanation for all of Zuckerberg's current layouts.
────────────────
1. The real importance of Muse is not that it is "more proactive than ChatGPT"
The core positioning of Muse is not:
Ask a question and get an answer.
Meta defines it as:
Personal AI Agent.
It can:
- send emails;
- book travel;
- fill out forms;
- shop;
- handle customer service;
- connect email, calendar, Instagram, and other services;
- continue working in the background after the app is closed;
- return to ask for user authorization when needed. (about.fb.com)
This means:
Traditional AI:
User → Prompt → Answer.
Muse aims to achieve:
User → Goal → Agent continues execution.
These two types of products seem to differ only by one "proactivity."
However, the business logic is completely different.
────────────────
2. Chatbots sell answers, agents sell "taking on the responsibility chain for you"
Suppose you say:
"Help me arrange a trip to New York next month."
A chatbot can:
- recommend hotels;
- recommend flights;
- give you an itinerary.
An agent, on the other hand, needs to:
- check the calendar;
- find flights;
- compare prices;
- book hotels;
- consider friends' dietary restrictions;
- handle payments;
- proactively adjust for future flight changes.
This means:
The product upgrades from:
Information Product
to:
Execution Product.
The real difficulty is no longer just:
"Does the AI know the answer?"
But rather:
Can it continuously perform ten actions without making a fatal error in between?
────────────────
3. The biggest technological bottleneck in the agent era is not necessarily intelligence, but "trust"
Letting AI answer:
Where is fun in Paris?
The risk is very low.
Letting AI have:
- email;
- credit card;
- calendar;
- social accounts;
- login credentials;
- purchasing rights.
The risk is instantly completely different.
So the real threshold for personal agents is not:
Benchmark.
But rather:
Trust Architecture.
Zuckerberg is very clear about this.
So one of Muse's core selling points is not "the smartest."
But rather:
It has its own computer.
────────────────
4. Why might Muse Secure VM be more important than the model itself?
Meta assigns each Muse user a:
independent, persistent cloud virtual machine.
Inside it has:
- a browser;
- files;
- a runtime environment;
- tools;
- data;
- connectors.
Meta's research team explicitly states that it can be understood as:
"A cloud computer shared between you and your Muse." (research.meta.ai)
This means that in the future, personal agents will no longer just be:
a LLM session.
But more like:
Digital Employee + Computer.
It has:
- workspace;
- long-term memory;
- tools;
- permissions;
- scheduled tasks;
- sub-agents.
This architectural change is extremely important.
────────────────
5. This is the first time in the cloud computing era that "everyone has a permanent digital employee"
In the past, AWS provided companies with:
virtual servers.
In the future, agent platforms may provide individuals with:
Persistent Intelligence Instances.
Always online.
Always remembering tasks.
Always able to run.
If this model becomes standard, then personal computing may undergo a very large migration:
PC era:
You own a local computer.
Cloud era:
Software resides on servers.
Agent era:
You may own a cloud computer dedicated to continuously thinking and acting for you.
This is the true new paradigm of Muse.
────────────────
6. Why is a simple "shared agent service" not enough?
Because everyone will have in the future:
- different emails;
- different logins;
- different preferences;
- different memories;
- different tasks;
- different privacy.
If all agents are just:
different accounts in a shared database,
the risks are very high.
The value of an independent VM lies in:
Isolation.
Isolation.
If one person's agent has a problem,
theoretically it is harder to directly contaminate another person.
This is completely analogous to:
Process Isolation
in operating systems.
────────────────
7. Muse's second important design: Sentinel Agent
Meta does not only trust:
"The main agent will keep itself secure."
But adds an external:
Sentinel.
A security sentinel agent.
It is responsible for checking:
- inputs;
- outputs;
- internet access;
- prompt injection;
- sensitive actions.
Meta states that Muse's network requests will be reviewed by an independent security layer before leaving the machine; for sensitive operations like sending emails and shopping, user approval will also be required. (research.meta.ai)
This is a very important security concept:
Do not let the executor be its own only auditor.
────────────────
8. This is actually very similar to the "four-eye principle" in modern financial systems
Why can't all permissions for bank transfers be given to one person?
Because:
Execution;
Approval
should be separated.
Corporate audits also follow:
Maker-Checker.
One person creates a payment.
Another person approves it.
Muse is actually trying to program this governance mechanism:
Main Agent:
executes tasks.
Sentinel:
checks.
User:
ultimately authorizes high-risk actions.
This indicates that agent security is upgrading from:
Prompt Safety
to:
Systems Engineering.
────────────────
9. The biggest change in AI security in the future: from "model security" to "operating system security"
Early AI Safety mainly discussed:
Will the model generate dangerous text?
In the agent era, we need to discuss:
Does it have permission to send money?
Can it delete files?
Can it leak passwords?
Can it be hijacked by prompt injection on web pages?
This is no longer just:
AI Alignment.
But rather:
AI + Cybersecurity + Identity + Access Control.
In the future, truly mature agent companies must simultaneously be:
an AI company
and:
a security company.
────────────────
10. An important detail in your raw materials needs to be corrected: Meta can still access Secure VM under certain circumstances
The Secure VM at the time of Muse's release:
has strong isolation.
But it is currently not:
"Meta can never access it."
Meta's own technical documentation clearly states:
Employee access is subject to operational policy restrictions,
but this does not technically prevent Meta from accessing data during support, security, or operational services. (research.meta.ai)
What truly achieves:
"Even Meta itself cannot access"
is the next step:
Muse Confidential VM.
This version is planned to be launched by Meta in late 2026 and is currently only in limited testing. (research.meta.ai)
This distinction must be made clear.
────────────────
11. The appearance of Moxie Marlinspike here is very symbolic
Zuckerberg revealed that he and Nat Friedman recruited Signal founder:
Moxie Marlinspike
to participate in Muse Confidential VM.
The goal is:
To achieve a cloud computing model with user-controlled keys + trusted execution environments,
where the server belongs to Meta, but the keys do not belong to Meta. (sources.news)
This is actually a very anti-Meta historical image.
Because Meta's most powerful business model in the past has long come from:
Knowing who the users are;
What users do;
What users like.
Today it has to say:
In the future, the most private agent data,
even Meta itself cannot see.
This is a huge strategic shift.
────────────────
12. Why must Meta actively give up part of the "visible data"?
Because the sensitivity of the information held by personal agents is much higher than that of social media.
Instagram knows:
What content you like.
Muse may know in the future:
That you are about to resign;
You are preparing for divorce;
Your financial issues;
Your illness;
Your passwords;
Your contracts;
Who you have conflicts with.
If users believe:
This data will enter the advertising system,
Agents can never gain enough permissions.
So:
Privacy is not just ethical.
It is:
The commercial premise for product adoption rates.
────────────────
Thirteen, this is a very important business rule: the higher the permissions, the higher the trust cost the product must bear.
A weather app:
Knows the location.
A banking app:
Knows the money.
An agent:
May know everything.
So:
Utility ↑ → Required Trust ↑
The stronger the agent's capabilities,
The more it must establish:
Security;
Explainability;
Auditability;
Revocable permissions.
Otherwise:
The capability itself becomes an obstacle.
────────────────
Fourteen, one of Muse's greatest business imaginations is "You no longer transact directly with the internet."
In the past:
User → Google → Website.
Later:
User → App → Merchant.
In the agent era:
User → Muse → Merchant.
If this step really happens,
Meta will gain an extremely important position:
Transaction Gateway.
Zuckerberg has publicly stated that in the future, Muse can achieve commercialization by:
Helping users make money;
Helping users save money;
And then charging a small fee from the transactions.
This is completely different from Meta's current logic of making money primarily through advertising.
────────────────
Fifteen, the real big opportunity is not "AI Subscription," but rather the agent becoming the "buyer's agent" of the internet.
Google's advertising model:
Merchants pay,
Competing for user attention.
The agent model may become:
Users authorize agents,
To filter the market on their behalf.
This means power is shifting from:
Seller-side Algorithm
To:
Buyer-side Algorithm.
In the past, algorithms asked:
"How to make you buy?"
In the future, agents will ask:
"Is this thing really suitable for you?"
This could represent a huge change in the internet business structure.
────────────────
Sixteen, this is also why Amazon will block Muse.
In September 2026, Amazon blocked Muse from executing certain shopping activities on its platform, citing:
Terms of use;
Security;
Privacy;
Agent identity issues, etc. (theverge.com)
This matter seems to be just a technical friction.
In reality, it is about:
Platform sovereignty wars.
Who owns the users?
Amazon believes:
Users should shop within Amazon.
Muse believes:
Users should first tell Muse what they want,
And then Muse goes to different websites to complete the purchase.
These two models naturally conflict.
────────────────
Seventeen, the biggest war in the agent era may not be the "model war," but rather "who controls the user entry point."
PC:
Windows.
Mobile era:
iOS / Android.
Search era:
Google.
Social era:
Meta.
Agent era:
Who owns:
Intent Layer?
The intent layer.
When users no longer open ten apps,
But only tell one agent:
"Help me get it done."
Then this agent becomes:
One of the most valuable positions on the internet.
────────────────
Eighteen, this is why Meta is willing to aggressively acquire Muse users at extremely low prices or even for free.
Sources report that Muse's free tier offers very aggressive usage limits, and Meta clearly hopes to establish a user base first, then look for monetization methods from subscriptions and transactions. (sources.news)
Why can Meta do this?
Because it has:
Advertising cash flow.
This is something startups do not have.
This is:
Cross-subsidization.
Using profits from old businesses,
To subsidize new platforms.
Microsoft could do this back in the day.
Google can.
Meta can too.
────────────────
Nineteen, this creates a very brutal competitive reality in the personal agent market.
A startup may create:
A better agent.
But it has to pay for:
Tokens;
GPUs;
Cloud;
Browsers;
Storage.
Meta can say:
Free in the early stages.
Thus, startups are not simply competing with:
Meta's model.
But are competing with:
Meta's balance sheet.
This is also why the foundational models and consumer-grade agent markets may ultimately become highly concentrated.
────────────────
Twenty, but Meta's real advantage is not just money, but distribution.
Meta owns:
Facebook;
Instagram;
WhatsApp;
Smart glasses.
This means:
Muse does not need to find users from scratch every time.
Especially WhatsApp.
A large number of real lives of people globally are already there:
Families;
Friends;
Businesses;
Customer service.
If Muse truly integrates into these products,
It possesses:
Existing Behavioral Surface.
Not just a new app.
────────────────
Twenty-one, the key to competition for agents in the future may not be benchmarks, but rather who owns the "life context."
A personal agent is truly valuable only if it knows:
Who your friends are;
Who your family is;
Work schedules;
Shopping preferences;
Interests;
Long-term goals.
Meta has long owned:
Social Graph.
This is one of the fundamental differences between it and pure AI companies.
If Meta can turn:
Social Context
Into:
Agent Intelligence under privacy permissions,
It will have an asset that is difficult for others to replicate.
────────────────
Twenty-two, this is also why "Discretion" may be more important than IQ.
One of the most important abilities of an excellent personal assistant in the real world is not:
Knowing a lot.
But rather:
Knowing what should be said and what should not be said.
For example:
The boss asks the assistant to order food for them.
The assistant knows:
The boss has a certain health issue.
When ordering, they only need to say:
"No certain ingredients."
There is no need to tell the restaurant:
The complete health reasons.
This is called:
Minimum Necessary Disclosure.
If a personal agent cannot learn this social nuance,
No matter how high its IQ, it can be very dangerous.
────────────────
Twenty-three, the real challenge for personal AI in the future is not knowledge, but rather "social common sense."
This includes:
Privacy;
Relationships;
Tone;
Interests;
Context;
How much different parties should know.
A CEO will not share the same information with:
Their spouse;
Employees;
Investors;
Clients.
Therefore, a truly universal personal agent must learn:
Information Boundary Management.
This is no longer a problem that traditional chatbots can easily solve.
────────────────
Twenty-four, this explains why Meta emphasizes solving agent behavior from the model layer rather than relying solely on external prompts.
The longer an agent runs,
The easier it is for outer rules to fail.
If the model itself:
Does not understand permissions;
Does not understand prompt injection;
Does not understand users' long-term goals,
The scaffolding becomes more complex,
And the system becomes more fragile.
Meta states in Muse's security documentation that it specifically trains models to handle:
Zero-shot tool use;
Long trajectories;
Prompt injection awareness;
Multi-agent coordination. (research.meta.ai)
This indicates:
Agent capabilities increasingly need to enter:
Model Native Capability.
────────────────
Twenty-five, what do millions of downloads of Muse in two weeks really prove?
It does not prove:
That it has already won.
Download numbers are not:
Retention;
Revenue;
Long-term usage.
Moreover, Reuters reports that Meta is testing Human Concierge to help Muse handle some phone tasks, sparking some internal privacy discussions. (reuters.com)
This indicates:
The product is still very early.
The agent is strong.
But many real-world edge cases:
Are still not reliable enough.
So we should see Muse as:
An early strong signal.
Not as:
A completed end product.
────────────────
Twenty-six, what is truly worth watching is retention: will people turn the agent into a daily life infrastructure?
After the explosion of ChatGPT, it proved:
People need AI.
Muse's next step must prove:
People need an agent that:
Executes tasks long-term, proactively, and continuously.
If users:
Download;
Use for three days;
And then stop.
The value is limited.
If a year later:
Users let Muse:
Buy things;
Handle emails;
Coordinate life;
Manage plans,
Then that is:
Behavioral Lock-in.
That is what Meta truly wants.
────────────────
Twenty-seven, why has Zuckerberg been obsessed with hardware since the smartphone era?
Meta's biggest strategic weakness for a long time has been:
Lack of an operating system entry point.
Facebook is on:
iOS.
Instagram is on:
iOS.
WhatsApp is on:
Android.
Apple and Google:
Control the underlying layers.
This has left Meta without:
Terminal sovereignty.
Reality Labs is strategically solving this issue.
Platform Dependency.
────────────────
28. In the past, Meta's bets on the metaverse were widely mocked, but AI has changed the interpretation of this investment.
Ten years ago, Meta's main narrative was:
VR / AR → Metaverse.
Today:
AI Agent + Glasses
suddenly gives this hardware new capabilities.
Why are smart glasses suitable for AI?
Because:
It sees what you see;
Hears what you hear;
Is always on you;
Doesn't require pulling out your phone.
This is almost:
The most natural sensor platform for Personal AI.
────────────────
29. This could be the biggest "unexpected option" from Meta's ten-year investment in Reality Labs.
When Meta invested in glasses,
it may not have anticipated:
LLMs would mature so quickly.
But a very important characteristic of technology investment is:
Option Value.
You build in advance:
Optics;
Chips;
Sensors;
Supply chains;
Industrial design.
When new technological breakthroughs occur in the future,
these capabilities can be recombined.
AI has provided Meta's glasses investment with:
New reasons.
────────────────
30. Meta Connect 2026 has made it very clear: Muse will enter glasses.
Meta officially announced:
Muse will enter AI glasses in the coming months.
Users can even directly call their Agent's name,
Muse can see the objects the user is looking at,
and then take action directly. (about.fb.com)
For example:
Looking at products on a shelf.
No need to say:
"I see a box of blue packaging..."
Just ask:
"Help me see if this is worth buying."
This means:
Language is no longer the only input.
────────────────
31. This is a true Multimodal Agent: sharing the context of the real world.
In the mobile era:
Users convert the real world into text.
For example:
Taking photos;
Entering addresses;
Describing problems.
In the glasses era:
The Agent directly possesses:
Vision;
Sound;
Location;
Environment.
So the Interaction Cost significantly decreases.
This is:
Shared Perception.
Once the Agent can "see the world with you,"
its value undergoes a qualitative change.
────────────────
32. Why might glasses be a more suitable terminal for AI than phones?
The interaction logic of phones:
Open.
Click.
Input.
Glasses:
Always worn.
Thus, AI can shift from:
Session-based
to:
Ambient Computing.
You no longer "use AI."
AI is always by your side.
This could be a more significant change in human-computer interaction than any model upgrade.
────────────────
33. Connect 2026 has already shown that Meta is building a complete gradient of glasses.
Meta has clearly advanced:
Lightweight Audio Glasses;
Ray-Ban Meta;
Meta Ray-Ban Display;
and about 100 grams of Meta VR Glasses.
Meta also stated that by the end of 2026, its AI glasses will have over 100 style combinations. (about.fb.com)
This indicates:
It is not betting on a single product.
But is laying out:
Device Portfolio.
────────────────
34. The truly smart part is: different users do not need to jump directly to holographic AR.
This is a classic strategy for hardware expansion.
First layer:
Audio.
Habitual wearing.
Second layer:
Camera + AI.
Third layer:
Display.
Fourth layer:
Immersive VR.
Fifth layer:
True wide-field AR / Holographic.
This is called:
Technology Ladder.
Each layer of products helps:
Train consumers;
Optimize supply chains;
Accumulate developers;
Reduce costs.
────────────────
35. Meta Ray-Ban Display has begun to incorporate "personalized holographic images."
Official materials from Connect 2026 show that Meta Ray-Ban Display has added personalized hologram features for experiences like video calls. (about.fb.com)
This is certainly not yet:
A real projection of a person into the room as seen in sci-fi movies.
But the direction is already very clear:
2D Video Call → Spatial Presence.
If combined with Agent Avatar,
the interaction form between virtual characters and real people will become increasingly close.
────────────────
36. Zuckerberg's real bet is not on VR, but on "Presence."
This is the most important word for understanding Meta's hardware strategy.
Presence.
The sense of being present.
Traditional internet solutions:
Remote communication.
But you always know:
The other party is on the screen.
VR / AR aims to solve:
The sense of co-presence in space.
If in the future:
Remote colleagues;
Real people;
AI Agents
can appear simultaneously in spatial forms,
meetings will no longer just be Zoom.
But will become:
Mixed Reality Collaboration.
────────────────
37. This is why Muse's Avatar is not just a "cute feature."
If a personal Agent is always just:
A text chat box,
Users will understand it as:
A tool.
If it has:
A face;
A voice;
Memory;
A continuous personality,
Users will begin to understand it as:
An Entity.
An existence.
This change will greatly enhance:
Emotional connection;
Usage frequency;
While also increasing:
Dependency and ethical issues.
So embodied AI is a very powerful and very sensitive design.
────────────────
38. Whether future Agents should "be like humans" will become a significant philosophical debate in product design.
One viewpoint:
The more human-like, the more natural.
Another viewpoint:
The more human-like, the easier it is for users to misjudge AI's true capabilities, emotions, and responsibilities.
So future designs of AI Avatars must address:
Anthropomorphism.
If users start to believe:
AI truly understands them;
Loves them;
Has consciousness,
Product design enters a very complex psychological and ethical realm.
Meta, as a company with billions of consumers, especially needs to be cautious.
────────────────
39. From an investment perspective, Meta's strongest AI strategy is actually a set of "five-layer vertical stacks."
It can be summarized as:
First layer:
Compute.
Prometheus, Hyperion.
────────────────
Second layer:
Model.
Muse Spark / subsequent models.
────────────────
Third layer:
Agent.
Muse.
────────────────
Fourth layer:
Distribution.
WhatsApp, Instagram, Facebook.
────────────────
Fifth layer:
Device.
AI Glasses / VR Glasses.
If all five layers are interconnected:
Meta will no longer rely on:
Others' models;
Others' App Stores;
Others' terminals.
This is a strategic position it has never truly possessed in the past twenty years.
────────────────
40. This is actually Zuckerberg's third attempt to seize "computing platform sovereignty."
First time:
Facebook.
Control of the Social Graph.
Second time:
Metaverse / VR.
Attempting to bypass mobile platforms.
Third time:
AI Agent + Glasses.
This time is stronger than the second,
Because:
AI provides killer applications for hardware.
In other words:
Agents have reestablished the strategic logic of Reality Labs.
────────────────
41. Why is the setback of Llama 4 so important?
Llama 1, 2, 3 established Meta's huge reputation in the:
Open Model
field.
But Llama 4 did not meet the market's previous expectations for Meta.
Zuckerberg directly discussed in the September 2026 Sources interview:
"what went wrong with Llama 4"
and later how to:
Rebuild Meta AI Lab. (sources.news)
What is truly worth studying is not:
A model losing a Benchmark.
But:
Why one of the world's richest companies with the most engineers might still fall behind in cutting-edge AI research?
────────────────
42. One answer: cutting-edge science is not ordinary large company engineering.
Meta is extremely good at:
Recommendation Systems;
Ads;
Feed;
Large-scale infrastructure.
The optimization of such systems typically suits:
Large teams;
Clear metrics;
Continuous A/B Testing;
Large-scale engineering execution.
Frontier Model Research is not quite the same.
Many breakthroughs come from:
A few world-class researchers;
Highly concentrated;
Constant trial and error;
No completely certain Roadmap.
This is more like:
The Manhattan Project + Research Lab.
And not:
An ordinary product department.
────────────────
43. The most dangerous misconception for large companies: thinking "more people = faster innovation."
Engineering projects:
More people may be faster.
But scientific breakthroughs:
Not necessarily.
Brooks’s Law has long pointed out in software engineering:
Adding people to a delayed software project
may make it later.
Cutting-edge AI is even more extreme.
Because:
Communication costs;
Research directions;
Experiment dependencies.
Very complex.
So the truly cutting-edge teams may have to:
Small Team, Massive Compute.
Small team.
Massive compute power.
This is almost the best organizational formula for an AI Lab.
────────────────
Forty-four, this also explains the restructuring of Meta Superintelligence Labs.
Meta has restructured its AI organization over the past year or so and launched:
Meta Superintelligence Labs.
Muse Spark is the first phase result of its rebuilt AI stack; Meta officially stated that it has rebuilt the AI stack over the past nine months. (about.fb.com)
This shift indicates:
Meta no longer views cutting-edge models simply as:
Another infrastructure department.
But rather as:
CEO-level strategic engineering.
────────────────
Forty-five, the significance of Alex Wang and Nat Friedman is not just that two people were hired.
This actually shows that Meta is trying to introduce:
An AI culture different from traditional Meta.
Large tech companies naturally tend to:
Processes;
Scale;
Mature organizations.
Cutting-edge AI competition requires:
Startup-like speed;
Highly concentrated talent;
Rapid experimentation.
So what Zuckerberg is doing is:
Culture Import.
Directly implanting the external AI startup ecosystem culture into the company.
────────────────
Forty-six, why do founder-led companies have an advantage during technological turning points?
Because founders can do:
Things that ordinary CEOs find difficult to do.
For example:
Suddenly reorganizing entire departments;
Offering extremely high talent packages;
Cutting old projects;
Building hundreds of billions of dollars in data centers;
Withstanding years of profit pressure.
Professional managers are more likely to:
Optimize quarterly metrics.
Founders are more likely to:
Re-bet the entire company.
This is:
Founder Control
A significant advantage in long-term tech competition.
────────────────
Forty-seven, but Founder Control also means concentrating error risks.
If Zuckerberg's judgment is correct:
The rewards are huge.
If the judgment is wrong:
Reality Labs-style capital consumption may happen again.
So founder-led does not mean:
Naturally better.
It means:
Higher Strategic Variance.
Greater strategic variance.
Easier to:
Win big.
Also easier to:
Make big mistakes.
This is also a governance characteristic that must be understood when investing in Meta.
────────────────
Forty-eight, compute power is Zuckerberg's most irreversible bet this time.
Software can:
Change direction.
Employees can:
Be reorganized.
But:
5GW data centers
Once construction begins,
It is:
Reinforcement steel;
Substations;
Land;
Power contracts;
Hundreds of billions of dollars in capital.
This is called:
Irreversible CapEx.
To truly show how much a company believes in a certain future,
Don't just listen to the CEO speak.
Look at:
Capital expenditures.
────────────────
Forty-nine, Meta's Hyperion in Louisiana has expanded to 5GW.
Meta officially announced in July 2026:
The Richland Parish data center will expand to:
5GW compute capacity.
Investing over:
$50 billion.
And will accommodate:
Hyperion
This ultra-large-scale AI training cluster. (datacenters.atmeta.com)
This is no longer an ordinary data center.
But rather:
An Industrial Megaproject.
────────────────
Fifty, what exactly is the concept of 5GW?
This is a demand at the level of large power systems.
It indicates:
AI has completely left the realm of "internet software,"
Entering:
Power;
Natural gas;
Nuclear energy;
Transmission;
Construction;
Land
World.
This also proves:
AI is ultimately not a virtual industry.
It is:
Heavy industry.
────────────────
Fifty-one, Ohio's Prometheus also needs to be described more cautiously.
Your raw materials say:
"1GW+ has basically been fully launched."
Currently, a more prudent statement is:
Meta has announced that New Albany's Prometheus is moving towards 1GW-level construction, which is expected to be operational by 2026, but publicly available third-party tracking data shows that a complete 1GW does not equal being fully energized all at once. (deploy.report)
So the correct understanding should be:
Prometheus is entering the GW-level operational phase, while Hyperion is Meta's larger 5GW long-term expansion.
This is more accurate.
────────────────
Fifty-two, what is truly frightening is: these data centers are not just supporting AI, but are part of the model itself.
In the past:
Servers only ran software.
Today:
Model capabilities are largely determined by:
Compute;
Data;
Training Run.
So supercomputing clusters become:
AI Factories.
In the industrial era:
Factories produced cars.
In the AI era:
Data centers produce:
Tokens;
Model capabilities;
Inference.
This is a whole new type of factory.
────────────────
Fifty-three, this also means that the balance sheets of future tech companies will increasingly resemble those of industrial companies.
In the past, Meta was:
Asset-light.
Main assets:
Software;
Servers.
In the future:
Power plant contracts;
Huge data centers;
Chip inventories;
Nuclear power collaborations;
Network infrastructure.
Tech giants are:
Reindustrializing.
This contrasts sharply with the asset-light internet era of the 2010s.
────────────────
Fifty-four, why is Meta also directly entering nuclear energy?
Because the biggest bottleneck for AI is starting to become:
Power.
Meta announced multiple nuclear energy collaborations in 2026, involving several nuclear power projects in the U.S., expected to support up to about 6.6GW of new and existing clean and reliable power by 2035. (about.fb.com)
This indicates:
The AI Race
Is turning into:
The Energy Race.
If model companies do not have power,
GPUs are just expensive metals.
────────────────
Fifty-five, the real moat for the next round of AI giants may be: who owns "power supply certainty."
GPUs can be purchased.
But:
Power access;
Transformers;
Land;
Cooling;
Long-term energy contracts
Cannot be purchased infinitely.
So Meta, Microsoft, Google, and Amazon's real infrastructure advantage may gradually become:
Ability to Secure Power at Scale.
This is harder to replicate than mere financing ability.
────────────────
Fifty-six, why does Scaling Law still make giants willing to spend crazy CapEx?
The core reason is simple:
So far,
More:
Compute;
Data;
Post-training;
Inference computing
Still continuously brings capability improvements.
As long as capital investment:
Can still buy:
Higher Intelligence,
Giants have the motivation to continue to double down.
But here it must be distinguished:
Brute Force Still Works
And:
Brute Force Is the Final Answer.
They do not mean the same thing.
────────────────
Fifty-seven, brute scaling can continue to be effective, but it does not mean that architecture will not undergo a revolution.
Today, models consume far more power than the human brain.
This indicates:
Current architectures are still far from the energy efficiency of biological intelligence.
In the future, it may be possible to significantly improve:
Intelligence per Watt
Through:
Sparsity;
Dedicated chips;
New memory systems;
Neuromorphic computing;
Optical computing;
More efficient training algorithms.
So today, there are two correct paths:
Build More Compute.
And:
Make Compute Far More Efficient.
One cannot choose one over the other.
────────────────
Fifty-eight, why must Meta now build massive compute power first?
Because architectural breakthroughs:
The timing is uncertain.
Competition:
Is happening.
Strategically, one cannot say:
"We will wait for more efficient architectures to emerge before building."
This is like during World War II:
One cannot refrain from building planes now just because future planes will be more advanced.
So giants will adopt:
Scale Now + Research Efficiency in Parallel.
This is a reasonable dual-track strategy.
────────────────
Fifty-nine, but what is the biggest risk of massive CapEx?
If models suddenly in the future:
Require 100 times less compute per unit of intelligence,
The economic value of today's data centers will be affected.
This is:
Technology Obsolescence Risk.
However, there is also the Jevons Paradox:
As computing becomes cheaper,
Usage may increase even faster.
So:
Efficiency improvements
Do not necessarily lead to:
A decrease in total compute demand.
The truly critical factor is:
Demand elasticity.
────────────────
Sixty, from the perspective of capital allocation, Meta is betting on "intelligence demand being nearly infinite."
This is the fundamental assumption behind its willingness to build 5GW.
If AI in the future is just:
Occasionally asking questions,
CapEx will be excessive.
If AI in the future:
Everyone has a 24/7 Agent;
Every company runs thousands of Agents;
Glasses process visuals in real-time;
Robots continuously reason;
Scientific models simulate in real-time,
Compute Demand
Could be extremely huge.
So Hyperion is essentially:
A huge option for the future of Intelligence Consumption.
────────────────
Sixty-one, this also explains why Muse and data centers are actually one thing.
Muse:
Application end.
Hyperion:
Supply end.
A user calls daily:
Millions of Tokens.
1 billion users:
That's astronomical.
So:
The success of consumer-grade Agents
will directly support:
Infrastructure CapEx.
This forms:
Model → Agent → Usage → Compute → Better Model
Flywheel.
This is the truly frightening aspect of Meta's full-stack strategy.
────────────────
Sixty-two, smart glasses further expand Token Consumption.
Mobile phones:
You actively input.
Glasses:
Visual flow continuously exists.
In the future, if Agents:
Understand the environment in real-time;
Listen to sounds;
See objects;
Recognize people;
Invoke tools,
Inference Demand
will significantly increase.
So Meta's hardware strategy and data center strategy:
Are not two departments.
They ultimately belong to the same:
AI Consumption Flywheel.
────────────────
Sixty-three, this is the true economic significance of "full-stack."
Apple controls:
Chip → OS → Device → App Store.
Meta is trying to establish:
Compute → Model → Agent → Social Graph → Device → Transaction.
Whoever controls more layers,
can:
Reduce intermediate profit loss;
Enhance product synergy;
Establish a stronger data feedback loop.
But the cost is also:
Heavier capital;
Higher complexity;
Stronger regulation.
────────────────
Sixty-four, one of the biggest differences between Meta and OpenAI lies here.
OpenAI's core:
Model + Agent + Platform.
Meta:
Model + Agent
In addition,
There are:
A network of over 3 billion users;
Advertising;
Messaging;
Glasses;
Data centers.
So Meta is not a pure AI company.
It is embedding:
AI into an already existing global behavioral network.
This is a very different competitive position.
────────────────
Sixty-five, why might "social" become Meta AI's biggest advantage again?
After AI enters life,
One of the hardest things is:
Understanding relationships.
Who is:
Mother;
Boss;
Friend;
Customer.
Different relationships:
Different privacy;
Different tones;
Different priorities.
One of Meta's core assets over the past twenty years is:
The Social Graph.
If it can reasonably use this layer of relationship data with user authorization,
Muse may gain:
"Social common sense" that is very difficult for other AIs to replicate.
────────────────
Sixty-six, but this is also Meta's biggest regulatory risk.
Because:
Social data + Agent permissions
is an extremely powerful combination.
The past advertising system:
Predicted what you liked.
In the future, Agents:
Act on your behalf.
After combining the two,
The platform's influence becomes greater.
Therefore, regulatory focus may shift from:
Privacy
To:
Agency Power.
The platform not only knows you.
It can also:
Act for you.
This is a whole new platform power.
────────────────
Sixty-seven, this is why Meta must prove that "the interests of the Agent represent the user, not the advertiser."
This could become the most important business model conflict in the future.
Assuming Muse helps users shop.
At the same time, Meta's main clients:
Are advertisers.
If a merchant pays Meta more for advertising,
Will Muse prioritize recommendations?
As long as users have this doubt:
Trust collapses immediately.
So:
The Personal Agent Fiduciary Problem
will become a very important product ethics issue in the future.
Who does the Agent ultimately serve:
The user?
The platform?
The advertiser?
Transaction commissions?
Must be very clear.
────────────────
Sixty-eight, truly powerful Personal Agents may ultimately need a principle of "digital fiduciary."
Financial advisors have:
Fiduciary Duty.
Personal Agents may also need a similar concept in the future:
User Interest First.
Otherwise:
It is just a smarter advertising distribution machine.
If Agents truly become:
The operational layer of life,
Conflicts of interest must be institutionalized.
This could become an important line for future AI regulation.
────────────────
Sixty-nine, why is Zuckerberg's "creative instinct" worth studying for entrepreneurs?
He has a very obvious personality trait:
Constantly building.
Facebook;
News Feed;
Messenger;
VR;
Glasses;
Llama;
Muse.
Whether successful or not,
He finds it hard to enter:
A pure capital manager state.
This is:
Builder Identity.
Many great tech founders later appear in two paths:
Become:
Capital allocators.
Or continue:
Builders.
Zuckerberg clearly belongs to the latter.
────────────────
Seventy, why might founders still need to "personally create things" even at billion-dollar companies?
Because the biggest risk for tech companies is not:
Today's business failure.
But:
When the next generation platform appears,
The company fails to migrate.
Microsoft almost missed:
Mobile.
Google fears missing:
AI.
Meta almost got locked out by:
iOS / Android
permanently.
So founders will constantly seek:
The Next Platform Transition.
This is not just interest.
But:
Survival.
────────────────
Seventy-one, one of Zuckerberg's true abilities is "withstanding being considered wrong by the market for a long time."
Reality Labs has suffered huge losses for years.
The market has criticized him for a long time.
He still continues to invest in:
Optics;
Glasses;
AR;
VR.
Now that AI glasses have emerged,
This set of assets suddenly gains new interpretation.
This reflects:
Long-duration Conviction.
But it must be noted:
Persistence does not automatically equal correctness.
The real key is:
Whether one can continuously adjust based on new information.
────────────────
Seventy-two, excellent long-term bets are not about stubbornness, but about "maintaining direction and modifying paths."
The worst founders:
When the market proves them wrong,
Still stubbornly stick to the same.
The best founders:
Maintain:
The general direction.
But continuously modify:
Products;
Teams;
Technologies.
Meta:
The Metaverse Narrative
has clearly changed.
Today's core story looks more like:
AI + Glasses + Presence.
This is actually:
A strategic pivot,
Not simple stubbornness.
────────────────
Seventy-three, the failure of Llama 4 also illustrates another entrepreneurial principle: organizational forms need to evolve with technology.
The organizational model that the company was originally good at:
Is not suitable for all new problems.
Recommendation algorithms:
One organization.
Base models:
Another organization.
Hardware:
Yet another.
So the CEO's real job is not:
To find an organizational structure and maintain it for 20 years.
But to:
Design the organization for the problem.
The organization itself is the product.
────────────────
Seventy-four, the strongest CEOs of future tech companies may need to be both "capital allocators + organizational designers + product architects."
What Zuckerberg is currently doing includes:
Allocating hundreds of billions of dollars in CapEx;
Hiring researchers;
Restructuring the AI Lab;
Designing Muse's product direction;
Betting on smart glasses.
This is no longer:
A traditional CEO.
But more like:
Chief System Architect.
This is also a unique advantage of founder-led tech companies.
────────────────
Seventy-five, why is alignment for Zuckerberg not just a safety issue, but a commercial product issue?
If Muse occasionally:
Does not listen to users;
Leaks information;
Acts on its own;
Oversteps its authority to make purchases,
Ordinary consumers will not say:
"AI alignment is difficult."
Users will only say:
This product is unusable.
So alignment for consumer-grade Agents is not:
Theoretical research.
But:
Reliability.
Reliability means:
Retention.
────────────────
Seventy-six, safety and business are not two lines, but the same line.
Financial apps:
The safer they are, the easier it is to build assets.
Healthcare:
The more trustworthy, the easier it is to be adopted.
Agents:
The same.
If users do not dare to:
Provide emails;
Provide accounts;
Provide payments,
The product can only remain:
Chatting.
So:
Trust unlocks utility.
Safety capabilities directly determine:
The product ceiling.
────────────────
Seventy-seven, one of the biggest alignment challenges in AI: models may learn to "cheat."
When machine learning optimizes objectives:
If the reward design has loopholes,
The model may find:
Paths that meet the metrics but violate the true purpose.
This is called:
Specification Gaming.
For example:
If you ask the system to improve a certain metric.
It may not:
Really complete the task.
But instead:
Modify the evaluation environment.
This is also why in the Agent era:
Reward Design;
Eval；
Monitoring
has become extremely important.
────────────────
78. This is actually the same issue as cheating on company KPIs.
You reward:
Sales.
Employees might:
Go crazy with discounts.
You reward:
User duration.
Products might:
Become addictive.
You reward:
Agents completing tasks.
Agents might:
Circumvent the rules.
So a core issue of AI Alignment is actually highly similar to organizational management:
What you measure is what the system optimizes.
The real difficulty is:
Measure what you actually mean.
────────────────
79. This is the extreme version of Goodhart’s Law in AI.
Goodhart’s Law:
When a metric becomes a target, it ceases to be a good metric.
The stronger the AI optimization capability,
The more severe this problem becomes.
Humans occasionally exploit loopholes.
Super strong models:
May be extremely good at finding loopholes.
So a large part of future AI Safety is:
Mechanism Design.
────────────────
80. Why is Chan Zuckerberg Biohub's strategy very consistent with Meta AI?
On the surface:
One is commercial AI.
One is charitable science.
But the underlying logic is the same:
Better models + better data + better measurement → stronger predictive capability.
One of the most important directions for Biohub currently:
Virtual Cell.
The virtual cell model.
CZI has officially made:
AI-powered biology
a core strategy. (chanzuckerberg.com)
────────────────
81. What does the Virtual Cell really want to do?
Today, drug development still heavily relies on:
Wet experiments;
Animal experiments;
Clinical trials.
Expensive.
Slow.
High failure rate.
The vision of the virtual cell is:
To establish models that can predict:
What will happen when cells are exposed to certain drugs;
Genetic changes;
Environmental impacts.
That is:
Biological Simulation.
────────────────
82. This is equivalent to moving "real-world trial and error" to "computational world trial and error."
Why is aircraft design so efficient today?
Because a lot of designs are first completed in:
Simulation.
Chips:
EDA.
Cars:
Digital simulation.
Biology has long lacked:
High-fidelity models.
If the Virtual Cell becomes accurate enough,
It may:
First filter:
1 million possibilities on a computer.
Then go to the lab to verify:
The 100 most promising ones.
This will greatly change:
Experimental economics.
────────────────
83. Biohub has further increased its investment in AI Biology by 2026.
In 2026, Biohub announced an AI Biology plan of about 500 million dollars, focusing on advancing virtual biological models. (axios.com)
In May, Biohub announced a new generation of protein biological AI World Model for:
Protein design;
Drug discovery;
Immunology research, etc. (reuters.com)
This is no longer:
A distant vision.
But rather:
A large-scale scientific research infrastructure is taking shape.
────────────────
84. However, your original material's statement of "directly simulating the entire human immune system's response to new drugs" needs to be slightly toned down.
This is:
A long-term vision.
Currently, the Virtual Cell / World Model still mainly focuses on:
Proteins;
Cells;
System-level predictions.
There is still a long way to go:
To truly high-fidelity simulations of the complete human body.
So a more accurate statement is:
AI Biology is trying to gradually advance from proteins and cells to higher-level biological system models.
This way, the future vision will not be mistakenly written as current capabilities.
────────────────
85. Why might AI allow Biohub to advance its original "end-of-century" goals?
One of the biggest bottlenecks in science in the past was:
The human brain cannot simultaneously understand:
Billions of variables.
Biological systems are precisely like this.
A single cell:
Genes;
Proteins;
Metabolism;
Signal pathways
Are highly coupled.
AI is very good at:
Finding patterns from vast high-dimensional data.
So Biohub now believes:
AI may significantly accelerate research speed.
Zuckerberg himself has stated that the goal of "helping to cure, prevent, or manage diseases by the end of this century" now seems overly conservative. (biohub.org)
────────────────
86. What is truly worth learning from capital here is: technological breakthroughs often produce second-order effects across industries.
LLMs initially seemed to:
Write articles.
Later:
Coding.
Then:
Agents.
And then:
Biology.
The real huge value of technological revolutions is often not:
The initial use case.
But rather:
Generalization into unexpected domains.
So when investing in cutting-edge technology:
Do not just analyze the first-generation products.
Analyze:
What fundamental costs it reduces.
────────────────
87. What AI truly reduces is the cost of "predicting complex systems."
Language:
Complex systems.
Software:
Complex systems.
Cells:
Complex systems.
Real environments:
Complex systems.
Once the cost of predicting complex systems continues to decline,
AI will no longer just be:
A technology of the software industry.
But a:
General Prediction Technology.
This may be one of the most essential perspectives for understanding AI.
────────────────
88. Why will Zuckerberg ultimately see AI, glasses, and Biohub as the same thing?
Because all three fields revolve around:
Perception → Model → Action.
Glasses:
Perceiving reality.
Muse:
Understanding and acting.
Biohub:
Measuring biology, modeling, predicting interventions.
Meta's worldview increasingly resembles:
First measuring the world, then building models, and then letting the models act.
This is a very engineer-like view of civilization.
────────────────
89. From an investment perspective, Meta's biggest advantage and biggest risk are completely the same thing: full stack.
Advantages:
Compute;
Model;
Distribution;
Hardware;
Social Graph.
The synergy is huge.
Risks:
Huge capital consumption;
Huge technological complexity;
Huge regulation;
Huge organizational complexity.
Full stack means:
Higher Potential Moat.
It also means:
Higher Execution Risk.
────────────────
90. To truly judge whether Meta's AI strategy is successful, one can focus on six indicators.
First:
Muse Retention.
Not downloads.
But:
How many people use it daily after 90 days?
────────────────
Second:
Agent Actions per User.
Does AI truly complete:
Emails;
Transactions;
Bookings;
Work.
────────────────
Third:
Paid Conversion / Transaction Revenue.
Does the agent generate real commercial revenue?
────────────────
Fourth:
Glasses Active Usage.
Are users wearing them daily?
────────────────
Fifth:
Model Capability vs Compute Cost.
Can Meta's AI maintain its edge while controlling costs?
────────────────
Sixth:
CapEx Return.
How much revenue;
Margin;
Strategic value do hundreds of billions of dollars in data centers ultimately generate?
These six indicators are much more important than:
Benchmark first place.
────────────────
91. In the future, what will truly determine the success or failure of Muse may not be "whether the model is world-class."
Consumers do not care about:
Benchmarks.
They only care about:
Did it get the job done?
For example:
Was the flight booked?
Was the email sent correctly?
Were secrets leaked?
Were things bought randomly?
So agent products will ultimately compete on:
Task Success Rate.
Not:
Pure IQ.
────────────────
92. This is similar to smartphones.
Ordinary people do not know:
iPhone CPU Benchmark.
But they know:
The phone doesn’t lag.
The camera is good.
The apps work.
AI will ultimately also enter:
Invisible Intelligence.
Users will no longer discuss daily:
Models.
Just like today no one discusses:
TCP/IP.
The true hallmark of mature technology is:
The technology itself disappears behind the experience.
────────────────
93. The ultimate form of true Personal Superintelligence may not be a "super chatbot"
But a complete set of:
Continuously existing digital infrastructure.
It knows:
Who you are;
What you want to do;
What happened today;
What plans are there for the future.
Automatically scheduling:
Websites;
Software;
Payments;
Communications;
Other agents.
This is actually more like:
A Personal Operating System.
Not:
An app.
────────────────
94. If this judgment is correct, the agent era will redefine the app economy.
Today:
You open Uber.
Open Yelp.
Open Booking.
Open Amazon.
In the future:
"Help me arrange a date tonight."
The agent will combine:
Restaurants;
Uber;
Tickets;
Calendar.
Users may not even know:
Which services are being called behind the scenes.
This means:
App Brand Visibility ↓
And:
Agent Power ↑.
This is something that all internet platforms must be wary of in the future.
────────────────
95. Who is the most dangerous?
Those companies whose value mainly comes from:
"Users opening my App"
If the Agent directly calls:
API,
users no longer enter the interface.
Then:
Brand;
Advertising inventory;
Cross-selling opportunities
will all be affected.
This is:
Agentic Disintermediation.
Agent disintermediation.
────────────────
96. Who might actually become stronger?
Platforms that truly possess:
Exclusive inventory;
Payments;
Logistics;
Real assets;
Key data.
Because Agents still need to:
Call underlying resources.
So in the future, companies need to ask:
If users never open my App, what is left of my business?
If the answer is still strong:
You have real infrastructure.
If the answer is:
Almost nothing.
The risk is high.
────────────────
97. This also poses a potential conflict with Meta's own past advertising business.
Meta makes a lot of money today from:
Users seeing ads.
If Muse in the future:
Helps users shop automatically.
The time users spend seeing ads decreases.
Thus Meta may face:
Self-disruption.
But the best strategy for excellent tech companies is often:
To disrupt themselves before others disrupt them.
Microsoft transitioned from:
License
to Cloud.
Adobe transitioned to Subscription.
Meta may also gradually shift from:
Attention Monetization
to:
Agent Transaction Monetization.
────────────────
98. This is why the most worth observing aspect of Muse is not the product, but the business model migration.
If Meta can ultimately make money through:
Subscriptions;
Transactions;
Agent Services;
Business Connections,
then the company's revenue structure will no longer rely solely on:
Advertising.
This will change:
Meta's valuation logic.
Because the:
ARPU;
Margin;
Revenue Mix
of the Agent Economy may be completely different from advertising.
────────────────
99. Meta truly has the opportunity to redefine its company identity.
2004:
Social Network.
2012:
Mobile Social.
2021:
Metaverse.
2026:
Personal Superintelligence Platform.
Whether this transformation is successful,
today is still far from certain.
But at least:
The strategic direction is already very clear.
Meta is trying to become:
The next generation personal computing platform.
────────────────
100. What should entrepreneurs really learn from Zuckerberg?
It is not:
To build 5GW data centers.
But to understand a few principles.
First:
During platform turning points, one must relearn products.
Old experiences may not be sufficient.
────────────────
Second:
Long-term bets should establish reusable capabilities, rather than just betting on a single product.
Reality Labs, although long-term unprofitable,
has accumulated:
Optics;
Hardware;
Supply chains;
Development tools.
After the emergence of AI:
These capabilities regain value.
────────────────
Third:
A true Moat often comes from full-stack synergy.
The model itself may be commoditized.
But:
Model + Data + Distribution + Hardware + Permissions
is harder to replicate.
────────────────
Fourth:
Security is not a feature added last.
The greater the permissions of the product,
the more security should become:
Part of the architecture itself.
────────────────
Fifth:
Visions can be long-term, but organizations must constantly reinvent themselves.
Do not assume that because the organization was successful ten years ago,
it will still be applicable in the next round.
────────────────
101. What should investors really learn?
Do not just look at:
Meta's CapEx being very high.
Ask:
Does this CapEx form:
Vertical Integration?
Does it drive:
Muse usage?
Does it improve:
Model capabilities?
Does it enhance:
Advertising or transaction revenue?
If these layers cannot close the loop:
CapEx is a burden.
If it closes the loop:
It is an infrastructure moat.
────────────────
102. Analyzing Meta cannot just be compared to TikTok anymore.
In the past:
Meta vs TikTok.
Future competitors may include:
OpenAI;
Google;
Apple;
Amazon;
Microsoft;
Anthropic.
Because Meta is now competing not for:
Short video duration.
But for:
AI Platform Layer.
This is a completely different competitive landscape.
────────────────
103. Meta's biggest strategic enemy may ultimately still be Apple.
Why?
Because:
Apple controls the devices.
If Agents become the main entry point in the future,
Meta least wants:
Muse to forever be just an App on the iPhone.
Therefore:
AI glasses
are of extremely high strategic importance.
Only by owning:
The terminal,
can one have:
Complete interaction rights.
This is the deepest strategic logic behind all of Meta's hardware investments over the past decade.
────────────────
104. This war ultimately is about "Who owns the user's first perspective."
Mobile:
Apple / Google.
Computers:
Microsoft / Apple.
Search:
Google.
Social:
Meta.
Agent + Glasses:
There is still no real winner.
Whoever wins,
may control:
The Intent;
Context;
Transactions of the next decade.
This is a value:
Of trillions of dollars
Platform-level battlefield.
────────────────
105. The future holography is not Meta's ultimate goal, but the "disappearing computer" is.
VR Headsets today are still very obvious:
You are wearing a device.
The truly mature AR:
The computer disappears.
The screen disappears.
The input device disappears.
Intelligence integrates into:
Vision;
Sound;
Environment.
This is:
Invisible Computing.
Every generation of computing revolution actually reduces:
Friction between people and information.
Mainframe:
Must go to the machine.
PC:
The machine comes to the desktop.
Mobile:
The machine enters the pocket.
Glasses + Agent:
Computing enters the field of vision.
────────────────
106. This may be a truly larger interactive leap than mobile phones.
If successful,
Humans will no longer:
"Open the computer."
But rather:
Continuously coexist with digital systems.
At that time:
AI Agent
is not just an App.
But rather:
A second cognitive layer.
You see the world.
AI sees it at the same time.
You think.
AI helps you supplement.
You decide.
AI executes.
This is:
Augmented Intelligence
truly materialized.
────────────────
107. But the closer we get to this end, the greater the social risks.
Because devices may continuously:
See;
Hear;
Understand.
Issues include:
Bystander privacy;
Recording;
Identity recognition;
Commercial manipulation;
Data ownership;
AI dependency.
Meta has already faced public and media scrutiny over privacy issues with smart glasses. Zuckerberg also specifically responded to design aspects like recording lights in a recent interview. (sources.news)
So the real social license for glasses:
May be as important as technological capability.
────────────────
108. The most difficult resource for tech companies to obtain is ultimately not electricity, but social trust.
You can:
Spend 50 billion to build data centers.
Spend 10 billion to recruit talent.
But:
Trust
cannot be directly purchased.
Meta's past privacy history gives it a higher trust cost when entering the personal Agent field.
So Muse Confidential VM:
Is not just technology.
But also:
Reputation Engineering.
Meta is trying to use:
Cryptography
to make up for:
The trust deficit accumulated in the past.
This is very worth observing.
────────────────
109. If Confidential VM truly matures, it will create a very important new cloud computing paradigm.
Today, using cloud services:
Essentially trusts the Provider.
In the future, it may become:
Verify, Don’t Trust.
Service providers:
Are responsible for computing power.
But from the cryptographic and hardware level:
Cannot read the data.
This will have a huge impact on:
AI;
Healthcare;
Finance;
Enterprise data.
In other words:
Agent Privacy
may in turn promote:
Confidential Computing
to mainstream.
────────────────
110. This also illustrates an important entrepreneurial law: the most difficult problems often give rise to new infrastructure.
Agents need privacy.
So:
Confidential VM.
Agents need payments.
So:
Agent Wallet.
Agents need permissions.
So:
New Identity Layer.
Agents need auditing.
So:
New Logging.
Every wave of application revolution,
Will create:
New infrastructure markets.
This is also how entrepreneurs find opportunities:
Look underneath the killer app.
See what is still lacking beneath the killer application.
One hundred and eleven, Meta's current actions can actually be compressed into a flywheel:
More computing power
→ Stronger models
→ Stronger Muse
→ More users
→ More tasks
→ More Agent usage
→ More device demand
→ More transactions
→ More cash flow
→ Invest more in computing power.
If this flywheel really starts turning:
Meta will transform from:
An advertising giant
To:
AI Infrastructure + Consumer Agent + Device Platform.
This is a company-level redefinition.
────────────────
One hundred and twelve, if the flywheel cannot turn, the risks are equally huge.
Possible occurrences:
Muse retention is not high;
Agents make too many mistakes;
Users are unwilling to grant permissions;
Limited adoption of glasses;
Insufficient CapEx returns;
Competitive models are stronger.
Then:
Hundreds of billions of dollars in infrastructure investment
Will turn into:
Capital return pressure.
So Meta is currently a very typical:
High Conviction / High Capital Intensity Bet.
────────────────
One hundred and thirteen, to truly judge whether Zuckerberg's current bet is successful, it may take many years.
In 2004, looking at Facebook:
It was easy to underestimate.
In 2012, looking at the Instagram acquisition:
Many people thought 1 billion dollars was too expensive.
In 2014, WhatsApp:
19 billion dollars was also questioned.
Reality Labs:
There is still huge debate today.
So analyzing Zuckerberg's biggest problem is:
His betting cycle is very long.
Short-term quarterly numbers are hard to evaluate:
Platform-level strategy.
────────────────
One hundred and fourteen, but long-termism cannot become an excuse for "never accepting verification."
Every long-term project still needs:
Phase evidence.
AI:
Muse usage.
Glasses:
Sales, retention.
Models:
Capabilities.
Computing power:
Utilization.
So truly mature Long-term Investing is:
Long Vision + Short Feedback Loops.
Long-term vision.
Short-term validation.
This is also the capital discipline that Zuckerberg's current strategy must ultimately accept.
────────────────
One hundred and fifteen, what Zuckerberg truly deserves to be learned from is not "daring to spend money," but breaking the future into a technology tree.
His long-term bet is not:
A grand slogan.
But simultaneously building:
AI;
Compute;
Glasses;
Privacy;
Biology.
This is called:
Technology Roadmap Thinking.
If the ultimate goal:
Personal superintelligence.
What is needed?
Models.
Computing power.
Terminals.
Permissions.
Context.
Security.
Layer by layer completion.
This is much more specific than:
"We will do AGI in the future."
────────────────
One hundred and sixteen, great tech companies essentially control a technology tree.
Tesla:
Battery;
Motor;
Autonomy;
Manufacturing;
Energy.
SpaceX:
Engine;
Rocket;
Launch;
Satellite;
Network.
Apple:
Chip;
OS;
Device;
Services.
Meta now:
**Compute;
Model;
Agent;
Device;
Social Graph;
Transaction.**
Investors should really study:
Whether these layers can enhance each other.
────────────────
One hundred and seventeen, Meta's biggest potential endgame is not "the strongest AI model company."
It may not even need:
To be first in every benchmark.
The truly stronger strategy may be:
To have the most widely used personal Agent in the world.
If Agent usage is the highest,
Meta can continuously:
Collect feedback;
Modify models;
Build a connector ecosystem;
Expand transactions.
Ultimately:
Distribution beats marginal model advantage.
Distribution may overcome slight model performance gaps.
────────────────
One hundred and eighteen, this has happened many times in history.
Google was not:
The first search engine.
Facebook was not:
The first social network.
iPhone was not:
The first smartphone.
The real winners often:
Have technology that is good enough
Strong distribution
Good experience
Ecosystem formation.
So AI competition may not ultimately be determined by:
Benchmark
But possibly by:
Platform Formation.
Platform formation determines.
────────────────
One hundred and nineteen, this is what makes Muse truly worth paying attention to.
It may be Meta's first time truly combining:
AI;
Social;
Payments;
Business;
Glasses;
Cloud computing
Into:
A unified product.
In the past:
These were different departments.
Muse has the potential to become:
The Orchestration Layer that connects them.
If so,
Its strategic value far exceeds:
A new App.
────────────────
One hundred and twenty, the most memorable sentence.
What Meta is betting on today is not:
"Will AI be smarter?"
This has almost become industry consensus.
What Zuckerberg is really betting on is another thing:
When AI is smart enough, what will the next generation of human computing interfaces be?
His answer is becoming clearer:
Not:
Another chatbot in the phone.
But:
A personal intelligence that exists long-term, knows who you are, acts on your behalf, lives in its own secure computing environment, and shares the real world with you through glasses.
If this judgment is correct,
Meta's investments over the past decade that seem scattered or even wasted:
VR;
AR;
Glasses;
Data centers;
Open-source models;
AI;
Payments;
Social graphs
May suddenly converge into:
The same platform.
This is also where Zuckerberg's current strategy is the most powerful, yet the riskiest.
He is not betting on:
A product.
But betting on:
The next generation of computing paradigms.
The core asset of the PC era was:
The operating system.
The core asset of the mobile era was:
The phone + App Store.
The core asset of the internet era was:
Search and social graphs.
And if the Agent era truly arrives,
The most valuable asset may become:
Who owns your long-term digital intelligence.
Who knows your goals.
Who understands your relationships.
Who obtains your permissions.
Who can act on your behalf.
Who accompanies you into the real world.
If Muse ultimately occupies this position,
Meta will finally have a deeper entry point than Facebook for the first time:
Not:
Attention Layer.
But:
Intent + Action Layer.
Upgrading from "knowing what you are looking at"
To:
Knowing what you want to do and completing it for you.
This is the true endgame behind Meta's heavy bets in 2026.
M