About a month ago, when Meta invested over $10 billion in Scale AI and its CEO switched to Meta, I began following their every move. They have since poached top talent in the large model field—people like Pang Ruoming and Zhao Shengjia—and it’s rumored that Pang Ruoming could earn up to $200 million over four years.
This is clearly outrageous. Meta is now pulling out all the stops, but after spending so much money and poaching so many people, what exactly are they planning to do?
This morning, I saw a video of Zuckerberg introducing the future vision of the Superintelligence Lab. In short, it centers on the “individual.”
Four Key Capabilities: helping individuals ① achieve their goals, ② create what they envision, ③ experience any adventure, and ④ become the companions and versions of themselves they aspire to be.
Zuckerberg explained that while most experts in the industry view AI as primarily a tool for “centralized automated production,” he disagrees (quite a distinctive stance). He believes that the direction of technology use should be guided by each individual.
It sounds lofty—so what does it really mean?
Centralized Automated Production vs. Individually Driven Technology Use
What exactly does “centralized automated production” refer to? How is it different from “individually driven technology use”? After all, tools like ChatGPT, Claude, and Gemini are used by everyone for their own purposes—doesn’t that count as the latter?
When we use current AI tools such as GPT or Gemini, the central brain and computing power are under the control of big companies; we are merely calling APIs. In other words, the results we see are produced by the platform and then redistributed.
Even when using GPT, you might run into issues like IP controversies leading to service restrictions. Basically, the AI isn’t truly in your hands—others can shut you off at any time.
In that regard, “centralized automated production” might mean:
- Single Superintelligence: Imagine a massive “cloud factory” where all reasoning, decision-making, content, and code are produced by one model.
- Embedded Values: A small group of developers or regulators sets the model’s objectives, and then that model generates most of society’s wealth.
- Human Positioning: Most people no longer work directly but receive “dividends” from AI output or a form of universal basic income. Zuckerberg calls this approach “centralized authoritarian control,” expressing concern over further concentration of power and resources.
Zuckerberg’s “individual-driven” idea, in contrast, envisions a decentralized or multi-agent ecosystem—where AI instances are embedded in smartphones, glasses, or private clouds, remain with the same user over time, maintain long-term memory focused on personal life, and are directly driven by and benefit the individual.
- Personal Superintelligence: Meta imagines AI as a long-term partner that knows your schedule, remembers family birthdays, helps you write code, and composes music.
- Hardware Platforms: Devices such as Ray-Ban AI glasses, smartphones, and AR headsets serve as the main interfaces, keeping AI “with you all the time” rather than as a distant hub.
- Multi-Agent Ecosystem: Instead of one “central brain,” countless specialized micro-models could collaborate.
So why do products like ChatGPT, which seem available to everyone, still fall under centralized control?
First, it’s the deployment model: all user requests go back to the centralized computing pools of OpenAI or Google, with decision-making power and data centralized. Then there’s the unified safety framework—content filtering and feature management are centrally controlled by the platform. This is precisely what Zuckerberg fears about excessive concentration of technological power.
He also mentioned a “bottom-up” approach where individual technology use drives prosperity.
When personal computers emerged in the 1980s, the Macintosh brought GUIs and creative tools to the masses, sparking the desktop publishing and digital design industries. Later, internet blogging and platforms like YouTube ignited a cultural explosion. Then came smartphones and app stores, where millions of independent developers delivered services to billions worldwide, giving birth to a host of super apps.
Each of these developments shifted control and innovative momentum from the center to the individual, resulting in exponential, long-tail growth in economic and cultural impact.
Meta portrays personal superintelligence (PSI) as the next leap—a way to condense nearly “Ph.D.-level” cognitive and creative ability into something you can carry in your pocket or wear on your glasses, empowering ordinary individuals to perform professional tasks.
Meta’s Rocky Road in AI
It all sounds grand. This time, the recruitment effort is massive—sending shockwaves through Silicon Valley as companies worry that their core personnel might be the next targets.
So why has Meta suddenly become so aggressive? The Financial Times described it as having “more cohesion” than before.
Consider this:
Today, Meta’s AI strategy is clear: with the newly established Meta Superintelligence Labs, a dual-track structure combining AI Products and AGI Foundations, and billions of dollars in computing power investments, they aim to align all research, product, and hardware narratives to one goal.
In contrast, back in 2023 the strategy resembled a jumble of loosely connected pieces. In February of that year, after the formation of the Generative AI Product Group, Meta soon released the open-source LLaMA model, while FAIR and Reality Labs operated without a unified product strategy. Externally, they juggled the "AI" and "metaverse" narratives simultaneously, diluting resources and investor attention.
This brings us to LLaMA’s humiliating history and Meta’s struggles with large language models—not to mention that they even avoid using LLaMA internally. 🤣
| Date | Key Verified Event | Impact and Details |
|---|---|---|
| 2023-03-03 | Less than a week after the LLaMA-1 paper was released, complete weight packages were uploaded to a 4chan seed forum and spread rapidly. | - The research license became void instantly, allowing anyone to retrain or modify the model. <br>- Meta attempted to issue GitHub takedown notices, but it was too late; the model became “the starting point for open source for everyone.” <br>- U.S. senators questioned Meta over safety concerns. |
| 2023-07-18 | Llama 2 switched to the “Llama Community License”: semi-commercial use is allowed, but users with over 700 million monthly active users need additional authorization. | - The OSI explicitly stated it was “not truly open source,” and the restrictive terms caused community backlash. <br>- Entrepreneurs and developers shifted to more lenient newcomers like Mistral or DeepSeek, resulting in LLaMA downloads being outpaced. |
| 2024-04-02 | The internal coding assistant, Metamate, was launched; it defaults to using LLaMA and falls back to GPT-4 for more complex tasks. | - Employees joked that “LLaMA ≠ Meta LLM,” reflecting diminished confidence in the in-house model. <br>- It exposed LLaMA’s limitations in code reasoning and long-chain operations compared to GPT-4. |
| 2024-12-04 | An anonymous memo from a departing employee revealed “direction drift and a lack of consensus,” criticizing management for focusing solely on “milestone releases.” | - Wall Street analysts downgraded expectations for LLaMA commercialization, and Meta’s stock fell 8% that month. |
| 2025-04-05 | Three MoE versions of LLaMA 4 were released: Behemoth (2T parameters, pending completion), Maverick, and Scout. | - Officially, they claim these models rival GPT-4o, but independent evaluations show Maverick lagging behind GPT-4o and DeepSeek V3 in complex reasoning and programming tasks. <br>- The community accused Meta of “benchmark chasing,” saying multiple tests were tainted by data “pollution”; Meta’s leadership had to clarify overnight. |
In just over a year, LLaMA suffered a triple setback—“leak, licensing controversy, and performance lag”—missing the best window to capture the ecosystem. Confidence from employees and investors in LLaMA dropped, leading to a talent exodus (11 out of the 14 original authors left).
Now, with time running out, Meta feels compelled to spend big and move aggressively.
So what does this “increased cohesion” specifically look like?
| Old Model (≤ 2023) | New Model (from 2025) |
|---|---|
| Multi-centric: FAIR, Gen-AI, and Reality Labs each managing their own area | Meta Superintelligence Labs (MSL) becomes the sole “top-level lab” |
| Long chain from model research to product migration | AI Products and AGI Foundations operate in parallel, with a demo iteration every two weeks and a unified metric system |
| Computing power allocated by each business unit | A Hyperion-grade 5 GW data center pool is established, reserved directly for MSL |
| Separate recruitment processes | Top executives are directly poaching talent, reportedly offering nine-figure contracts to 50+ top scientists |
Zuckerberg has even stepped in personally to recruit.
Moreover, superintelligence holds enormous authority—it functions like an internal startup, unbound by typical corporate constraints while overseeing all AI-related operations in the company. Zuckerberg has also promised to inject significant funding.
Why Does Meta Dare?
They Have the Money
After research by BNP Paribas, it is estimated that hiring for superintelligence could lead to an extra $1.5–3.5 billion in spending. In an interview with The Information, Zuckerberg bluntly stated that the company has the funds and that the current business model can support such endeavors.
In Q2 2025, Meta’s advertising revenue reached $4.6 billion, maintaining high levels over recent years—with 97% of revenue coming from ads. The profit margins in advertising are enormous.
With Q1 net profits of $16.6 billion and Q2 at $18.3 billion—$35 billion earned in just half a year—even an additional annual spend of $1.5–3.5 billion would only be 8–19% of a single quarter’s profit. Financially, they can handle it.
They’re buying talent at a premium and spending heavily on data centers—the money is flowing.
In its May financial report, Meta raised its annual capital expenditure forecast by as much as 10%—between $64 billion and $72 billion—citing “additional data center investment” to support AI development and “increased expected hardware costs.” Furthermore, Meta even purchased a nuclear power plant’s electrical output for the next twenty years to support its AI projects.
Yet despite the apparent opulence, there are whispers that Meta is seeking to raise $29 billion to fund its full-scale AI strategy, and that it is looking to private capital firms to help finance US data center construction.
They Have the Resources
Beyond money, what else does Meta have? When you think of Meta, you naturally think of Facebook, Instagram, and WhatsApp—products that combined have over three billion daily active users. Imagine how quickly a product can scale once launched, especially with advertising as their ace.
“As the saying goes, 'A craftsman must have sharp tools.'”
They are also building up three key resource dimensions simultaneously—computing power, data, and hardware endpoints—which integrate seamlessly with their existing social-advertising ecosystem.
For computing, Meta is planning multiple “Titan-class” data centers. For example, Prometheus in Ohio is designed to be the first petawatt-scale training supercomputer, with over 500,000 GPUs and an on-site dual gas-powered generator setup.
Why such scale? For speed.
A single campus housing over 500,000 GPUs on a low-latency InfiniBand/400 Gbps network can train a model with over 10 trillion parameters in parallel, drastically reducing cross-site synchronization and compressing training cycles from months to weeks.
This scale also attracts top researchers. Who else can deploy resources so lavishly?
Consider the talent Zuckerberg has recruited:
Noam Shazeer (founder of Switch-ML and former Google Brain member) and Boris Power (former Anthropic MoE chief engineer) are experts in MoE model architectures—a field where Meta has limited experience.
Haoliang Jiang, the technical lead for long-context in Claude, can help solve issues like gradient explosions in extremely long contexts and inefficiencies in position encoding and caching—crucial for building a personal “memory assistant” that must process massive event sequences.
Alexander Kirillov (of SAM) and the Jaegle brothers (of the Perceiver series) are multimodal experts who can transform data from Ray-Ban glasses into training datasets.
There are also specialists in privacy and on-device fine-tuning.
One might compare it to racing: top talent are the drivers, the computing-powered data centers are the cars, and data—like fuel—determines how fast and how long they can run. Perhaps this is why Meta spent $14.3 billion to invest in Scale AI—to lock in a high-quality data annotation team and reduce reliance on external vendors.
Even the best car needs a driver to harness its full potential, and the quality and quantity of fuel determine how fast and far it can go.
Finally, let’s not forget Meta’s ongoing work on hardware, like the Ray-Ban Meta smart glasses (with projected shipments exceeding 2 million units in 2024–25). These personal devices are the carriers of personal superintelligence, equipped with 24-hour multimodal sensors and private domain contextual inputs.
There are also smartwatches and headsets that provide immersive rendering and input from gestures or physiological data, enabling you to “experience any adventure.” With glasses selling over 2 million units annually and more than 4 million active cameras on endpoints, the infrastructure is already scaling—if they don’t accelerate now, when will they start?
AI + Meta??
In his public letter, Zuckerberg mentioned that their goal is to help users achieve their ambitions, create their ideal world, and experience any adventure. This is reminiscent of deeply integrating large AI models with VR/AR hardware to create something akin to a metaverse.
The promise of adventure and creation immediately brings to mind the film “Ready Player One.”
Meta envisions smart glasses as the future’s primary computing device. These devices are naturally suited to harness the functions of a personal AI assistant (PSI): equipped with cameras and microphones, the glasses allow AI to “see what we see and hear what we hear,” constantly understanding the user’s environment and offering assistance.
With AR glasses and AI, users can overlay their desired digital information and virtual elements onto the real world, effectively “remodeling” their environment. For example, you might project a virtual cinema screen in your living room or have an AI generate a cute virtual pet that interacts with you at your desk.
In the future, what you see through your glasses could be a unique augmented reality: you could choose to render your surroundings in a sci-fi style or as a forest path, with AI instantaneously transforming buildings into your preferred style and displaying your personal digital art in the sky.
Zuckerberg’s vision is to have personal AI help people “create the world you want to see and experience any adventure.” In Meta’s envisioned metaverse, everyone could become a creator—customizing the style of a virtual home, building a personal social island exclusive to one’s friend circle, or even letting AI generate unique worlds based on your dreams for exploration.
The metaverse would then extend beyond a large social platform to become an extension of reality and a personalized projection.
Of course, these are early ideas—I’m not exactly sure how they plan to implement them.
While the letter emphasizes “individual-driven” development, it still relies on their own supercomputers, likely resulting in a hybrid center-edge model. The so-called “individual-driven” approach may really be more about “privacy isolation + long-term personalization + on-device fine-tuning,” rather than being completely decentralized.
Moreover, building a personal super assistant for billions of people is an extremely ambitious goal. The required computing resources and the affordability of such devices might pose significant challenges. For example, a pair of Meta glasses might cost several hundred dollars, and if additional on-device computing power is needed, the price could climb further. Prices may eventually drop over time—as they have with smartphones—but initially, this could limit adoption.
We might end up with a situation where resource-rich individuals enjoy a 24/7 AI mentor and a personal lab, accelerating their talent development. In contrast, those without such resources remain tied to traditional education, further amplifying information gaps. Wealthy users can instantly generate code, 3D models, and content with zero barriers, even using AI for financial management, creating a positive feedback loop.
The ultimate outcome could be an exacerbation of “AI-driven wealth inequality,” intensifying the Matthew Effect.