Meta's AI Shows Early Signs of Self-Improvement, Zuckerberg Says Superintelligence Is in Sight

The Journey Toward Artificial Superintelligence

On July 30, 2025, the same day Meta reported its second-quarter results, Mark Zuckerberg posted a short letter on Meta's website under the title "Personal Superintelligence." Its opening lines were the ones that traveled: "Over the last few months we have begun to see glimpses of our AI systems improving themselves. The improvement is slow for now, but undeniable. Developing superintelligence is now in sight"¹.

Quick Summary

On July 30, 2025, Mark Zuckerberg published a letter titled "Personal Superintelligence." He wrote that Meta had begun to see "glimpses of our AI systems improving themselves," that the improvement "is slow for now, but undeniable," and that "developing superintelligence is now in sight"¹. He gave no date and published no technical details.

The letter followed a month of moves: Meta invested in Scale AI and hired its founder, Alexandr Wang¹², created Meta Superintelligence Labs on June 30¹³, and on July 30 said it expected 2025 capital expenditures of $66 billion to $72 billion¹⁶.

The letter came at the end of a decade of AI research at Meta, two and a half years of openly released language models, and a summer in which the company reorganized its AI teams, made a multibillion-dollar investment in a data-labeling company and raised its spending forecast again. This article sets out what Meta said, what it built on the way, what it is spending, and what it has not disclosed.

A Decade of Groundwork

According to Meta, its Fundamental AI Research lab, FAIR, launched in late 2013². Much of Meta's standing in AI research came from tools it released publicly. When it launched Llama 2 in 2023, the company described PyTorch as "today's leading AI framework created by Meta and the AI community" and noted that Meta and Microsoft were among the founding members of the PyTorch Foundation³. In the same announcement Meta pointed to React, its framework for web and mobile applications, as another example of internal engineering that became "commonly used infrastructure for the entire technology industry"³.

Zuckerberg has framed that history as a business decision as much as a research one. In a July 2024 letter he wrote that Meta had "saved billions of dollars by releasing our server, network, and data center designs with Open Compute Project and having supply chains standardize on our designs," and that it had "benefited from the ecosystem's innovations by open sourcing leading tools like PyTorch, React, and many more tools"⁴.

The hardware came alongside the software. In 2022 Meta described its AI Research SuperCluster, which had 16,000 Nvidia A100 GPUs and, according to Meta engineers, "played and continues to play an important role in the development of Llama and Llama 2"⁵. In March 2024 Meta announced two new clusters of 24,576 Nvidia H100 GPUs each, which it said it was using to train Llama 3. The same post stated the goal plainly: "Meta's long-term vision is to build artificial general intelligence (AGI) that is open and built responsibly so that it can be widely available for everyone to benefit from"⁵.

The engineers also described the day-to-day load that justifies that hardware. "At Meta, we handle hundreds of trillions of AI model executions per day," they wrote⁵.

From LLaMA to Llama 4

On February 24, 2023, Meta released LLaMA, short for Large Language Model Meta AI, to researchers under a noncommercial license, with access granted case by case to academic, government, civil society and industry research groups⁶. Meta later said the model drew "more than 100,000 requests for access"³.

Five months later, on July 18, 2023, Meta and Microsoft introduced Llama 2. This time the model was "free for research and commercial use," with model weights and starting code for both the pretrained model and conversational versions. Microsoft became the preferred partner, Llama 2 went into the Azure AI model catalog, and it was also offered through Amazon Web Services and Hugging Face³. Meta said its fine-tuned models had been red-teamed internally and externally and published a Responsible Use Guide and an Acceptable Use Policy alongside the release. It also launched an Open Innovation AI Research Community for academic researchers and a Llama Impact Challenge to encourage public, non-profit and for-profit groups to use the model on "environmental, education and other important challenges"³.

In January 2024 Zuckerberg told The Verge that Meta's goal had widened. "We've come to this view that, in order to build the products that we want to build, we need to build for general intelligence," he said. He added that "a lot of the best researchers want to work on the more ambitious problems"⁷. In the same interview he said Meta would own more than 340,000 Nvidia H100 GPUs by the end of 2024, and almost 600,000 GPUs in total once other chips were counted⁷.

On April 18, 2024, Meta upgraded its assistant, Meta AI, with Llama 3, calling it "the most intelligent AI assistant you can use for free." It rolled Meta AI out in English to more than a dozen countries outside the United States, including Australia, Canada, Nigeria, Pakistan, Singapore and South Africa, added it to search in Facebook, Instagram, WhatsApp and Messenger, and opened a standalone website, meta.ai⁸.

In July 2024 Meta released Llama 3.1 405B, which Zuckerberg called "the first frontier-level open source AI model," together with updated 70B and 8B models⁴. He wrote that "starting next year, we expect future Llama models to become the most advanced in the industry"⁴.

On April 5, 2025, Meta released Llama 4 Scout and Llama 4 Maverick as open-weight models. According to Meta, both use a mixture-of-experts design with 17 billion active parameters; Scout has 16 experts and a context window of 10 million tokens, and Maverick has 128 experts. Meta also described a much larger teacher model, Llama 4 Behemoth, with 288 billion active parameters and nearly two trillion total, which was still training at the time⁹. On April 29, 2025, Meta launched a standalone Meta AI app built with Llama 4 and merged it with the companion app for its Ray-Ban Meta glasses¹⁰. The next day, reporting first-quarter results, Zuckerberg said Meta AI "now has almost 1 billion monthly actives"¹¹.

The Summer of 2025

The steps that led directly to the letter came quickly:

June 12, 2025: Scale AI announces "a significant new investment from Meta" that values Scale at over $29 billion, and says founder Alexandr Wang is joining Meta¹².

June 30, 2025: Zuckerberg announces Meta Superintelligence Labs (MSL) in an internal memo later published by CNBC¹³.

July 14, 2025: Zuckerberg says Meta is building Hyperion, a data center that will scale to five gigawatts¹⁴.

July 25, 2025: Shengjia Zhao, formerly of OpenAI, is named MSL's chief scientist¹⁵.

July 30, 2025: Meta reports second-quarter results and Zuckerberg publishes "Personal Superintelligence"¹⁶,¹.

From Artificial General Intelligence to Superintelligence

The words are used loosely across the industry. The philosopher Nick Bostrom offered a definition in a paper first published in 1998: "By a 'superintelligence' we mean an intellect that is much smarter than the best human brains in practically every field, including scientific creativity, general wisdom and social skills"¹⁷. He added that the definition "leaves open how the superintelligence is implemented"¹⁷. The same paper argued for "believing that we will have superhuman artificial intelligence within the first third of the next century," a forecast Bostrom based on estimates of the brain's processing power and how quickly computer hardware would catch up¹⁷.

Zuckerberg has avoided a precise definition. "I don't have a one-sentence, pithy definition," he told The Verge in 2024. "You can quibble about if general intelligence is akin to human level intelligence, or is it like human-plus, or is it some far-future super intelligence. But to me, the important part is actually the breadth of it"⁷. He also said he was "not actually that sure that some specific threshold will feel that profound"⁷.

Others had already put the word in their plans. In June 2024 Ilya Sutskever, Daniel Gross and Daniel Levy announced Safe Superintelligence Inc. with the line "Superintelligence is within reach," describing "the world's first straight-shot SSI lab, with one goal and one product: a safe superintelligence"¹⁸. A year later Gross joined Meta¹³.

Inside Meta, the long-term research group has used its own term. In June 2025, a few weeks before the letter, FAIR released V-JEPA 2, a "world model" trained on video that Meta said helps robots and other AI agents "understand the physical world and predict how it will respond to their actions." Meta described it as "meaningful progress toward our ultimate goal of developing advanced machine intelligence (AMI)"¹⁹. In Meta's lab tests, robots used the model to reach for, pick up and move objects they had not seen before¹⁹. TechCrunch noted that FAIR focuses on "techniques that may be used five to 10 years from now"¹⁵.

The July 2025 letter does not set a timeline. It says only that "in the coming years, AI will improve all our existing systems and enable the creation and discovery of new things that aren't imaginable today"¹. It gives no date for human-level AI or for superintelligence.

What Meta Has and Has Not Disclosed

The self-improvement claim rests on two sentences in the letter¹. The letter did not come with a paper, benchmark or system description explaining what that improvement is or how it was measured, so any account of the mechanism goes beyond what Meta said. The earnings release published the same day repeated the ambition in one line from Zuckerberg: "I'm excited to build personal superintelligence for everyone in the world"¹⁶.

What Meta did disclose that day was financial: its spending plans, its headcount and its user numbers, covered below. Readers following the story should separate the two kinds of statement. The spending and the hires are documented. The capability claim is Meta's own description.

Understanding the Superintelligence Vision

The Personal AI Revolution

Most of Zuckerberg's letter is about who superintelligence should serve. "Meta's vision is to bring personal superintelligence to everyone," he wrote. "We believe in putting this power in people's hands to direct it towards what they value in their own lives"¹. He contrasted that with "others in the industry who believe superintelligence should be directed centrally towards automating all valuable work, and then humanity will live on a dole of its output"¹.

He described personal AI "that knows us deeply, understands our goals, and can help us achieve them," and argued that "an even more meaningful impact on our lives will likely come from everyone having a personal superintelligence that helps you achieve your goals, create what you want to see in the world, experience any adventure, be a better friend to those you care about, and grow to become the person you aspire to be"¹.

The same idea already shapes Meta's products. The Meta AI app launched in April 2025 was billed as "the assistant that gets to know your preferences, remembers context and is personalized to you." Users can tell it to remember details about them, and in the United States and Canada it can draw on information from their Facebook and Instagram profiles and activity to personalize answers¹⁰.

Real-World Applications Today

The personal assistant Zuckerberg describes already exists in an early form. When Meta upgraded Meta AI with Llama 3 in April 2024, it gave examples of what people could ask: a restaurant "with sunset views and vegan options," concerts for a Saturday night, or help explaining "how hereditary traits work" before a test⁸. It added Meta AI to search in its apps, so someone planning a ski trip in a Messenger group chat could ask it to "find flights to Colorado from New York and figure out the least crowded weekends to go" without leaving the app⁸. It also sped up its Imagine image generator so pictures appeared and changed as people typed⁸.

The Meta AI app added a Discover feed where people can share and remix prompts, voice conversations, and a test of "full-duplex speech technology" that generates voice directly instead of reading out written answers. Meta said the voice demo "doesn't have access to the web or real-time information" and warned users they "may encounter technical issues or inconsistencies"¹⁰.

Outside Meta, developers use the open models in their own products. When Llama passed one billion downloads in March 2025, Meta said Spotify had used it "to help deliver customized, contextualized recommendations for new songs, artists, podcasts or audiobooks" and to enrich commentary from its AI DJs²⁰.

Glasses as the Next Computer

Zuckerberg's letter predicts that "personal devices like glasses that understand our context because they can see what we see, hear what we hear, and interact with us throughout the day will become our primary computing devices"¹.

Meta has been building toward that for several years. In September 2023, with EssilorLuxottica, it announced a new generation of Ray-Ban Meta smart glasses starting at $299, with a 12 MP camera, a five-microphone array, livestreaming to Facebook and Instagram, and the ability to talk to Meta AI by saying "Hey Meta"²¹. In its 2024 annual report Meta said it had unveiled Orion, "a pair of true AR glasses that overlay content on top of the physical world," as a prototype²².

Meta's own announcement of Orion, in September 2024, was candid about the difficulty. Building holographic displays, new input methods and AR software into something that looks like ordinary glasses was, Meta wrote, "so challenging that we thought we had less than a 10% chance of pulling it off successfully." It said Meta AI runs on Orion and "understands what you're looking at in the physical world," giving the example of opening a refrigerator and asking for a recipe based on what is inside²³. Meta gave access only to employees and "select external audiences" and said it was building toward a consumer product line²³.

In the June 2025 memo announcing MSL, Zuckerberg listed glasses among Meta's advantages: "We are pioneering and leading the AI glasses and wearables category that is growing very quickly"¹³. When Meta launched the Meta AI app, it said "glasses have emerged as the most exciting new hardware category of the AI era"¹⁰.

Economic and Societal Transformation

The letter places superintelligence in a long history of productivity gains. "As recently as 200 years ago, 90% of people were farmers growing food to survive," Zuckerberg wrote. "Advances in technology have steadily freed much of humanity to focus less on subsistence and more on the pursuits we choose"¹.

He predicted that "if trends continue, then you'd expect people to spend less time in productivity software, and more time creating and connecting," and closed with a claim about timing: "The rest of this decade seems likely to be the decisive period for determining the path this technology will take, and whether superintelligence will be a tool for personal empowerment or a force focused on replacing large swaths of society"¹. These are predictions, and the letter presents them as such. It ends with a statement of intent: "We have the resources and the expertise to build the massive infrastructure required, and the capability and will to deliver new technology to billions of people across our products"¹.

Technical Infrastructure and Investment Strategy

Massive Computing Power Requirements

Meta's spending plans rose three times in the first seven months of 2025. In January the company reported capital expenditures of $39.23 billion for 2024 and forecast $60 billion to $65 billion for 2025. At that point it said "the majority of our capital expenditures in 2025 will continue to be directed to our core business"²⁴.

In April Meta raised the forecast to $64 billion to $72 billion. It said the "updated outlook reflects additional data center investments to support our artificial intelligence efforts as well as an increase in the expected cost of infrastructure hardware"¹¹.

In July it narrowed the range to $66 billion to $72 billion, "up approximately $30 billion year-over-year at the mid-point." Second-quarter capital expenditures alone were $17.01 billion. Meta also said it expected "another year of similarly significant capital expenditures dollar growth in 2026"¹⁶. That guidance covers Meta's total capital spending and is not a budget for one lab.

Operating costs are rising too. Meta expected 2025 total expenses of $114 billion to $118 billion, growth of 20% to 24%, and said employee compensation would be the second-largest driver of 2026 expense growth "as we add technical talent in priority areas"¹⁶. Headcount was 75,945 at the end of June 2025, up 7% from a year earlier¹⁶.

Meta can fund this because its advertising business is very profitable. In the second quarter of 2025 revenue rose 22% to $47.52 billion, ad impressions rose 11% and the average price per ad rose 9%. Meta forecast third-quarter revenue of $47.5 billion to $50.5 billion and ended June with $47.07 billion in cash, cash equivalents and marketable securities¹⁶. Its Family of Apps segment earned $87.11 billion in operating income in 2024, while Reality Labs, the hardware and metaverse division that makes its glasses and headsets, lost $17.73 billion²². In the second quarter of 2025 Reality Labs lost another $4.53 billion¹⁶. Zuckerberg has rejected the idea that AI replaced the metaverse as Meta's priority. "I don't know how to more unequivocally state that we're continuing to focus on Reality Labs and the metaverse," he told The Verge in January 2024⁷.

Data Center Expansion and Innovation

On July 14, 2025, Zuckerberg said Meta was building a data center called Hyperion. A Meta spokesperson told TechCrunch it would be in Louisiana, likely in Richland Parish where Meta had earlier announced a $10 billion development, and that Meta planned to bring two gigawatts of capacity online there by 2030 and scale to five gigawatts over several years. Zuckerberg said Hyperion's footprint would be large enough to cover most of Manhattan¹⁴.

Scaling up has engineering costs that Meta has described in detail. When it built its 24,576-GPU clusters, its engineers wrote that "our out-of-box performance for large clusters was initially poor and inconsistent, compared to optimized small cluster performance," with utilization ranging from 10% to 90% until they changed job scheduling, network routing and software⁵. The same post said Meta remained "the largest and primary contributor to PyTorch" and a founding member of the Open Compute Project, where it shares hardware designs such as its Grand Teton GPU platform⁵. The engineers also said "identifying a problematic GPU that is stalling an entire training job becomes very difficult at a large scale," and that they had cut the startup time for large PyTorch jobs "from sometimes hours down to minutes"⁵.

He also said a one-gigawatt cluster called Prometheus, in New Albany, Ohio, would come online in 2026, which would make Meta one of the first tech companies to control an AI data center of that size¹⁴. TechCrunch later noted that one gigawatt "is enough energy to power more than 750,000 homes"¹⁵.

Meta's annual report lists the dependency this creates. Its ability to develop and deploy AI, the company wrote, "is dependent on access to specific third-party equipment and other technical and physical infrastructure, such as processing hardware, network capacity, computing power, and related energy requirements, as to which we cannot control the availability or pricing"²².

Open Source Versus Proprietary Development

The Shift in AI Development Philosophy

Meta built its AI reputation on openly released models. When it launched Llama 2 in 2023, it wrote: "We believe an open approach is the right one for the development of today's AI models, especially those in the generative space where the technology is rapidly advancing"³. Its 2024 annual report said "by making our Llama models openly available, we aim to accelerate AI research and development, improve our own products, and to foster collaboration and innovation within the broader tech community"²².

Zuckerberg's July 2024 letter, "Open Source AI Is the Path Forward," laid out the business case. He compared AI to operating systems, arguing that open source Linux eventually beat closed versions of Unix, and wrote: "I believe that AI will develop in a similar way"⁴. He argued that openness protected Meta from rivals' platforms: "One of my formative experiences has been building our services constrained by what Apple will let us build on their platforms"⁴. And he said releasing models did not hurt Meta's revenue because "selling access to AI models isn't our business model"⁴.

The strategy relied on partners. For Llama 3.1, Zuckerberg said Amazon, Databricks and Nvidia were launching services to help developers fine-tune and distill the models, that Groq had built "low-latency, low-cost inference serving," and that the models would be available on all major clouds, including AWS, Azure, Google and Oracle⁴. By March 18, 2025, Meta said Llama had been downloaded more than one billion times²⁰.

He listed what he said developers wanted: to "train, fine-tune, and distill our own models," to avoid being "locked into a closed vendor," and to protect sensitive data that "can't send to closed models over cloud APIs." He also made a cost claim: "Developers can run inference on Llama 3.1 405B on their own infra at roughly 50% the cost of using closed models like GPT-4o"⁴.

He also addressed the argument that open models help rival nations. "Some people argue that we must close our models to prevent China from gaining access to them, but my view is that this will not work and will only disadvantage the US and its allies," he wrote⁴. "Our adversaries are great at espionage, stealing models that fit on a thumb drive is relatively easy, and most tech companies are far from operating in a way that would make this more difficult"⁴. His proposed answer was for leading American companies to "work closely with our government and allies" while keeping the ecosystem open⁴.

Llama 4 Scout and Maverick were released as open-weight models in April 2025⁹. Even in January 2024, though, Zuckerberg had left room to change course. "For as long as it makes sense and is the safe and responsible thing to do, then I think we will generally want to lean towards open source," he told The Verge. "Obviously, you don't want to be locked into doing something because you said you would"⁷.

Safety and Control Considerations

The July 2025 letter is more cautious than the 2024 one. "We believe the benefits of superintelligence should be shared with the world as broadly as possible," Zuckerberg wrote. "That said, superintelligence will raise novel safety concerns. We'll need to be rigorous about mitigating these risks and careful about what we choose to open source"¹. He did not say which models Meta would keep closed.

Meta had published a more detailed statement in February 2025, and it opened with a defence of openness: "Open sourcing AI is not optional; it is essential for cementing America's position as a leader in technological innovation, economic growth and national security"²⁵. Its Frontier AI Framework, which it said followed a commitment made at the AI Seoul Summit, "focuses on the most critical risks in the areas of cybersecurity threats and risks from chemical and biological weapons." The framework describes identifying catastrophic outcomes to prevent, running threat-modeling exercises with outside experts where needed, and setting "risk thresholds based on the extent to which our models facilitate the threat scenarios"²⁵.

Zuckerberg's 2024 letter had divided AI harms into two groups: "unintentional and intentional." He argued that "open source should be significantly safer" against unintentional harms "since the systems are more transparent and can be widely scrutinized," and described Meta's process as "rigorous testing and red-teaming to assess whether our models are capable of meaningful harm, with the goal of mitigating risks before release"⁴.

Meta's annual report is blunter about what it cannot control. Because it licenses AI technology to third parties, it wrote, "we cannot guarantee that third parties will not use such AI technologies for improper purposes," listing examples from harmful content and intellectual property infringement to "cybersecurity attacks including spear phishing attacks"²².

Talent Acquisition and Research Teams

Meta spent the summer of 2025 hiring. In the June 30 memo, Zuckerberg said Alexandr Wang had joined as Chief AI Officer to lead MSL. "Alex and I have worked together for several years, and I consider him to be the most impressive founder of his generation," he wrote¹³. Former GitHub CEO Nat Friedman joined "to partner with Alex to lead MSL, heading our work on AI products and applied research"¹³.

Scale AI said Wang would remain on its board, that its chief strategy officer, Jason Droege, would become interim CEO, and that Meta would hold a minority of Scale's equity after the investment¹². Scale, which supplies training data and data services to AI labs, said it would use part of the proceeds to pay out its shareholders and vested employees. Wang said in the announcement that Meta's investment "recognizes Scale's accomplishments to date"¹². CNBC reported that Meta hired Wang and some of his colleagues as part of a $14.3 billion investment in Scale AI, and that it also hired Daniel Gross, who had been CEO of Safe Superintelligence¹³.

The memo named new researchers and their previous work. The memo named eleven researchers and described each one's past work. Most had helped build OpenAI or Google models:

Trapit Bansal: "pioneered RL on chain of thought" and co-created OpenAI's o-series models¹³.

Shuchao Bi: co-creator of GPT-4o voice mode and o4-mini, and previously led multimodal post-training at OpenAI¹³.

Huiwen Chang: co-creator of GPT-4o's image generation, who earlier invented the MaskGIT and Muse text-to-image architectures at Google Research¹³.

Ji Lin: helped build o3, o4-mini, GPT-4o, GPT-4.1, GPT-4.5 and the reasoning stack for Operator¹³.

Joel Pobar: worked on inference at Anthropic, after 11 years at Meta¹³.

Jack Rae: "pre-training tech lead for Gemini," who earlier led the Gopher and Chinchilla language model work at DeepMind¹³.

Hongyu Ren: co-creator of GPT-4o, 4o-mini, o1-mini, o3-mini, o3 and o4-mini, who previously led a post-training group at OpenAI¹³.

Johan Schalkwyk: a former Google Fellow, an early contributor to Sesame and technical lead for Maya¹³.

Pei Sun: worked on post-training, coding and reasoning for Gemini at Google DeepMind, after building Waymo's perception models¹³.

Jiahui Yu: co-creator of o3, o4-mini, GPT-4.1 and GPT-4o, who previously led the perception team at OpenAI¹³.

Shengjia Zhao: co-creator of ChatGPT, GPT-4 and OpenAI's mini models, who previously led synthetic data at OpenAI¹³. Zuckerberg wrote that MSL would include "all of our foundations, product, and FAIR teams, as well as a new lab focused on developing the next generation of our models"¹³.

On July 25, Zuckerberg named Zhao chief scientist. "Shengjia co-founded the new lab and has been our lead scientist from day one," he wrote on Threads¹⁵. TechCrunch noted that this gave Meta two chief AI scientists, Zhao and Yann LeCun, who leads FAIR, and that FAIR is designed to focus on long-term research¹⁵.

The pay involved became news in its own right. CNBC reported that OpenAI CEO Sam Altman said on a podcast that Meta had offered signing bonuses as high as $100 million. Meta technology chief Andrew Bosworth told CNBC that "the market is setting a rate here for a level of talent which is really incredible and kind of unprecedented in my 20-year career as a technology executive"¹³. TechCrunch reported that Meta had "reportedly offered some researchers eight- and nine-figure compensation packages" and that Zuckerberg had sent personal emails to researchers¹⁵.

The competition for researchers had been building for years. "We're used to there being pretty intense talent wars," Zuckerberg told The Verge in 2024. "But there are different dynamics here with multiple companies going for the same profile"⁷.

Challenges and Potential Risks

Technical Hurdles in AI Development

Meta's own statements point to open questions. Zuckerberg called the self-improvement "slow for now"¹. In April 2025 the largest Llama 4 model was still in training⁹. In the June memo he said the new lab would "start research on our next generation of models to get to the frontier in the next year or so"¹³.

The company's annual report includes a direct warning to investors: "There are significant risks involved in developing and deploying AI and there can be no assurance that the usage of AI will enhance our products or services or be beneficial to our business, including our efficiency or profitability"²². It adds that its AI initiatives "depend on our access to data to effectively train our models"²².

Some of the lessons Meta has described came from surprises. Zuckerberg told The Verge that the company had doubted coding mattered for its apps, "because it's not like a lot of people are going to ask coding questions in WhatsApp," but found that "coding is actually really important structurally for having the LLMs be able to understand the rigor and hierarchical structure of knowledge"⁷.

Research Methodologies and Innovation Processes

Meta's public descriptions of how it works emphasize building and testing at full scale. "As we push the limits of AI systems, the best way we can test our ability to scale-up our designs is to simply build a system, optimize it, and actually test it (while simulators help, they only go so far)," its infrastructure engineers wrote in 2024⁵.

On the research side, Meta has paired model releases with public benchmarks. When FAIR released V-JEPA 2 in June 2025, it also released "three new benchmarks to evaluate how well existing models can reason about the physical world from video," saying it wanted "to give researchers and developers access to the best models and benchmarks to help accelerate research and progress"¹⁹. With Llama 2, Meta published a transparency section in its research paper that it said "discloses known challenges and issues we've experienced"³. The self-improvement claim in the July 2025 letter did not come with that kind of documentation¹.

Ethical and Safety Concerns

The letter names safety as a concern without describing specific measures¹. The Frontier AI Framework sets out the categories Meta says it tests for, cyber and chemical and biological risks²⁵, and the annual report lists the ways third parties could misuse openly licensed models²².

The infrastructure also has local costs. TechCrunch noted that Prometheus and Hyperion together will use enough energy to power millions of homes, and cited a New York Times report that residents near a Meta data center project in Newton County, Georgia, had seen their taps run dry¹⁴.

Financial Pressure

The spending itself is a risk that Meta discloses. Its annual report says "we expect our AI initiatives will require increased investment in infrastructure and headcount" and that "if our investments are not successful longer-term, our business and financial performance could be harmed"²². In July 2025 Meta said it expected 2026 expense growth to be above the 2025 rate¹⁶.

Industry Impact and Competition

Meta is one of several companies spending on the same scale. TechCrunch described the Hyperion announcement as part of Meta's effort to compete with OpenAI, Google DeepMind and Anthropic, and pointed to other large projects such as OpenAI's Stargate with Oracle and SoftBank and xAI's Colossus supercomputer¹⁴. Zuckerberg has also accused larger rivals of using safety arguments for competitive advantage. "The biggest companies that started off with the biggest leads are also, in a lot of cases, the ones calling the most for saying you need to put in place all these guardrails on how everyone else builds AI," he told The Verge in 2024. "I'm sure some of them are legitimately concerned about safety, but it's a hell of a thing how much it lines up with the strategy"⁷.

The Verge noted in 2024 that OpenAI's stated mission is to create artificial general intelligence and that Demis Hassabis, who leads Google's AI efforts, has the same goal⁷.

The other large platforms reported their own AI plans in late July 2025. On July 23, 2025, Alphabet said it was increasing its 2025 capital expenditures to about $85 billion, and CEO Sundar Pichai said: "We are leading at the frontier of AI and shipping at an incredible pace." Alphabet tied the extra spending to demand for Google Cloud, whose annual revenue run-rate it put at more than $50 billion²⁶. On July 30, Microsoft said Azure had passed $75 billion in annual revenue, and CEO Satya Nadella said "cloud and AI is the driving force of business transformation across every industry and sector"²⁷. On July 31, Amazon CEO Andy Jassy said: "Our conviction that AI will change every customer experience is starting to play out," pointing to Alexa+ reaching millions of customers and a shopping agent "used by many millions of customers"²⁸.

Zuckerberg's case for Meta rests on scale. In the MSL memo he wrote that Meta "is uniquely positioned to deliver superintelligence to the world," because "we have a strong business that supports building out significantly more compute than smaller labs" and "deeper experience building and growing products that reach billions of people"¹³. Meta reported 3.48 billion daily active people across its apps in June 2025¹⁶, and the memo said Meta AI was "used by more than 1 billion monthly actives across our apps"¹³.

Conclusion

Zuckerberg's letter is a statement of direction backed by real spending: a new lab, a wave of senior hires, a minority stake in Scale AI and capital expenditure guidance of up to $72 billion for 2025. It sits on a decade of groundwork, from FAIR and PyTorch to four generations of Llama models.

The claim at its center, that Meta's AI is showing "glimpses" of improving itself, is Meta's own description, and Meta did not publish evidence alongside it. The company has also signaled that its long commitment to open models now comes with conditions. The useful way to follow the story is through what Meta ships, what it releases openly and what it reports to investors next. For more on the terms, see our glossary entries on AI alignment and machine learning .

Frequently Asked Questions

What is Meta superintelligence?

It is Meta's term for the advanced AI it says it is working toward. Zuckerberg's letter does not define it precisely; it describes giving each person a "personal superintelligence" that helps them reach their goals. He set out the idea in a July 30, 2025 letter and organized Meta's AI teams into Meta Superintelligence Labs on June 30, 2025.

What is Meta Superintelligence Labs?

Meta Superintelligence Labs (MSL) is the organization Zuckerberg created on June 30, 2025 to hold all of Meta's AI foundations, product and FAIR teams, plus a new lab working on the next generation of models. It is led by Chief AI Officer Alexandr Wang with Nat Friedman, and Shengjia Zhao is its chief scientist.

Is Meta's AI able to improve itself?

Zuckerberg wrote that Meta had seen "glimpses" of its AI systems improving themselves and that the improvement was "slow for now." Meta has not published technical details, a paper or benchmarks describing this.

How much has Meta invested in its superintelligence project?

Meta has not published a separate budget for superintelligence. It said it expected total 2025 capital expenditures of $66 billion to $72 billion, up from $39.23 billion in 2024, and CNBC reported a $14.3 billion investment in Scale AI.

What are Hyperion and Prometheus?

They are Meta's largest planned AI data centers. Prometheus, in New Albany, Ohio, is a one-gigawatt cluster Meta expected to bring online in 2026. Hyperion, in Louisiana, is planned to reach two gigawatts by 2030 and scale to five gigawatts over several years.

When does Meta expect to achieve superintelligence?

Meta has not given a date. The letter says superintelligence is "now in sight" and that AI will improve existing systems "in the coming years."

Is Meta still releasing open source AI models?

Meta released Llama 4 Scout and Maverick as open-weight models in April 2025. In his July 2025 letter, Zuckerberg wrote that superintelligence will raise novel safety concerns and that Meta must be "careful about what we choose to open source."

How is Meta recruiting talent for its AI projects?

Meta hired Scale AI founder Alexandr Wang as Chief AI Officer, former GitHub CEO Nat Friedman and former Safe Superintelligence CEO Daniel Gross, plus researchers from OpenAI, Google DeepMind and Anthropic. Shengjia Zhao became MSL's chief scientist in July 2025.

What are the main risks of Meta's superintelligence project?

Zuckerberg wrote that superintelligence "will raise novel safety concerns." Meta's annual report warns there is "no assurance" AI will benefit its business and that third parties could misuse its openly licensed models. The data centers behind the effort also draw heavily on local energy and water, as reporting on Meta's projects has noted.