Meta’s $65 Billion Gamble on Superintelligence - Sync #526
Plus: OpenAI and Oracle struck $30B deal; Apple considers OpenAI and Anthropic for new Siri; Amazon deployed its millionth robot; a BCI to treat severe depression; and more!
Hello and welcome to Sync #526!
This week, Meta made waves in the AI industry by opening its wallet and poaching top researchers from OpenAI and other labs to join its new Superintelligence Labs. We’ll take a closer look at what happened as the main story in this week’s issue.
Elsewhere in AI, OpenAI has struck a deal to rent about 4.5 gigawatts of data centre power from Oracle, Apple is considering either OpenAI or Anthropic to power the new Siri, and Gartner predicts that 40% of agentic AI projects will fail by 2027.
Over in robotics, Amazon has deployed its millionth robot. Meanwhile, Tesla Optimus is reportedly in shambles, while China hosted a football tournament for humanoid robots.
This week’s issue of Sync also features neural implants to treat severe depression, an open-source exoskeleton, an exploration of why nuclear non-proliferation is the wrong framework for AI governance, a project to synthesise the human genome, and more!
Enjoy!
Meta’s $65 Billion Gamble on Superintelligence
What does it take to win the AI race? For Meta, the answer is simple: billions in new data centres, and an aggressive campaign to recruit the world’s top AI talent.
Since 2024, the world’s largest tech companies—OpenAI, Microsoft, Google, and others—have poured billions into AI infrastructure. The competition for high-end chips and cloud contracts is fierce, and the scramble for data is just as intense. After all, the AI scaling laws say that the bigger the model and the more compute power and data you throw at it, the better it performs.
But even the best hardware and largest datasets are not enough on their own. While competition for compute and data has dominated headlines, the battle for top AI researchers has mostly played out behind the scenes—until now. In a bid to win the AI race, Mark Zuckerberg has brought the fight for talent front and centre.
Meta’s Position in AI: Not Looking So Hot
Before we explore Meta’s new AI ambitions and Zuckerberg’s hiring spree, it is worth examining where the company currently stands in the AI landscape.
Meta is best known in the AI community for being an unlikely champion of an open approach to AI with its Llama family of open models. While OpenAI, Google and Anthropic keep their models closed, Meta released Llama for free for almost everyone to download and use as they wish. Llama quickly became the go-to model for researchers and people experimenting with large language models. With each iteration, Meta has steadily improved its open models, and by the release of Llama 3, its performance was approaching that of the top closed models on many benchmarks. “Llama 3 is competitive with the most advanced models and leading in some areas. Starting next year, we expect future Llama models to become the most advanced in the industry,” wrote Mark Zuckerberg in his Open Source AI is the Path Forward post last year.
Those plans, however, did not materialise. Llama 4 failed to meet expectations. The launch was plagued by technical setbacks, and the company was accused of manipulating benchmarks. Moreover, the release of DeepSeek R1 earlier this year greatly undermined Meta’s position as a leader in open models. And if that wasn’t enough, Meta’s flagship model, Llama 4 Behemoth, has been delayed until at least autumn due to ongoing internal issues, according to The Wall Street Journal.
Meta has also suffered a significant talent drain. Of the fourteen researchers credited on the 2023 Llama paper, only three remain at Meta. The rest—including several who were instrumental in shaping Meta’s early open-source strategy—have moved on.
Zuckerberg’s Big Bet on Superintelligence
Just as the internet created Google and Amazon, social media gave rise to Facebook, and mobile revived Apple, the world’s biggest tech companies now compete to win the AI race. Mark Zuckerberg, whose empire was built on winning the social media race and capitalising on the mobile revolution, is determined not to lose the AI race. His goal: to build Artificial General Intelligence (AGI)—AI that can reason, learn, and perform at or above human level across multiple domains.
Meta already has access to heaps of data from its social media apps to train new AI models. But data alone is not enough—they also need compute and talent. To catch up with Google and OpenAI in compute power, Meta is making its largest-ever investment: raising $29 billion in private financing for a new network of AI-optimised data centres across the US, as part of a planned $65 billion spending on AI infrastructure, compute, and research.
Meta also knows that talent is just as important as hardware. That’s why Zuckerberg is taking extraordinary steps to assemble what he hopes will be the most formidable team of AI researchers and engineers in the world.
The Battle for AI Talent
The creation of Meta Superintelligence Labs (MSL) is Zuckerberg’s answer to the talent war and his bid to place Meta at the very forefront of AI research. MSL consolidates all of Meta’s major AI projects—including Llama language models, augmented reality, smart glasses, and next-generation agents—under a single, unified leadership. With MSL in place, Zuckerberg set out to hire the best minds in the field.
The world of top AI researchers is a relatively small community. Many have worked at multiple labs, including OpenAI, Google DeepMind, and Anthropic, and regularly (and mostly quietly) move between companies to pursue new challenges, seniority, or higher pay.
Acquiring talent through “acqui-hires”—buying smaller companies primarily for their staff—has also become a common practice among tech giants. Notable examples include Google, Microsoft and Amazon, which acqui-hired, or plundered for talent, CharacterAI, Inflection and Adept, respectively (more on that here).
However, what Mark Zuckerberg has done is crank all those tactics to 11 by throwing mindboggling amounts of money to lure top AI talent to join Meta’s new AI lab. Zuckerberg has made building an elite AI team his personal priority and has been actively involved in recruiting, reportedly maintaining a “dream team” list of the world’s top AI researchers and hosting them at his homes. As a result, Meta has been able to recruit a remarkable roster for Meta Superintelligence Labs.
Leading the new lab is Alexandr Wang, the founder and former CEO of Scale AI, brought in through a $14 billion investment. Meta also sought to acquire Safe Superintelligence (SSI), a startup co-founded by Ilya Sutskever, but was ultimately unsuccessful (Sutskever, however, said SSI is “flattered by their attention” but is focused on seeing its work through). Nonetheless, Meta succeeded in hiring SSI’s CEO, Daniel Gross, along with Nat Friedman, the former CEO of GitHub. At the time of writing, Meta has poached at least fourteen senior researchers and engineers from OpenAI, DeepMind, Anthropic, and Google, offering compensation packages rumoured to be in millions of dollars. Many of Meta's new hires played pivotal roles in developing world-class models like GPT-4o, Gemini, o1, and o3. Some reportedly received compensation packages in the tens of millions.
The Fallout of Meta’s Poaching Spree
Meta’s aggressive recruitment strategy has left a mark on the wider AI industry. Some companies have lost CEOs, and others have lost “talented and hardcore” colleagues. The departures have had an emotional impact. Wired obtained an internal memo from OpenAI’s chief research officer, Mark Chen, who wrote: “I feel a visceral feeling right now, as if someone has broken into our home and stolen something.”
The situation has also highlighted the extraordinary pressure and burnout in the field, as companies push for a relentless pace to stay ahead. In many labs, the mission to build AGI has created a high-stress environment where eighty-hour weeks are not uncommon, and the pursuit of artificial intelligence can border on the cult-like (“Feel the AGI,” anyone?). For some researchers, the lure of a higher salary and a slightly less frantic work environment may have been enough reason to jump ship.
In response to losing people to Meta, OpenAI was reportedly shut down for a week to give its employees some time to rest, alongside boosting compensation and pausing internal deadlines, to stave off burnout and further defections. Other AI companies may, at least for a brief time, also reconsider how much they push their employees.
Will the Bet Pay Off?
Meta can quickly catch up in terms of computing power, and it already sits on heaps of data from Facebook, WhatsApp and Instagram. With the creation of a superteam at Meta Superintelligence Labs, it seems Meta now has all the ingredients for a breakthrough in AI—and is betting its future on it, just as it once did with VR.
However, superteams tend to be fragile and unstable. There are plenty of examples in sports where superteams, despite their star power, fail to deliver on the hype. A team built on massive pay packages and intense competition could also prove unstable, especially if momentum stalls, the mission becomes unclear, or egos and internal politics take over.
It is too early to predict whether Zuckerberg’s gamble on building a superteam at Meta Superintelligence Labs will allow Meta to leapfrog the competition, or if it will ultimately fizzle out like the company’s ambitions to make VR the next big thing. The answer may come when Llama 5—or whatever Meta’s next flagship model will be—arrives, or if the team implodes before then.
What do you think? Does Zuckerberg have the right ingredients to win the AI race, or will this huge gamble end up like the Metaverse?
If you enjoy this post, please click the ❤️ button or share it.
Do you like my work? Consider becoming a paying subscriber to support it
For those who prefer to make a one-off donation, you can 'buy me a coffee' via Ko-fi. Every coffee bought is a generous support towards the work put into this newsletter.
Your support, in any form, is deeply appreciated and goes a long way in keeping this newsletter alive and thriving.
🦾 More than a human
Next-Gen Brain Implants Offer New Hope for Depression
In search of a treatment for severe depression, scientists are exploring advanced deep brain stimulation (DBS) technology, which uses implanted electrodes to monitor and adjust brain activity in real-time. Unlike traditional “set-it-and-forget-it” DBS, new AI-driven systems can detect early signs of relapse and automatically adapt stimulation, offering more personalised and effective treatment. Although DBS is not yet widely approved for depression, ongoing trials and innovations are moving the field towards responsive, precise therapies that could help those who have not found relief from existing options.
Gene therapy restored hearing in deaf patients
Researchers from Sweden and China have shown that gene therapy can significantly improve hearing in both children and adults with congenital deafness. The study, which involved ten patients aged 1 to 24, used a synthetic virus to deliver a healthy gene to the inner ear, resulting in notable hearing gains within a month and substantial improvement at six months, particularly among younger children. The therapy was safe and well-tolerated, with no serious side effects reported, marking a major advance in genetic treatments for deafness.
OpenExo
Meet OpenExo, an open-source exoskeleton platform designed for mobility and rehabilitation research. Developed by a team at Northern Arizona University, OpenExo offers freely available mechanical designs, electronics, and software for researchers seeking to advance wearable robotics, making it an accessible resource for the global rehabilitation and assistive technology community.
🧠 Artificial Intelligence
Oracle, OpenAI Expand Stargate Deal for More US Data Centers
Bloomberg reports that OpenAI has struck a deal to rent about 4.5 gigawatts of data centre power from Oracle in the US as part of its $500 billion Stargate AI infrastructure project. Earlier this week, Oracle announced a single cloud deal worth $30 billion in annual revenue beginning in fiscal 2028, without naming the customer. According to the report, the Stargate agreement with OpenAI makes up at least part of this contract.
Apple Weighs Using Anthropic or OpenAI to Power Siri in Major Reversal
Apple Intelligence is not going well, to say the least, as the company struggles to keep pace with competitors in generative AI. As Mark Gurman at Bloomberg reports, Apple is reportedly considering replacing its in-house AI models with technology from Anthropic or OpenAI to power a new version of Siri. The company has held talks with both firms about running custom versions of their models on Apple’s own cloud servers, though no final decision has been made.
xAI raises $10B in debt and equity
Morgan Stanley announced that Elon Musk’s AI company, xAI, has raised $10 billion in a combination of debt and equity. The funds are intended to support the development of advanced AI solutions, including a major data centre and its Grok platform. This latest injection, which follows a $6 billion fundraising round in December, brings xAI’s total capital raised to approximately $17 billion.
What comes next for AI copyright lawsuits?
Recent court rulings in copyright cases in favour of Anthropic and Meta (which were featured in last week’s issue of Sync) have changed the game. However, although tech companies scored a major win, the courts’ rulings were nuanced and limited, with judges finding either that the use of copyrighted works was transformative or that the plaintiffs had not proved market harm—leaving many legal issues unresolved as dozens of similar lawsuits continue and questions over the use of pirated materials remain.
At Amazon’s Biggest Data Center, Everything Is Supersized for A.I.
The New York Times takes us to New Carlisle, Indiana, where Amazon is building a massive data centre complex designed with a single customer in mind: Anthropic. This project, part of Amazon’s “Project Rainier”, will eventually feature around 30 data centres packed with hundreds of thousands of AI chips and require 2.2 gigawatts of electricity—enough to power a million homes—while drawing millions of gallons of water for cooling each year. The scale of the operation, already drawing both local support and protest, highlights the extraordinary energy and resources being devoted to the AI boom, as tech giants race to construct ever-larger computing facilities across the US.
Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027
Agentic AI is the current hot thing. However, Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls. Additionally, Gartner argues that many agentic AI initiatives are immature, with only a small number able to demonstrate meaningful return on investment. The firm advises organisations to cut through the hype and make strategic decisions about where to implement agentic AI, emphasising the importance of focusing on genuine productivity gains and aligning projects with clear business outcomes.
The AI Backlash Keeps Growing Stronger
This article explores the growing public backlash against the use of AI in any shape or form, highlighting widespread anger over automation replacing human workers and increasing distrust towards AI’s role in society, the environmental impact of AI, mental health effects, and the exploitation of creative work. The article predicts that resistance to AI will only continue to grow, and that real-world protests against AI may soon match the intensity of online backlash.
Nuclear Non-Proliferation Is the Wrong Framework for AI Governance
This article argues that using nuclear non-proliferation as a model for AI governance is fundamentally flawed, as AI differs from nuclear technology in its broad applicability, lack of excludable resources, and continuous rather than binary strategic value. While prominent figures such as Demis Hassabis and Sam Altman have called for an “IAEA for AI,” the article contends that such analogies misrepresent the challenges of controlling AI’s spread and divert attention from more effective regulatory approaches. Instead, it recommends adapting existing standards and regulatory bodies to oversee AI, focusing on clear, practical measures tailored to AI’s unique characteristics.
▶️ The LLM's RL Revelation We Didn't See Coming (15:33)
This video argues that recent research has undermined the belief that reinforcement learning (RL), particularly reinforcement learning from verifiable rewards (RLVR), enables large language models to develop new reasoning abilities; instead, RL mostly amplifies knowledge already present in the base models, with apparent advances often failing to generalise across different AI systems. The video concludes that true innovation in language models still relies on large-scale pre-training and that distillation, not RL, appears more promising for introducing genuinely new capabilities.
🤖 Robotics
Amazon launches a new AI foundation model to power its robotic fleet and deploys its 1 millionth robot
Amazon reached a historic milestone by deploying its one-millionth robot, delivered to a fulfilment centre in Japan. Alongside this achievement, Amazon unveiled a new generative AI model, DeepFleet, which will boost robot fleet travel efficiency by 10%.
Tesla Optimus is in shambles as head of program exits, production delayed
Elon Musk set a goal for Tesla to produce 5,000 to 10,000 Optimus humanoid robots this year. However, this goal is now in jeopardy, as recent reports suggest production has been delayed. One reason for this is the recent departure of the programme’s leader, Milan Kovac. Other reasons include a major redesign, hardware challenges such as overheating and low load capacity, and a two-month pause in supplier orders. Nevertheless, Tesla is expected to push forward and unveil an updated Optimus at its shareholders’ meeting.
Tesla says it made its first driverless delivery of a new car to a customer
Tesla has announced its first driverless delivery, with a brand new Model Y from its Austin Gigafactory driving “fully autonomously” to a local customer, according to Elon Musk. The company has not disclosed which technology was used or when it will be available to the public, and the demonstration comes amid ongoing regulatory scrutiny and competition from established robotaxi services like Waymo. This news arrives as Tesla faces brand backlash and declining sales in key markets.
China hosts first fully autonomous AI robot football match
Recently, there was a football (soccer if you are from the US) tournament in Beijing in which not humans but humanoid robots competed. The event, intended as a testing ground for advancing humanoid robots, saw university teams bringing their own robots, with Tsinghua University’s THU Robotics defeating the China Agricultural University’s Mountain Sea team 5–3 in the final. Despite the comical falls, experts noted the impressive progress in robot development, though human players remain far ahead for now.
H2L, a Japanese company specialising in advanced human-robot interaction technology, presents the Capsule Interface—an interesting device that enables full-body, real-time synchronisation with robots and virtual avatars. According to the company, you simply sit or lie down and connect to a robotic body, while the device detects and transmits your muscle movements and intentions to control the robot or avatar. Potential use cases include attending business meetings without the need to travel, remotely operating robots in hazardous environments, or simply experiencing real or virtual worlds through hardware or software avatars.
🧬 Biotechnology
▶️ A Billion Years of Evolution in a Single Afternoon — George Church (1:34:28)
Dwarkesh Patel sat down with George Church, a legendary synthetic biologist, to discuss the future of bioengineering. In this conversation, Church discusses rapid advancements in biotechnology, including CRISPR, de-ageing, and de-extinction. He predicts significant lifespan increases by 2050, driven by exponential progress in biotechnologies and breakthroughs in reversing ageing. He also touches on the potential for biobots and the dual-use nature of synthetic biology, emphasizing the need for robust biodefense strategies. The interview highlights the potential for AI to accelerate biological research and the importance of genetic counselling alongside gene therapy. It is a must-watch if you are interested in biotech and bioengineering.
NATO fund backs biotech startup in push to counter biological threats
The NATO Innovation Fund has made its first investment in a biotechnology firm, co-leading a $35 million funding round for UK-based Portal Biotech, which specialises in using AI-driven protein sequencing to rapidly detect and counter biological threats. This move, part of NATO’s wider $1 billion investment strategy to strengthen defence technologies, aims to boost biosecurity and enable on-site detection of pathogens, potentially improving pandemic prevention, drug discovery, and precision medicine.
UK scientists to synthesise human genome to learn more about how DNA works
Researchers in the UK have launched the Synthetic Human Genome (SynHG) project with the goal of building large sections of human genetic material in the lab over the next five years to better understand DNA and develop new medical therapies. By creating and testing synthetic chromosomes in human cells, scientists hope to unlock treatments for diseases, such as developing virus-resistant cells for transplantation. The project, involving several leading universities, will also explore the ethical and social issues of synthetic genomes, including concerns about environmental risks and the potential for designer babies, though experts say such possibilities remain distant.
Arc Institute Launches Virtual Cell Challenge to Accelerate AI Model Development
The Arc Institute has launched the Virtual Cell Challenge, an open competition aimed at accelerating the development of AI models that can predict gene expression changes in cells under various conditions, crucial for advancing therapeutic discovery. Sponsored by Nvidia, 10x Genomics, and Ultima Genomics, the competition offers a $100,000 grand prize and tasks participants with building models that generalise to new cell contexts using a newly generated single-cell dataset. The initiative aims to set industry benchmarks—similar to the CASP competition in protein structure prediction—to foster collaboration and innovation in virtual cell modelling, with the ultimate goal of improving drug development and treatment for complex diseases.
The Failures and Futures of Cancer Vaccines
This article explores the challenges and recent advances in the field of cancer vaccines, highlighting why effective personalised therapies remain elusive despite decades of research and some early clinical successes. It outlines key scientific hurdles as well as the high cost and complexity of manufacturing personalised vaccines. The article also discusses emerging technologies and approaches that could help overcome these barriers and transform cancer vaccines into a more widely accessible and effective treatment option in the future.
Thanks for reading. If you enjoyed this post, please click the ❤️ button or share it.
Humanity Redefined sheds light on the bleeding edge of technology and how advancements in AI, robotics, and biotech can usher in abundance, expand humanity's horizons, and redefine what it means to be human.
A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!
My DMs are open to all subscribers. Feel free to drop me a message, share feedback, or just say "hi!"