Meta’s Llama Troubles - Sync #521
Plus: how Amazon is using AI to spy on workers; Nvidia is still riding the AI wave high; Gabe Newell's BCI startup emerges from the stealth; Tesla targets June 12 to launch robotaxi service in Austin
Hello and welcome to Sync #521!
This week, we’re taking a closer look at what’s happening inside Meta and Llama, as recent reports suggest the situation is far from ideal.
Elsewhere in AI, Nvidia reported stronger-than-expected earnings, and xAI has paid Telegram $300M to integrate Grok. We also feature an important article exposing the scale of AI systems used at Amazon to spy on workers.
In the world of robotics, Chinese company EngineAI, which is developing humanoid robots, is targeting a $1B valuation. Meanwhile, Tesla aims to launch a robotaxi service in Austin, Texas, on 12 June. Hugging Face has unveiled two new open-source robots (one of which is a humanoid), and Unitree hosted the world’s first fight between two humanoid robots.
This week’s issue of Sync also features Gabe Newell’s brain–computer interface startup, which has emerged from stealth mode, while Neuralink is reportedly raising funds at a $9B valuation. We’ll finish with a look at AI-designed drugs that could hit the market within the next five to six years, and explore the potential of biomanufacturing in space.
Enjoy!
Meta’s Llama Troubles
Of all tech companies, few expected Meta to emerge as the champion of open source AI. After the accidental leak of the first Llama model in 2023, the social media giant fully embraced the open approach to AI with the release of Llama 2. The launch of Llama 3 then cemented Meta’s position as the leader in open AI. Llama 3 represented a massive improvement, placing Meta’s models within reach of the top closed models in 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.
However, that vision has not materialised as planned.
The first major disruption came with the release of DeepSeek R1. Almost overnight, the relatively unknown Chinese AI lab launched an open-source reasoning model that was not only more efficient but also matched or exceeded leading models on multiple benchmarks, all while being trained and operated at a fraction of the cost. R1’s impact was immediate: within days, it shot to the top of community rankings and became one of the most downloaded models on Hugging Face.
DeepSeek R1 sent shockwaves through the AI world—especially inside Meta, where engineers reportedly began to worry about falling behind. Meta, once the leader in open AI, suddenly looked like it was playing catch-up with a new Chinese rival.
With its reputation on the line, Meta moved quickly to respond. In early April, the company launched its much-anticipated Llama 4 models—Scout and Maverick—touting advances such as massive context windows, multimodality, and greater efficiency thanks to a mixture-of-experts architecture. However, the launch swiftly ran into trouble. Users soon noticed that the version of Llama 4 Maverick Meta submitted to the LM Arena benchmarking site was not the same as the one made publicly available. The experimental “chat optimised” version submitted to the leaderboard outperformed the model released to the public, sparking accusations of “gaming” the benchmarks.
To make matters worse, the launch was plagued by technical setbacks. Despite Meta’s claims that Llama 4 Scout could handle a 10-million-token context window, developers found that real-world usage was far more limited, largely due to the extremely high hardware requirements, as well as inconsistent behaviour across environments and frequent crashes. Third-party providers restricted context windows to a small fraction of the advertised maximum, and community testing produced mixed results. While the underlying models remain technically impressive, the rollout felt rushed and unstable, further undermining trust in Meta’s open AI efforts.
And the problems do not stop there. According to The Wall Street Journal, Meta’s flagship model, Llama 4 Behemoth, has been delayed until at least autumn this year due to ongoing internal setbacks. Engineers are reportedly struggling to deliver improvements significant enough for a public release, with senior executives placing the blame on the Llama 4 team for a lack of progress.
One of the clearest signals of Meta’s internal struggles is the exodus of top AI talent. Of the fourteen researchers credited on the 2023 paper introducing the original Llama model, only three still work at Meta. The rest—including several who were instrumental in shaping Meta’s early open source strategy—have moved on. The departure of so many experienced team members, some after more than five years at Meta, has left the company with the difficult task of defending its AI lead without the core group that built it.
In response to these new challenges, Meta has undergone internal restructuring of its AI teams to accelerate product development and regain its competitive edge against fast-moving rivals like OpenAI, Google, and DeepSeek. As reported by Axios, Meta divided its AI operations into two main groups: an AI Products team, focused on delivering user-facing applications and features, and an AGI Foundations unit, responsible for advancing the core technology behind models such as Llama. The Fundamental AI Research group (FAIR) remains a separate entity within Meta.
Artificial intelligence sits at the heart of Meta’s future plans. During a recent annual meeting, Mark Zuckerberg highlighted that Meta AI already reaches one billion users every month, and the company’s goal is to transform it into the world’s top personal AI assistant. Achieving that vision, however, depends on Meta’s ability to produce robust, capable models that can power the next generation of AI-powered products and features. Yet, with mounting internal problems, delivering on these promises might be difficult. Meanwhile, the competition is not slowing down. OpenAI, Google, and Anthropic continue to push the boundaries of what’s possible in artificial intelligence, while Chinese challengers such as Qwen3 from Alibaba and the latest DeepSeek R1 are setting the pace in the open-source space. The next six months will be crucial for Meta as it seeks not only to steady itself but to reclaim its leadership and chart a clear path forward.
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
Valve CEO Gabe Newell’s Neuralink competitor is expecting its first brain chip this year
At GDC 2019, Gabe Newell, the legendary founder of Valve and the saviour of PC gaming, unveiled his interest in brain-computer interfaces (BCIs). Six years later, his BCI startup, Starfish Neuroscience, has emerged from stealth to share its plans. Unlike its competitors, which implant a single device, Starfish aims to implant multiple smaller, less invasive devices accessing different brain regions. The initial product is a custom “electrophysiology” chip for recording and stimulating brain activity, not a full implant yet, scheduled to debut in late 2025. You can read more about Starfish’s approach on their blog.
Elon Musk’s Neuralink raises fresh cash at $9B valuation
Semafor reports that Neuralink has raised $600 million, bringing its valuation to $9 billion, according to anonymous sources. Elon Musk’s BCI startup was last valued at $3.5 billion in late 2023.
Surgeons in California perform first ever successful bladder transplant
Surgeons in California have performed the world’s first successful bladder transplant on a 41-year-old patient, who previously lost his bladder and kidneys to cancer and kidney disease. The procedure enabled the patient to come off dialysis and marks a major milestone in medicine, though long-term outcomes remain uncertain.
🧠 Artificial Intelligence
Busting Unions with AI: How Amazon Uses AI to Crush Labor Movements
In this article,
Nvidia beats on earnings and revenue as data center sales jump 73%
Nvidia reported stronger-than-expected earnings, with revenue rising 69% to $44.06 billion. This growth was driven by a 73% annual increase in its data centre business, fuelled by global demand for AI chips. Despite facing significant headwinds from US export restrictions to China, which cost Nvidia billions in lost sales and inventory charges, the company’s net income jumped 26% to $18.8 billion, and shares rose 6% in extended trading. Jensen Huang said the AI chip market in China is now "effectively closed" to US firms, but Nvidia continues to see robust growth across its data centre, gaming, automotive, and visualisation divisions.
Duolingo CEO walks back AI-first comments: ‘I do not see AI as replacing what our employees do’
Duolingo CEO Luis von Ahn has walked back earlier comments that AI would replace contract workers, now stating the company will continue hiring and support employees in adapting to AI as a tool rather than a replacement. This shift follows user criticism and reflects a wider trend, with companies like Klarna also tempering their “AI-first” promises amid evidence that AI’s productivity benefits remain limited outside specific roles.
Nvidia Pushes Further Into Cloud With GPU Marketplace
Nvidia is launching DGX Cloud Lepton, a new service that broadens access to its sought-after AI chips by connecting developers directly with a range of cloud providers, not just major tech giants. The platform serves as a marketplace where AI developers can select from various GPU cloud vendors, including smaller partners such as CoreWeave and Lambda, to train and deploy their models and reduce friction in accessing AI computing resources.
Oracle to buy $40 billion of Nvidia chips for OpenAI's US data center, FT reports
Oracle will spend around $40 billion on Nvidia’s high-performance chips to power a new OpenAI data centre in Abilene, Texas, as part of the Stargate Project, according to the Financial Times. The centre is expected to be fully operational by mid-next year and will use approximately 400,000 Nvidia GB200 chips, which Oracle will lease to OpenAI.
Elon Musk Tried to Block Sam Altman’s Big AI Deal in the Middle East
The Wall Street Journal reports that OpenAI and other American tech giants secured a major deal to build one of the world’s largest AI data centres in Abu Dhabi, despite efforts by Elon Musk to block the agreement unless his own company, xAI, was included. Musk warned UAE officials that the deal would not get Trump’s approval without xAI’s involvement. Despite Musk’s warnings and lobbying during President Trump’s Middle East trip, the deal moved forward with xAI listed as a potential future participant.
xAI to pay Telegram $300M to integrate Grok into the chat app
Telegram’s CEO, Pavel Durov, has announced on X a partnership with xAI to bring the Grok chatbot to all Telegram users and integrate it into apps on the platform for one year, in a deal worth $300 million in cash and equity. Telegram will also receive 50% of revenue from xAI subscriptions purchased through the app. Grok will be accessible from the chat search bar, offering features such as writing suggestions, chat summaries, sticker creation, and moderation support.
Nvidia ‘considering’ new AI chip for China but rules out Hopper series due to US export ban
Nvidia has confirmed that it does not yet have a new AI chip ready for the Chinese market, despite considering such plans amid ongoing US export controls that have ended its Hopper data centre business in China. While speculation has grown about a lower-cost AI chipset specifically for China, Jensen Huang clarified that nothing is ready to announce, though analysts see potential for a new product later this year.
Nick Clegg says asking artists for use permission would ‘kill’ the AI industry
As the UK Parliament debates new legislation to increase transparency over copyrighted content used by tech firms, Nick Clegg, former UK Deputy Prime Minister and Meta executive, warned that requiring artists’ consent before using their work to train AI models would be unworkable and could "kill" the UK AI industry. While major artists back an amendment demanding disclosure of training data, MPs recently rejected it, though campaigners insist the battle for greater AI transparency is not over as the bill heads back to the House of Lords in June.
Inside the First Stargate AI Data Center
In this article, Bloomberg takes us to Abilene, Texas, where the first of many massive data centres is rising as part of the ambitious $500 billion Stargate Project—a collaboration between OpenAI, Oracle, and SoftBank, with backing from President Donald Trump. The project aims to rapidly build a nationwide network of AI infrastructure, beginning with a $12 billion site in Abilene.
LM Arena, the organization behind popular AI leaderboards, lands $100M
LM Arena, a popular crowdsourced AI benchmarking project used by major AI labs, has raised $100 million in a seed funding round led by Andreessen Horowitz and UC Investments, valuing the organisation at $600 million. Founded in 2023 and primarily run by UC Berkeley-affiliated researchers, LM Arena has rapidly become central to the AI industry, partnering with firms such as OpenAI, Google, and Anthropic. Recently, LM Arena has faced accusations of enabling labs to game its leaderboards, which the company denies.
DeepSeek-R1-0528
DeepSeek has released an updated version of its R1 model, which significantly improves its depth of reasoning and inference capabilities, resulting in the model achieving comparable results on some benchmarks to those of OpenAI’s o3 or Google’s Gemini 2.5 Pro.
Claude 4 tops ARC-AGI Leaderboard
The recently released Claude Sonnet 4 and Claude Opus 4 models from Anthropic have been added to the ARC-AGI leaderboard and swiftly claimed the top spots. Claude Opus 4 (Thinking 16K) leads with a score of 8.6% on the ARC-AGI-2 benchmark, followed by Claude Sonnet 4 (Thinking 16K) with 5.9% in second place, and Claude Opus 4 (Thinking 8K) with 4.5% in third. Meanwhile, Claude Sonnet (Thinking 8K) achieved a score of 2.1%, ranking twelfth on the list.
▶️ The AI Revolution Is Underhyped | Eric Schmidt (25:37)
In this interview, Eric Schmidt, former CEO and executive chairman of Google, argues that AI is still underhyped given its growing capabilities, especially in planning and strategy. He also highlights the challenges around the immense energy and data demands of advanced AI systems, as well as ethical dilemmas, including the dual-use nature of AI for civilian and military applications, the risks of open-source models, and the need for robust guardrails rather than development bans. He remains cautiously optimistic about AI’s potential to transform society, stressing the importance of individual freedom, societal values, and embracing new technology.
Fortnite’s AI-Generated Darth Vader Controversy, Explained
Fortnite has introduced a version of Darth Vader whose voice lines are dynamically generated by AI in response to players’ actions. However, Epic Games did not adequately safeguard the AI, allowing players to break character and make Vader say offensive things in the iconic style of James Earl Jones. The use of an AI-generated voice, created with the help of Jones’ estate, has also reignited ethical debates about digital recreations and the impact of AI on voice acting jobs, prompting SAG-AFTRA to file a legal complaint over the lack of negotiation with human performers.
🤖 Robotics
Tesla Targets June 12 Launch of Robotaxi Service in Austin
Tesla is set to launch its long-anticipated robotaxi service in Austin on 12 June, according to a source familiar with the matter, as reported by Bloomberg. The launch follows recent public road tests in the city with no driver in the seat, although a Tesla engineer was present as a passenger. While the rollout date could still change, the initial fleet will consist of around 10 self-driving vehicles, with plans for rapid expansion. Texas’s relatively relaxed rules on autonomous driving have facilitated Tesla’s progress, although the company is not yet listed as a licensed rideshare provider in the state.
EngineAI Targets $1 Billion Valuation as Chinese Robots Heat Up
EngineAI, a Shenzhen-based robotics start-up founded in 2023, is seeking to raise $139 million at a $1 billion valuation later this year. The company plans to use these funds to accelerate the development of its humanoid robots and projects to ship 1,000 units in 2025, ramping up to 10,000 units by 2028, with a target for future entry into the household market. Supported by China’s comprehensive supply chain and a national push for robotics innovation led by President Xi Jinping, EngineAI and peers such as Unitree are racing to outpace global competitors like Boston Dynamics in bringing advanced, human-like robots to the mainstream.
▶️ Unitree Combat Competition Highlights (0:53)
On 25 May 2025, Unitree hosted the world’s first fight between two humanoid robots. Two Unitree G1 robots, equipped with boxing gloves and head protection gear, entered the ring to face off in this historic match. I’ve linked above to the highlights reel. The full livestream, which is over 1.5 hours long, is available here.
Hugging Face unveils two new humanoid robots
As part of its growing move into robotics, Hugging Face has announced two new open-source humanoid robots, HopeJR and Reachy Mini. HopeJR is a full-sized, highly mobile robot expected to cost around $3,000, while Reachy Mini is a smaller desktop unit priced between $250 and $300. Both are designed to be affordable and open source, allowing anyone to build and modify them. The first units are expected to ship by the end of the year. I’ve had a chance to see these robots in person at the Humanoid Summit in London (more on that soon)—they look very promising, and I’m happy to see initiatives like these making robotics more accessible.
▶️ Perception and Adaptability | Inside the Lab with Atlas (7:58)
In this video, engineers from Boston Dynamics give us a glimpse into how Atlas sees the world and how it adapts to the unpredictable environments it might be operating in. As they explain, something that seems easy for us humans, like picking up random objects, can be extremely hard for robots, requiring a lot of work to make it happen smoothly and naturally.
Robots Are Starting to Make Decisions in the Operating Room
Could your future surgeon be a robot? Advances in autonomous surgical robots like the Smart Tissue Autonomous Robot (STAR) suggest this could soon be a reality. These robots are now able to perform complex surgical tasks with exceptional precision and minimal human input, potentially reducing complications and helping to address the shortage of skilled surgeons. While technical, ethical, and regulatory challenges remain, the integration of AI and advanced imaging is moving robotic surgery closer to becoming a new standard in operating rooms, promising safer and more consistent outcomes for patients.
The small robot company with big global ambitions
Meet Robot.com, the Colombian robotics company leading Latin America’s push into the global robotics market. Despite facing high production costs, limited investment, and minimal government support at home, Robot.com has deployed more than 500 robots internationally and formed major partnerships, including with Amazon Web Services. Like many Latin American robotics startups, it’s recognised for engineering expertise but often looks abroad for growth due to scarce local funding.
🧬 Biotechnology
AI-Generated Drugs to Reach Market by 2030, Says Insilico CEO
AI-designed drugs could be commercially available within five to six years, according to Alex Zhavoronkov, CEO of Insilico Medicine, who says his company is leading the way with over 40 AI-driven drug programs. While no AI-developed drug has been approved yet, industry interest is rising, with companies like Takeda nearing late-stage trials. Insilico recently filed for a Hong Kong stock exchange listing after being valued at over $1 billion.
Boosters and Biologics: Is Space-Based Biomanufacturing Real?
This article explores the fascinating possibility of manufacturing drugs in Earth’s orbit using constellations of steel bioreactors. While the concept has shown promise through experiments on the International Space Station, commercial space-based biomanufacturing faces steep challenges, including high costs, regulatory uncertainty, and scalability issues. Nevertheless, declining launch costs and the high value of pharmaceuticals could soon make industrial drug production in orbit a reality, though mainstream adoption is likely still at least a decade away.
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!"
Thank you for sharing my work. Love your stuff always