World Humanoid Robot Games - Sync #533
Plus: OpenAI eyes $500B valuation; Meta restructures its AI teams; the dangers of vibe medicine; US government takes 10% stake in Intel; a holographic policeman in South Korea; and more!
Hello and welcome to Sync #533!
In this week’s issue of Sync, we take a look at the first World Humanoid Robot Games in Beijing, where humanoid robots—and the teams behind them—competed in various events, showcasing not only how far humanoid robotics has come but also their current limitations.
Over in AI, OpenAI is eyeing a $500 billion valuation and has launched a sub-$5 ChatGPT plan in India. Anthropic now allows Claude to end conversations and is in talks to raise $10 billion, while Meta has paused its AI talent hiring campaign, reorganised its AI teams, and partnered with Midjourney.
In robotics, Boston Dynamics has demonstrated Atlas’s impressive ability to manipulate objects, even when humans try to make it more difficult. Meanwhile, Waymo has secured a permit to begin testing autonomous vehicles in New York City.
Also in this issue, we feature a warning against using ChatGPT for medical advice, a look at a holographic policeman in South Korea, and news that the US government has taken a 10% stake in Intel.
Enjoy!
World Humanoid Robot Games
The inaugural World Humanoid Robot Games showed how far humanoid robots have come—and how much they still have to go
In mid-August 2025, at the National Speed Skating Oval in Beijing, built for the 2022 Winter Olympics, 280 teams from 16 countries gathered to see who could conquer a variety of athletic challenges. But it wasn’t humans competing—it was more than 500 humanoid robots taking part in the inaugural World Humanoid Robot Games.
It was a remarkable showcase of how far humanoid robotics has come. A decade ago, the best robots in the DARPA Robotics Challenge were slow, clumsy, and often unable to complete the simplest tasks. Today’s machines are faster, smarter, and more reliable—and in Beijing, they showed just how much progress has been made.
While Chinese teams dominated many events, the Games drew teams from all over the world. Teams from Germany, Portugal, the Netherlands, Brazil, Malaysia, and beyond brought diverse approaches and shared a common goal: advancing the field of humanoid robotics.
Glory, Glitches, and Giggles
Across three days of competition, robots went head-to-head in events including 100m and 400m sprints, obstacle course, football (or soccer, if you are in the US), kickboxing, martial arts, and even practical tasks, such as sorting medicine and cleaning. There was also a cultural twist, with some robots performing dances, playing drums, and showing off choreographed kung fu routines.
On the track, Chinese firm Unitree Robotics proved dominant, taking gold in the 1500m, 400m, 4x100m relay, and 100m hurdles. In total, Unitree scored 7 gold, 3 silver and 4 bronze medals. In the 400m race, Unitree H1 clocked 1:28.03, which is an impressive result, but still far behind Wayde van Niekerk’s 45.03-second world record. The same can be said about the 100m sprint and every other competition the robots took part in—they showed remarkable performance, but the gap to human performance is still noticeable.
Of course, the Games delivered plenty of slapstick moments, too. One robot ran into a person off track during a sprint, knocking him down and then continued the race like nothing happened. Another robot had to withdraw from the 1500m after its head flew off midway through. In the boxing ring, robots threw more punches at thin air than at each other, much to the crowd’s delight.
Football matches were equally unpredictable. Robots frequently stumbled, tripped over the ball, or crashed into one another. Still, Tsinghua University’s Hephaestus team secured the 5v5 football title with a rare long-range goal against Germany’s HTWK Robotics+Nao Devils. In the 3v3 event, China Agricultural University’s Mountain & Sea team took home the championship.
Building the “Third Olympics”
Today’s humanoids can walk, run, and even perform complex acrobatic moves, but they remain limited in reasoning, dexterity, and safe interaction with humans. Folding laundry, navigating a cluttered kitchen, or safely assisting an elderly person are all tasks far more complex than sprinting 100 metres or sparring in the ring.
For researchers and companies, the Games were more than just a spectacle. They offered a chance to push the robots and the teams behind them to the limits, showing what they are capable of, but also what needs to be improved. The Games were not only an opportunity to compete but also to collaborate. Teams routinely shared insights and even codebases after competitions, all united in the shared pursuit of advancing humanoid robotics.
Organisers envision the World Humanoid Robot Games becoming a recurring international event, a “third Olympics” that combines sport, science, and industry. For now, the Games are part research platform, part cultural showcase, and part entertainment.
For China, the Games were a chance to stake a claim to leadership in humanoid robotics. The field has been pushed to the forefront of the country’s national AI agenda, supported by billions in government funding and incentives. By hosting a global event, China showcased its rapid progress in “embodied AI” while signalling its ambition to lead the industry.
The Games were also a chance to showcase the best in humanoid robotics to the public and inspire a new generation of engineers. In its article about the Games, The Guardian shared a story of a mother who brought her two daughters to watch the competitions. Whether they become engineers or not, it’s easy to imagine that someone in the stands will look back on this event as the moment they decided to pursue a future in robotics.
The next World Humanoid Robot Games are set for August 2026, also in Beijing. If the past year’s progress is any indication, the robots we see next summer will be faster, smarter, more agile, and maybe even more reliable under pressure.
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.
🧠 Artificial Intelligence
Meta Freezes AI Hiring After Blockbuster Spending Spree
Meta has paused hiring in its AI division after a major recruitment campaign that brought in over 50 researchers and engineers from companies like OpenAI and Google, The Wall Street Journal says. The freeze also limits internal transfers as Meta focuses on building a clearer structure and controlling costs. The pause follows investor worries that Meta’s high spending on AI talent could hurt shareholder returns if it doesn’t lead to significant breakthroughs.
Meta Restructures AI Group Again in Pursuit of Superintelligence
Meta is restructuring Meta Superintelligence Labs (MSL) into four teams—TBD Lab (large language models, led by Alexandr Wang), FAIR (fundamental AI research, led by Robert Fergus), Products and Applied Research (consumer applications, led by Nat Friedman), and MSL Infra (AI infrastructure, led by Aparna Ramani)—to accelerate its pursuit of superintelligence. While no layoffs are planned, the shake-up dissolves the former AGI foundations group and redistributes leadership roles.
OpenAI Staffers to Sell $6 Billion in Stock to SoftBank, Other Investors
Current and former OpenAI employees are preparing to sell around $6 billion in shares to investors, Bloomberg reports. The deal would value OpenAI at $500 billion, making it the world’s most valuable startup, surpassing SpaceX. The sale will also help retain top AI talent, giving employees a chance to cash out while competing companies like Meta offer lucrative packages to lure experts away.
Gemini Live: A more helpful, natural and visual assistant
Google is upgrading Gemini Live to make it more helpful and natural. The update adds real-time visual guidance when you share your camera, starting with the Pixel 10 on 28 August. Gemini now works better with Google Calendar, Keep, and Tasks, with support for Messages, Phone, Clock, and Maps coming soon.
AI Mode in Search gets new agentic features and expands globally
Google has expanded its AI Mode in Search to over 180 countries. The update also adds new agentic capabilities, like allowing users to book restaurant tables or find event tickets. In the US, users in the Labs experiment can also get personalised suggestions and share search results. More languages and regions will be added soon.
Claude Opus 4 and 4.1 can now end a rare subset of conversations
Anthropic updated Claude Opus 4 and 4.1 with a feature that lets the AI end conversations in rare cases of persistent harmful or abusive behaviour. This safeguard, part of ongoing research into AI welfare and safety, allows Claude to exit upsetting interactions or end a chat if the user requests it, while still letting users start new chats or revisit old ones. The feature is experimental and will be improved based on user feedback.
OpenAI launches a sub-$5 ChatGPT plan in India
OpenAI has launched ChatGPT Go in India, a cheaper plan costing ₹399 ($4.60) per month compared to the ₹1,999 ($23) Plus plan. It offers 10 times more usage for messages, image generation, and file uploads than the free version, plus better memory for personalised responses. India, OpenAI’s second-largest market with over 29 million app downloads in the past 90 days, is the first to get this plan, with potential expansion to other regions in the future.
Meta partners with Midjourney on AI image and video models
Meta has teamed up with Midjourney to use its AI image and video generation technology in future products and models, and to keep up with competitors like OpenAI and Google. Midjourney, founded in 2022, has quickly become a leader in AI-generated images and recently launched its first video model. While the deal’s terms haven’t been shared, Midjourney will stay independent, said its CEO in a post on X.
Anthropic in Talks to Raise Up to $10 Billion in New Funding
Anthropic is reportedly close to raising up to $10 billion in new funding, which would significantly boost its valuation beyond the $61.5 billion reached earlier this year.
▶️ The World's First ChatGPT Poisoning (16:56)
Dr Rohin Francis (aka Medlife Crisis) discusses in this video the growing risks of relying on AI chatbots for medical advice and the dangers of vibe medicine. He shares the first published case of a patient who was harmed after following guidance from ChatGPT and explains how AI is adding to the spread of medical misinformation, much like wellness influencers do. While some fear AI will replace doctors, Dr Francis suggests that, for now, it is creating more work for medical professionals who must correct AI-driven misconceptions.
xAI Was About to Land a Major Government Contract. Then Grok Praised Hitler
Earlier this month, the General Services Administration (GSA) awarded contracts to OpenAI, Anthropic and Google, allowing federal workers access to their AI tools as part of the US government's push to adopt AI. Notably missing from that list was xAI. As Wired has learned, xAI was initially set to join the initiative, but the deal collapsed after its Grok chatbot began posting antisemitic conspiracy theories and offensive content on X.
Nvidia reportedly halts production on its H20 AI chips
The Information reports that Nvidia has paused production of its H20 AI chips after Beijing warned Chinese companies not to use them over security concerns and pushed for domestic alternatives. This comes soon after Nvidia was cleared to sell China-specific AI chips. The company insists its chips are secure and that the production pause is just routine supply management.
DeepSeek-V3.1 Release
DeepSeek-V3.1 is the latest version of DeepSeek’s flagship language model, featuring a hybrid design that supports both thinking and non-thinking modes for flexible task handling. It offers smarter tool usage, faster and more efficient reasoning, and improved performance in agent-based tasks. DeepSeek-V3.1 is available to download on Hugging Face. In terms of performance, benchmark results published by Artificial Analysis place the new model with reasoning mode roughly on par with Claude Sonnet 4 Thinking and OpenAI’s GPT-OSS-128B (high).
Introducing Gemma 3 270M: The compact model for hyper-efficient AI
Google has released Gemma 3 270M, a compact, 270-million-parameter model designed for efficient, task-specific fine-tuning. It is intended for high-volume, well-defined tasks such as text classification, sentiment analysis, entity extraction, and query routing, as well as creative and privacy-sensitive applications that can run entirely on-device. Gemma 3 270M is available for download from Hugging Face, Ollama, Kaggle, LM Studio, and Docker.
▶️ Demis Hassabis on shipping momentum, better evals and world models (31:09)
This conversation with Demis Hassabis is essentially a snapshot of where AI stands today and what’s coming next. He talks about the rapid pace of innovation, from Genie 3, which builds interactive, dynamic environments, to the push for universal assistants like Project Astra and Gemini Live. Hassabis highlights the challenge of “jagged intelligence,” where models have impressive strengths but noticeable gaps, and stresses the need for better ways to measure real progress.
Superintelligence Deterrence Has an Observability Problem
The article looks at Mutual Assured AI Malfunction (MAIM), a framework proposed by Dan Hendrycks, Eric Schmidt, and Alexandr Wang, designed to prevent any one nation from gaining total control of superintelligent AI. Similar to the Cold War’s Mutually Assured Destruction model, MAIM relies on deterrence—but instead of nuclear threats, it uses surveillance of rivals and the risk of sabotage. However, the framework is seen as unstable due to poor monitoring, rapid and unpredictable AI advances, decentralised development, and the dangers of espionage. The article argues that stronger tracking methods and clearer rules are essential to maintaining global stability in the AI era.
▶️ Inside a NEOCLOUD AI Cluster with NVIDIA B200 GPUs (15:13)
ServeTheHome visits a new Nvidia B200 cluster built by Supermicro for Lambda and shows the behind-the-scenes challenges that come with running massive AI clusters, such as cooling, powering, and connecting thousands of GPUs. The video also offers a closer look at the Nvidia DGX B200 NVL72 rack, along with an in-depth overview of a single HGX B200 server containing eight GPUs. If you’re interested in how large-scale AI infrastructure works and the hardware that makes it possible, this video is for you.
🤖 Robotics
▶️ Getting a Leg up with End-to-end Neural Networks (2:39)
Boston Dynamics uploaded a new video showing the progress on developing Large Behaviour Models for Atlas. To show what the new model can do, researchers demonstrated how good the robot is at manipulating objects and adapting to changes caused by humans being a bit annoying to it.
Waymo granted first permit to begin testing autonomous vehicles in New York City
Waymo has secured approval to test self-driving cars in New York City, kicking off the city’s first official autonomous vehicle trials. Up to eight cars will be tested in Manhattan and Downtown Brooklyn until late September, with a driver required to stay behind the wheel. As part of the permit, Waymo must share data with the city and work closely with emergency services to keep the testing safe.
Today’s Humanoid Robots Look Remarkable—but There’s a Design Flaw Holding Them Back
This article explores the limitations of current humanoid robots, which rely heavily on advanced AI but have rigid, inefficient bodies that consume excessive energy and struggle in real-world environments. It argues that adopting Mechanical Intelligence (MI)—drawing inspiration from nature to create bodies that passively adapt—could enable robots to move more efficiently, handle complex tasks, and evolve from research prototypes into practical, real-world tools.
▶️ Are AI Humanoid Robots a Bubble? What K-Scale Knows that Elon Doesn’t (59:34)
In this video, Leon from the YouTube channel Leon Ex Machina visits K-Scale Labs, a young company developing open-source humanoid robots. Its founder, Benjamin Bolte, reveals the truth behind the humanoid hype, what it really takes to build a robotics company and the challenges that come with it, as well as how to stand up to competitors with 100 times more funding and build the next generation of embodied AI.
🧬 Biotechnology
AI designs antibiotics for gonorrhoea and MRSA superbugs
Researchers at MIT have used AI to design two new potential antibiotics that can kill drug-resistant gonorrhoea and MRSA. Unlike earlier methods that only tested existing chemicals, this approach created completely new molecules, some of which worked in lab and animal tests. The drugs still need more development and testing before they can be used in people. Additionally, making them in large quantities could be difficult. Experts say this could be the start of a “second golden age” of antibiotic discovery, though more research is needed.
💡Tangents
White House announces chipmaker Intel to give US government 10% stake
The US government will take a 10% stake in Intel with an $8.9 billion investment, funded by CHIPS and Science Act grants. Officials say the deal will boost US semiconductor production and protect national security. The deal is unusual but not without precedent, similar to the government’s temporary stake in General Motors during the 2008 financial crisis.
South Korea deploys hologram police officer to fight crime – and it’s working
Every two minutes in Jodong No 3 Park in Seoul, a holographic policeman appears to remind visitors they’re being watched by CCTV. And it seems to be working, as crime in the area has dropped by 22% since its installation in October 2024. Officials say the glowing figure has made the park feel safer, and they plan to expand the technology to other public spaces.
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!"