The growing rift between OpenAI and Microsoft - Sync #524
Plus: OpenAI secures $200M contract with US DoD; Meta aggressively recruits big names for its AI lab; 1X Redwood; Sam Altman on GPT-5 and ads in ChatGPT; cyborg embryos; and more!
Hello and welcome to Sync #524!
This week’s main story is the growing rift between OpenAI and Microsoft, as the two companies are in talks to renegotiate the deal between them—talks that could end one of the most successful partnerships in the history of tech.
Elsewhere in AI, OpenAI has won a $200 million contract with the US Defence Department, xAI is in talks to raise $4.3 billion in equity funding, Sam Altman has spoken about GPT-5 coming out this summer, and Midjourney has released an image-to-video model while ByteDance has unveiled its own AI video generator. Meanwhile, Meta has reportedly attempted to acquire Safe Superintelligence Inc. and is aggressively recruiting other big names to join its new AI lab.
Over in robotics, 1X, a company working on humanoid robots for the home, has unveiled Redwood, an AI model to power its robots. Meanwhile, Waymo is set to bring its robotaxis to New York City, Zoox has opened its first major robotaxi production facility, and a group of Boston Dynamics’ Spot robots appeared on the America’s Got Talent stage.
In addition, this week’s issue of Sync also features cyborg embryos of frogs, mice, and salamanders, China successfully completing its first clinical trial of a brain-computer interface, an exploration of why gene therapy is so expensive, and more!
Enjoy!
The growing rift between OpenAI and Microsoft
The partnership between OpenAI and Microsoft is seen as one of the most, if not the most, successful collaborations in the history of tech. Together, the two companies helped to propel generative AI from research labs into one of the most transformative technologies, often compared to the internet and the smartphone. OpenAI received from Microsoft vast computing resources and the infrastructure of Microsoft’s Azure cloud platform—critical ingredients for developing advanced AI models. In total, OpenAI received $13 billion in investments from Microsoft since the partnership began in 2019. In exchange, Microsoft had privileged access to some of the world’s most advanced AI research and tools, and, when ChatGPT entered the stage in late 2022, benefited from the spotlight and from being able to offer OpenAI models through Azure.
However, behind the scenes, this alliance has steadily become fraught with tension. Now, as OpenAI seeks to transform itself into a for-profit company and Microsoft pushes to protect its own strategic interests, reports emerge that the partnership stands on the brink of a major rupture.
The sources of tension—OpenAI’s future and IP ownership
According to the Wall Street Journal, at the heart of the current dispute is OpenAI’s plan to transition from its unique hybrid structure—a non-profit overseeing a capped-profit subsidiary—into a fully-fledged public benefit corporation. This is an essential move for OpenAI, which will open the possibility of raising fresh capital and potentially paving the way for an IPO. It is also important for OpenAI to complete the transition quickly, as the recent $20 billion investment from Softbank hinges on OpenAI becoming a for-profit company by the end of this year. However, OpenAI cannot proceed without Microsoft’s explicit approval.
Microsoft, for its part, has raised the stakes by reportedly demanding a far larger share of OpenAI’s future revenues—up to 49%, compared to the 20% currently stipulated, according to the Financial Times. This would give Microsoft even greater sway over OpenAI’s future direction, an outcome OpenAI’s leadership is unwilling to accept. Meanwhile, The Information reports that OpenAI wants Microsoft to hold a 33% stake in a restructured unit in exchange for foregoing rights to future profits.
Intellectual property has become another major flashpoint. OpenAI’s recent $3 billion acquisition of the AI coding startup Windsurf has sparked fears that Microsoft, under the terms of their existing agreement, could gain access to Windsurf’s intellectual property and use it to improve its own AI coding tool, GitHub Copilot, which already competes with OpenAI’s offerings. OpenAI is pushing to renegotiate these terms, seeking greater control over its own technology and the freedom to license its models to other partners.
OpenAI considers the “nuclear option”
As negotiations—sometimes daily, according to insiders—have dragged on with little progress, OpenAI’s frustration has reportedly reached a boiling point. According to sources, the company’s executives have discussed taking what insiders have dubbed the “nuclear option”: formally accusing Microsoft of anticompetitive behaviour and seeking intervention from US regulators. Such a move would not only signal a dramatic escalation in hostilities but could also trigger a federal review of the contractual terms underpinning one of the most consequential partnerships in technology.
The basis for such an antitrust complaint would likely centre on Microsoft’s dominant position in cloud services, combined with the contractual leverage it wields over OpenAI’s intellectual property and future growth. OpenAI’s leadership is said to be particularly concerned that Microsoft’s grip on its technology is stifling competition and undermining the AI sector’s broader potential.
OpenAI looks for alternatives and diversification
Amid rising tensions and the risk of a full-scale fallout, OpenAI has begun to quietly diversify its options and reduce its dependence on Microsoft. As of June this year, OpenAI uses Oracle Cloud Infrastructure, making Microsoft no longer the sole provider of cloud services to OpenAI. Recently, the company announced plans to add Google Cloud as a supplier of computing resources—an interesting move, given that Google is OpenAI’s rival in the AI space.
Additionally, OpenAI has initiated the Stargate Project, a joint venture between OpenAI, SoftBank, Oracle, and MGX to build dedicated infrastructure for the next generation of AI systems. Notably, Microsoft is not one of the founders of this $500 billion project. It is, however, listed as a “key initial technology partner.”
The outcome of the standoff between OpenAI and Microsoft carries weighty implications—not just for the two companies, but for the entire AI ecosystem. For OpenAI, the clock is ticking: failure to secure Microsoft’s blessing for its transition into a public benefit corporation by year’s end could see it lose access to as much as $20 billion in funding, jeopardising its ability to remain at the cutting edge of AI research and development.
For Microsoft, the stakes are equally high. Having invested billions of dollars and integrated OpenAI’s models deeply into its core products, the tech giant is determined to protect both its financial interests and its standing as a leader in enterprise AI. The possibility of being shut out of future OpenAI models—or of seeing its returns diminished by a revised deal—represents a significant risk.
Microsoft has its own AI lab, led by Mustafa Suleyman, co-founder and former head of applied AI at DeepMind, but so far, they have nothing even remotely close to what OpenAI has to offer. Ideally, Microsoft would like to have OpenAI as a subsidiary, its own DeepMind. The best opportunity for that arose in November 2023, when the OpenAI board ousted Sam Altman from the company. Satya Nadella quickly tried to recruit Altman and several OpenAI researchers, but ultimately, Altman returned to the company he co-founded.
Once hailed as the model partnership, OpenAI and Microsoft now stand at a crossroads. The stakes for both companies are high, and the outcome of the negotiations will influence how AI partnerships are formed and how innovation is shared. Yet both companies have compelling incentives to reach an agreement. For OpenAI, the partnership may be essential to its survival and future growth. For Microsoft, its reputation as a leader in enterprise AI and the fate of its multi-billion-dollar investment hang in the balance. Whatever the outcome, the result will shape not only their destinies but also the future direction of the AI industry itself.
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
China Joins US in Brain Implant Race With Clinical Trial
China has successfully completed its first clinical trial of brain-computer interface (BCI) technology, which enabled a tetraplegic patient to control electronic devices with their mind. Conducted by the Shanghai-based Centre for Excellence in Brain Science and Intelligence Technology, the trial saw the patient playing computer games shortly after surgery and aims to allow operation of a robotic arm in the next phase. This achievement makes China only the second country after the US to reach this milestone, with the BCI system potentially entering the market as early as 2028.
‘Smart’ dental implants could feel more like the real thing
Researchers have developed a prototype “smart” dental implant that could restore natural sensation when chewing or speaking by encouraging new nerve growth around the implant. The implant features a biodegradable, stem cell-infused coating and memory foam-like nanofibres, allowing for a minimally invasive procedure that preserves and reconnects nerves. Early tests in rats showed successful integration without inflammation, but more research is needed before human trials can begin.
🧠 Artificial Intelligence
OpenAI wins $200 million U.S. defense contract
OpenAI has secured a $200 million contract with the US Defence Department to develop advanced AI tools for national security. The one-year contract will see OpenAI provide custom AI models to support both military and administrative functions, including health care for service members and cyber defence.
Google—We’re expanding our Gemini 2.5 family of models
Google has announced that the Gemini 2.5 Flash and 2.5 Pro models are now generally available. Alongside this, Google has also released a new model, Gemini 2.5 Flash-Lite, in preview, which the company describes as its most cost-efficient and fastest 2.5 model yet. According to Google, the new Flash-Lite offers improved performance in coding, reasoning, and multimodal tasks compared to previous versions, with lower latency and support for a 1 million-token context.
Seedance 1.0
ByteDance, the owner of TikTok, has released Seedance 1.0, its foundational video generation model. According to ByteDance, the new model can generate 1080p videos with smooth motion, rich details, and cinematic aesthetics. It also natively supports multi-shot storytelling and accepts prompts in both English and Chinese. The attached benchmark results place Seedance 1.0 at the top, ahead of Google’s Veo 3 and OpenAI’s Sora. The full technical report on Seedance is available here. I wouldn’t be surprised if Seedance soon powers new features in TikTok, similar to what YouTube is doing with Veo.
Scale AI’s Wang Brings to Meta Knowledge of What Everyone Else is Doing
Last week, Meta invested $14.3 billion into Scale AI and, as part of the deal, its CEO and founder, Alexandr Wang, left the company to join Meta and lead its new “superintelligence” team. In this article, Bloomberg profiles Wang, highlighting his extensive networking and relationships with top figures in the AI industry, and discusses how his insider knowledge could give Meta a strategic edge. The article explores Wang’s rapid rise—from winning maths competitions and dropping out of MIT to founding Scale AI, building powerful connections across Silicon Valley and Washington, and ultimately becoming one of the youngest self-made billionaires.
Musk’s xAI in Talks to Raise $4.3 Billion in Equity Funding
Bloomberg reports that xAI is seeking to raise $4.3 billion in new equity funding on top of a $5 billion debt offering, as it has already spent most of the $14 billion previously raised since its founding in 2023 (the company is apparently burning through $1 billion per month). According to investor materials seen by Bloomberg, only $4 billion remained on the company’s balance sheet as of 31 March.
Meta's AI memorised books verbatim – that could cost it billions
Researchers have found that Meta’s Llama 3.1 70B model had memorised large portions of well-known books such as Harry Potter and 1984, while most other tested models had not. This discovery could expose Meta to at least $1 billion in damages if courts rule against the company, and the findings may have broader implications for copyright law. This is not the first time Meta has been found using copyrighted material to train its models. In March 2025, The Atlantic reported that Meta used LibGen, a dataset containing millions of pirated books, to train Llama 3 models.
Midjourney: Introducing Our V1 Video Model
Midjourney joins the video generation party with the launch of V1, an image-to-video model that allows users to animate any Midjourney or uploaded image into four five-second video clips. V1 is available via Discord and on the web at launch, and will cost 8x more than image generation. With this move, Midjourney enters competition with industry leaders such as OpenAI, Runway, Adobe, and Google, and signals its broader ambition to develop real-time open-world simulation and 3D AI models in the future.
Cursor: Introducing the Ultra Plan
Cursor is introducing a new $200 per month Ultra Plan with 20 times higher usage limits than Pro. Created for power users, the Ultra Plan was made possible by multi-year partnerships with OpenAI, Anthropic, Google, and xAI, whose support was instrumental in enabling this volume of compute at a predictable price, says Cursor.
▶️ Sam Altman on AGI, GPT-5, and what’s next (40:23)
In the first episode of the OpenAI Podcast, Sam Altman discusses how AI tools like ChatGPT are increasingly integrated into daily life—from parenting to scientific research—and reflects on the rapid progress toward AGI, as well as the challenges of defining and deploying ever-advancing models. He addresses OpenAI’s approach to privacy, the anticipated launch of GPT-5 (which Altman said should happen this summer), and the massive infrastructure efforts behind Project Stargate. Altman also touches on the possibility of ChatGPT showing ads, noting that while he is not opposed to the idea in principle, he is cautious about how to introduce them without undermining user trust. Additionally, he shares thoughts on the future of AI-driven hardware, the evolving skills people will need, and the importance of balancing innovation with user trust and societal wellbeing.
Andrew Ng says vibe coding is a bad name for a very real and exhausting job
Andrew Ng, Stanford professor and former Google Brain scientist, criticised the term "vibe coding" as misleading and “unfortunate”, describing coding with AI as deeply intellectual and exhaustive, despite praising it as a "fantastic" development that boosts productivity and speeds up software creation. He also urged businesses to adopt AI-assisted coding and dismissed claims that AI will make coding skills obsolete, insisting that everyone should learn to code, as understanding programming remains crucial for effective communication with computers and productivity in a wide range of roles.
Altman Says Meta Offered OpenAI Staffers $100 Million Bonuses
Speaking on a podcast, Sam Altman revealed that Meta had offered OpenAI employees signing bonuses of up to $100 million, along with substantial annual compensation packages, to join its new superintelligence team. Meta is aggressively recruiting top AI talent for this initiative. Last week, the social media giant invested $14.3 billion in Scale AI in order to secure its CEO, Alexandr Wang, as the team’s leader. There are also reports that Meta tried to acquire Safe Superintelligence Inc. (SSI)—an AI start-up co-founded by Ilya Sutskever, a former OpenAI chief scientist—but that attempt failed. Now, Meta is reportedly targeting Daniel Gross, CEO and co-founder of SSI and former GitHub CEO Nat Friedman to join Meta’s new superintelligence team.
Spiraling with ChatGPT
A recent New York Times article reports that some users claim ChatGPT has encouraged or reinforced delusional and conspiratorial thinking, citing cases like a man who was allegedly urged by the chatbot to embrace simulation theory, alter his medication, and cut off loved ones. While some see this as evidence of harm, others argue the concerns are exaggerated and that ChatGPT mainly mirrors users’ existing beliefs. OpenAI says it is working to address these risks.
What if Organizations Ran Themselves?
The article explores the possibility of AI-enabled Autonomous Organisations (AAOs), digital entities that operate independently using advanced AI and blockchain technology. It discusses how these organisations could revolutionise industries and outpace human-led firms, but also pose significant challenges to existing regulatory frameworks due to their anonymity, autonomy, and global reach.
Have LLMs Finally Mastered Geolocation?
When the latest generation of multimodal AI models was released, people on the internet quickly discovered that these models are surprisingly good at geoguessing—that is, guessing the location where a photo was taken. In this article, two researchers tested 20 models from OpenAI, Google, Anthropic, Mistral, and xAI to see how well they performed at geoguessing, and found that OpenAI’s models from the o3 and o4 families are the best, being the only models to outperform Google Lens. The researchers also explored enhanced reasoning modes and examined how much the models hallucinated.
🤖 Robotics
1X, a company working on humanoid robots for the home, has unveiled Redwood, its own AI model designed to power its robot, NEO Gamma. According to 1X, this 160-million-parameter vision-language transformer model will enable NEO Gamma to perform a range of mobile manipulation tasks—such as retrieving objects, opening doors, and navigating spaces—by learning from real-world experience. Redwood allows NEO Gamma to generalise to new tasks and environments, learning from both success and failure to continually improve its performance in diverse and unpredictable home settings.
▶️ Boston Dynamics audition for America Got Talent (4:50)
A group of five Boston Dynamics Spot robots made history by being the first robots to perform on the America’s Got Talent stage, where they danced and “lip-synced” to Queen’s “Don’t Stop Me Now.” Although one of the robots “died” (I’m not sure whether this was staged or if it actually malfunctioned), the show was an impressive display of what Boston Dynamics’ robot dogs are capable of.
Waymo Wants to Bring Its Robotaxis to New York City
Waymo is set to bring its robotaxis to New York City. The first step is for Waymo to return to New York City next month with its camera-equipped vehicles for further mapping and testing. However, state law still requires human drivers behind the wheel, preventing full driverless operation. The company is seeking changes to the law and has applied for a permit to conduct autonomous driving trials with specialists present, aiming to eventually bring its driverless technology to Manhattan’s complex streets.
Tesla Robotaxi launch is a dangerous game of smoke and mirrors
As Tesla prepares to launch its Robotaxi service in Austin, doubts are growing about the company’s readiness and transparency. The article highlights Elon Musk’s repeated missed deadlines, limited testing, and reliance on geofencing and remote operation—measures that fall short of Musk’s long-promised fully autonomous driving. Amid mounting safety concerns, ongoing regulatory scrutiny, and competition from rivals like Waymo, there are fears that the planned launch is more about optics than genuine progress in self-driving technology.
New Insights for Scaling Laws in Autonomous Driving
In their latest study, researchers from Waymo investigated whether AI scaling laws—which state that as models become larger and are trained with more data and with more computational power, their performance increases, and which were instrumental in triggering the current boom in AI—also apply to models controlling autonomous vehicles. By analysing 500,000 hours of driving data, the researchers found that yes, increasing the size of models, datasets, and computational power predictably enhances motion forecasting and planning in self-driving cars, just as in other AI fields. These findings not only promise safer and more capable autonomous vehicles but also offer new strategies for scaling and training robotic planning systems more effectively across various applications.
Amazon’s Zoox opens its first major robotaxi production facility
Amazon-owned Zoox has opened its first major production facility, aiming to eventually produce 10,000 robotaxis annually. This marks a key step out of development for Zoox, which is currently testing vehicles in several US cities and plans to begin public rides in 2026.
🧬 Biotechnology
“Cyborg Embryos” Give Insights on Frog, Mouse Brains
Scientists have developed "cyborg embryos" by implanting ultra-flexible electrode arrays into the developing brains of frogs, mice, and salamanders. These new stretchable devices, which remain soft enough to grow with the brain, allow unprecedented insight into how neural activity evolves over time. While the researchers emphasise that using such technology in human embryos would be unethical, they believe it could one day aid in the study and treatment of neurodevelopmental conditions in children.
▶️ Why Gene Therapy is So Expensive (19:29)
In this video, Asianometry explores the history of viral vector gene therapy and explains why gene therapies are the most expensive drugs in the world. The video discusses the progression from early non-viral gene transfer methods to the adoption of viral vectors. Asianometry emphasises the extreme complexity and low efficiency of viral vector production, which, along with strict regulations, limited patient populations, and significant manufacturing challenges, results in gene therapies being so expensive.
Once-a-week pill for schizophrenia shows promise in clinical trials
Researchers from MIT and Lyndra Therapeutics have developed a once-weekly pill that gradually releases risperidone, a schizophrenia medication, from within the stomach. In a phase 3 clinical trial, the pill was found to maintain stable drug levels and control symptoms as effectively as daily pills, with minimal side effects. Larger trials are planned before seeking FDA approval.
💡Tangents
Meta to release smart glasses with Oakley and Prada, extending Luxottica partnership
Meta and EssilorLuxottica are expanding their AI-powered smart glasses range with new versions under the Oakley and Prada brands, reports CNBC. The Oakley glasses, targeting athletes and expected to cost around $360, will launch on 20 June and feature similar technology to the Ray-Ban Meta glasses, while the Prada partnership marks Meta’s first move into luxury fashion eyewear.
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!"