Google DeepMind lost two high-profile researchers in a week marked by a flurry of departures across major AI labs.

Why it matters: The talent wars continue and have intensified at the highest levels as many AI developers believe that artificial general intelligence (AGI) is on the horizon.

Driving the news: Noam Shazeer announced last week that he was leaving Google DeepMind for OpenAI. Google previously paid more than $2 billion to acqui-hire Shazeer and part of his Character.ai team.

• Shazeer co-authored the pivotal 2017 paper "Attention Is All You Need," which introduced the transformer architecture — the "T" in ChatGPT.
• Two days later, John Jumper, who shared the 2024 Nobel Prize in Chemistry for AlphaFold, said he was also leaving Google DeepMind for Anthropic.
• Meanwhile, Barret Zoph, who left Thinking Machines in January after "alleged misconduct," rejoined OpenAI and is now departing the company for a second time.
• Nvidia also acqui-hired the team behind Essential AI, including AI researcher Ashish Vaswani.

The big picture: Even as companies race to automate more of AI research itself, they're placing extraordinary value on a tiny number of humans who know how to direct that work.

• Plus, one or two key hires can also help spur further hiring while a major departure can spur further defections.

Yes, but: Not all of the recent personnel moves are the same.

• Shazeer and Jumper are genuine lab-to-lab defections.
• Essential AI is more along the lines of the expected consolidation of startups.
• Zoph's latest departure remains unexplained.

Between the lines: The most coveted researchers offer more than technical knowledge contained in papers or code.

• They bring judgment about which ideas to pursue, experience running enormous experiments and the ability to recruit other sought-after scientists.
• For top researchers, deciding where to work often involves a complicated calculus: potential financial rewards, access to computing power, each company's prospects of leading the field, and whether its leadership will wield that power responsibly and in ways that align with their own beliefs.
• On the financial front, Anthropic and OpenAI have the advantage of forthcoming IPOs, which could have greater upside than can be offered by the other major players, which are already publicly traded.
• The other factors are more subjective, with perceptions around leadership and responsibility varying from person to person.

Reflection — an Nvidia-backed open-source AI startup — signed a major compute agreement with SpaceXAI, securing immediate access to chips and hardware from the SpaceX Colossus 2 data center.

Why it matters: If open-source AI companies are going to compete with the frontier AI labs, they need compute. Reflection is now getting it from the same source its competitors are.

Driving the news: After an initial ramp period, Reflection will pay SpaceXAI $150 million a month starting July 1, 2026, through 2029.

• The deal gives Reflection access to high-end reasoning GB300 chips and other hardware inside Colossus 2, expanding the compute available to train its models.
• Either company can end the deal with 90 days' notice after the first three months.
• The deal follows similar compute agreements. Both Anthropic and Google are expected to spend billions for access to Elon Musk's compute capacity.

Follow the money: It's a reminder that the AI boom's biggest players are increasingly investors, suppliers and customers to one another, often all at the same time.

• Nvidia invested $800 million in Reflection, which is now getting access to Nvidia chips purchased by SpaceX.
• Nvidia is helping fund its next generation of customers, while some startups are dodging the multibillion-dollar cost of building their own data centers by leasing compute from others.

Between the lines: Compute is the scarce resource fueling the AI race.

• Some investors have called Reflection the "DeepSeek of the West," the Wall Street Journal reported. The company is still training its models.
• Having more compute capacity while training its models could allow Reflection to compete more directly with frontier AI labs like OpenAI and Anthropic.

Zoom out: The deal comes amid a surge in interest surrounding open-source models.

• Anthropic shut down access to its most powerful models after the Trump administration threatened to block access for foreign nationals, raising questions about who gets to control access to intelligence.
• Unlike closed models, open models can often be downloaded, inspected and modified. Many of the strongest open models now come from Chinese labs.

What they're saying: "Recent events highlight how important open source is to the AI ecosystem, with more nations and enterprises recognizing the risks and costs associated with exclusively depending on closed models," a Reflection spokesperson said.

OpenAI is releasing a more permissive version of its cybersecurity model, which the company says is designed for advanced, authorized security work.

Why it matters: Even as Anthropic remains in limbo with the U.S. government, the race to get advanced AI models into the hands of cyber defenders continues to heat up.

Driving the news: OpenAI is updating its GPT-5.5-Cyber model — which is only available to vetted cybersecurity companies and researchers — so it is both "more permissive and more capable for advanced, authorized cybersecurity work," according to a blog post.

• The updated model can perform deeper analysis across large codebases, identify security-relevant components, validate likely vulnerabilities, and develop and test software patches.
• OpenAI says the updated GPT-5.5-Cyber achieved an 85.6% score on CyberGym, a benchmark that measures whether an AI agent can reproduce known software vulnerabilities in testing environments, compared with 81.8% for GPT-5.5.

The big picture: OpenAI is expanding access to its cybersecurity tools at a time when policymakers are paying closer attention to how advanced AI systems are evaluated, tested and deployed.

• AI developers face a difficult balancing act: getting powerful cyber capabilities into the hands of legitimate defenders and researchers while limiting opportunities for malicious use.

Between the lines: OpenAI is also rolling out a series of new programs and capabilities designed to let vetted cybersecurity companies use its models to help secure customer environments.

• The company is launching the OpenAI Daybreak Cyber Partner Program, which allows participating security vendors to use GPT-5.5 with Trusted Access for Cyber in the products and services they provide to customers.
• Previously, approved organizations primarily used the company's cyber models on systems they owned or were authorized to test.
• OpenAI is also helping fund Patch the Planet, an initiative founded with Trail of Bits and developed in collaboration with HackerOne and Calif, aimed at helping open-source maintainers manage and remediate vulnerabilities identified with AI-assisted tools.

• Meta paused its AI program, which tracked employee keystrokes, after finding that sensitive data on workers was left widely accessible. (Business Insider)
• SpaceX plans to raise $20 billion by selling bonds, after its $86 billion IPO. (Axios)
• Autodesk is committing $350 million to drive a broader AI upskilling effort, joining a growing list of tech companies investing in this.

Ina here. Forget Google Alerts for your name. Have you checked how prevalent you are in the weights of various AI models? You can type your name in at In The Weights to find out.