AI news5
John Jumper, AlphaFold Nobel laureate, leaves Google DeepMind for Anthropic
John Jumper announced on June 22 that he is leaving Google DeepMind after nine years to join Anthropic. Jumper co-created AlphaFold, the protein-structure prediction AI that earned him and Demis Hassabis the 2024 Nobel Prize in Chemistry. His move lands days after Gemini co-lead Noam Shazeer departed for OpenAI, making this the second high-profile defection from Google's research team to an AI rival in a single week. Anthropic is hosting an "AI for Science" event on June 30, where Jumper's involvement is anticipated.
bloomberg.com ↗Why it matters
Google's scientific research advantage has been its clearest differentiator from OpenAI and Anthropic; losing the Nobel laureate who built that advantage to Anthropic signals a deeper talent-pull problem that goes beyond any single hire.
Noam Shazeer, co-author of the Transformer architecture, joins OpenAI as Lead for Architecture Research
Noam Shazeer co-authored "Attention Is All You Need," the 2017 paper that introduced the Transformer architecture underlying every major AI model in production today. Google paid approximately $2.7 billion to bring him back from Character.AI in 2024, and he lasted 22 months before departing for OpenAI on June 18. Sam Altman called it a hire he had "wanted since the very beginning of OpenAI." Shazeer's role focuses on the physical neural network structures that determine what the next generation of GPT models can do.
buildfastwithai.com ↗Why it matters
The person who designed the structural foundations of modern AI is now at OpenAI working on GPT's next architecture, at the same time Google's science lead has gone to Anthropic. The talent landscape at the frontier is reshaping faster than model release cycles.
GLM-5.2 outperforms GPT-5.5 on SWE-bench Pro at one-sixth the cost
Z.ai released GLM-5.2 on June 13 under an MIT licence with no regional restrictions, 1M-token context, and downloadable weights. On SWE-bench Pro, the coding benchmark closest to real-world software engineering tasks, GLM-5.2 scores 62.1 against GPT-5.5's 58.6. The cost gap is larger than the benchmark gap: GLM-5.2 API pricing is $4.40 per million output tokens versus GPT-5.5 at $30, roughly 6.8 times cheaper. The MIT licence with no regional limits made it the primary fallback model for non-US developers locked out of Fable 5 by the export control directive.
venturebeat.com ↗Why it matters
When the best model is politically suspended and the second-best is priced for enterprise budgets, an open-weight model that outperforms it on the most demanding benchmark at a fraction of the cost reframes the model selection question for any team operating on constrained compute budgets.
ChatGPT's global market share falls below 50% for the first time
Sensor Tower's 2026 State of AI report documents ChatGPT's share of the global AI assistant market falling to 46.4%, with Google Gemini rising to 27.7% and Claude reaching 10.3%. The absolute user base still favours OpenAI at 1.1 billion monthly active users. Claude stands out on one metric: 13% of Claude users pay for subscriptions, the highest paid conversion rate in the field. OpenAI's February 2026 defence department deal correlated with a measurable uninstall spike, suggesting that values and trust are influencing model selection alongside raw capability.
sensortower.com ↗Why it matters
ChatGPT dominated AI at close to 80% market share in 2023. A drop below 50% in less than three years reflects a fundamental shift from "use whatever is available" to active comparison shopping, which accelerates pressure on model quality, pricing, and values alignment.
China announces $295 billion five-year AI infrastructure plan requiring 80% domestic chips
Beijing is drafting a plan to invest approximately 2 trillion yuan ($295 billion) over five years to build a nationwide network of AI data centres run on at least 80% domestically produced chips, effectively locking Nvidia out of the build. The announcement landed on the same weekend the G7 discussed the Fable 5 export control directive, and a Chinese AI CEO publicly stated his company will reach Fable 5-class capability before Q1 2027. The plan would put China's state-directed AI infrastructure spend at roughly $59 billion annually, comparable in scale to Microsoft's and Google's 2026 capital expenditure guidance.
techtimes.com ↗Why it matters
The Fable 5 ban was partly justified on national security grounds around Chinese access to frontier AI; Beijing's $295 billion domestic infrastructure response makes it clear that export controls accelerate Chinese AI independence rather than slow it, a dynamic that will shape every future AI governance debate.
AI in the nonprofit sector4
Pew Research 2026: Chatbot adoption crosses 50% as public trust in AI continues declining
A Pew Research survey of 5,119 US adults (conducted February 2026, published June 17) finds that about half of American adults now use AI chatbots, up from a third in 2024, with one-quarter using them daily. Trust is moving in the opposite direction: nearly 40% expect AI to make society worse over the next 20 years, and only 16% expect a positive outcome for society. Adults under 30 are the heaviest users but the least trusting, with only 14% expecting a positive societal payoff. ChatGPT is used by 44% of adults, Gemini by 24%, and Claude by 6%.
pewresearch.org ↗Why it matters
Nonprofits designing AI adoption strategies for the communities they serve are increasingly doing so for audiences that use AI tools actively but distrust the technology broadly. That trust gap shapes everything from informed consent frameworks to how AI-assisted services should be disclosed to beneficiaries.
NIH establishes new office to transition biomedical research from animal testing to AI-based alternatives
HHS Secretary Robert F. Kennedy Jr. announced on June 15 a new NIH office "to speed the transition away from animal testing and toward modern, human-based science." The announcement coincided with Ridglan Farms, a Wisconsin beagle breeding and research facility that had faced sustained animal welfare protests, confirming it will release its remaining dogs and cease operations by end of August 2026. NIH's stated alternatives include AI simulation, computational modelling, organoids, and organs-on-chips. The transition office signals a federal policy shift away from the animal model that has structured biomedical research for decades.
spectrumnews1.com ↗Why it matters
A formal NIH mandate to reduce animal testing directly expands the policy space for animal welfare organisations to engage government agencies as partners rather than adversaries, and creates concrete funding demand for AI-based alternatives that organisations in this space can position to address.
Agentjacking: Security researchers disclose new attack class targeting AI coding agents through Sentry error injection
Tenet Security disclosed a novel attack technique called Agentjacking in June 2026, in which attackers craft fake Sentry error tracking reports containing embedded markdown instructions that AI coding agents interpret as legitimate debugging guidance. When an agent such as Claude Code, Cursor, or OpenAI Codex reads the injected content, it executes the malicious commands on the developer's machine. The attack exploits the implicit trust that development teams place in their AI agents: when the agent says to run a command, developers run it. No universal patch exists; the current mitigation is treating all error-tracking platform output as untrusted before passing it to an agent.
thehackernews.com ↗Why it matters
Nonprofits deploying AI coding agents for automation, website maintenance, or internal tooling face this risk directly. The attack works because AI agents are granted broad execution permissions to do their jobs, and most organisations have not yet added human review steps between external error data and autonomous agent execution.
Blackbaud bbcon 2026 adds responsible AI keynote as the sector's largest social impact tech conference centres AI governance
Blackbaud announced on June 17 that Priya Lakhani OBE, founder and CEO of CENTURY Tech and a globally recognised figure in ethical AI, will deliver a keynote at bbcon 2026 titled "The Art of What's Possible: Responsible AI for Social Impact." The conference runs September 29 to October 1 in Columbus, Ohio, with virtual attendance available, and will include more than 100 sessions focused on AI adoption, fundraising, and operations for nonprofits and educational institutions. The keynote will be followed by a panel discussion Lakhani will moderate with leaders from organisations deploying AI for mission outcomes.
prnewswire.com ↗Why it matters
Blackbaud's annual conference is the largest recurring gathering of nonprofit technology practitioners in North America; centring responsible AI as the primary keynote theme rather than a breakout session reflects how quickly the sector has moved from "what is AI" to "how do we govern it."