Doctors scanned the thymus for decades. AI finally asked what it meant.🌿

For decades, clinicians largely ignored the thymus in adult patients. The gland, tucked behind the sternum, is most active in childhood, when it trains T cells to distinguish self from foreign. After puberty, conventional thinking held, it atrophies and fades from clinical relevance. Routine CT scans captured it incidentally. Nobody analyzed what they were seeing. Researchers at Mass General Brigham just changed that. Using AI to analyze CT scans from more than 25,000 adults across multiple hospitals, they found that thymus health, measured using a radiomics model trained on those scans, is one of the strongest predictors of lifespan and cancer outcomes the field has ever identified. The study was published in Nature. The numbers are striking. Adults with healthier thymuses had roughly 50% lower overall death risk and 63% lower cardiovascular death risk. Among patients with lung cancer specifically, a healthier thymus correlated with 36% lower risk of developing the disease. For patients already undergoing immunotherapy, the thymus difference was even more pronounced: 37% lower cancer progression and 44% lower mortality risk. The mechanism is immunological. A healthier thymus produces more active T cells, which are central to both baseline immune function and the effectiveness of checkpoint inhibitor therapies. The AI model didn't invent a new biomarker. It revealed that a measure of immune health we had been routinely capturing for years was, in fact, highly predictive, and we weren't using it. What we're still uncertain about: this is a retrospective observational study. Correlation between thymus health and longevity does not immediately tell us whether thymus health is modifiable, or whether improving thymus function translates into improved outcomes. The clinical utility depends on whether this becomes a standard read in CT interpretation and whether it changes treatment decisions in practice. Both are open questions.

Source: ScienceDaily

Apple selected 350 winners in its 2026 Swift Student Challenge, and the standout theme was accessibility-first AI. Distinguished winners built tools for navigating flood zones in Accra using real-time satellite imagery, playing viola without a physical instrument, drawing with hand tremors, and getting presentation coaching for neurodivergent speakers. Fifty winners were invited to WWDC. The cohort signals that a generation of developers is building for underserved users first, not as an afterthought.

The newly independent OpenAI Foundation announced a $250 million commitment to help workers navigate AI-driven job displacement. Three priorities: studying how labor markets are actually changing, supporting displaced workers directly, and identifying new wealth-distribution mechanisms. First programs are expected before the end of 2026. It is the largest single commitment from an AI lab directed at the workforce transition problem rather than at model capabilities.

A Columbia Climate School analysis argues that AI data centers, despite their electricity consumption, could act as multipliers for clean energy buildout. Their steady, bankable demand is making enhanced geothermal commercially viable for the first time: the technology has an estimated 100 GW of potential, but only 4 GW has been deployed. The piece reframes the "AI versus climate" narrative. The same power appetite raising alarms may be financing the grid infrastructure that the clean energy transition requires.

Source: Columbia Climate School

OpenAI built a biodefense model. What that means, carefully.

On May 29, OpenAI unveiled GPT-Rosalind, a specialized model built for biodefense and pandemic preparedness. The name is a deliberate nod to Rosalind Franklin, the crystallographer whose X-ray diffraction work was central to identifying the structure of DNA. The launch has received less coverage than OpenAI's consumer releases. It deserves more careful attention. The program has two tracks. The developer track gives biodefense researchers API access for pathogen surveillance, outbreak modeling, and public-health preparedness work. The government track, briefed to the White House before launch, is designed for national-security applications at cleared facilities. Launch partners include Lawrence Livermore National Laboratory, Johns Hopkins Applied Physics Laboratory, and CEPI, the Coalition for Epidemic Preparedness Innovations. This is not a product launch in the conventional sense. GPT-Rosalind is not publicly available, and it is unlikely to be. It is, instead, a deliberate structural choice: a frontier AI lab deciding to build and deploy a model specifically for biosecurity use cases, with curated access from day one.

Why the two-track architecture matters

The decision to separate developer access from government access is a design choice about risk containment. Biodefense AI occupies one of the most sensitive positions in the dual-use debate: any model capable of modeling pathogen behavior for defensive purposes is, by definition, trained on information that could be exploited for harmful purposes. The two-track structure is an attempt to create graduated access without fragmenting capability. Whether it works depends entirely on how the access controls are designed and audited, neither of which is yet public.

Our PM + Risk Manager lens

The product problem here is not capability. It is a trust architecture. The buyers, government agencies, national laboratories, and international preparedness organizations are among the most demanding and slow-moving procurement environments in the world. A credible deployment depends on establishing not just that the model can do useful biosurveillance work, but that the access controls and audit trails meet the standards those buyers require. LLNL and Johns Hopkins APL, as launch partners, are not marketing choices. They are credibility infrastructure: organizations whose participation signals that the security requirements have been, at minimum, seriously considered. The real PM question is whether OpenAI can build a durable institutional relationship with the national-security procurement ecosystem, which operates on timelines and accountability frameworks that have no equivalent in consumer AI.

The dual-use concern with biodefense AI is not theoretical. The information needed to build effective pandemic surveillance is closely related to the information that could, in principle, lower barriers to biological harm. The biosecurity community has been grappling with this longer than most AI risk frameworks have existed. GPT-Rosalind's government track is one architectural response, but it raises a second-order question: who audits the access controls? CEPI and LLNL are credible partners, but their participation does not constitute independent oversight of OpenAI's security architecture. An additional concern is political continuity: the program has been briefed to the current White House. What happens to the access controls, the oversight model, and the institutional relationships if the political environment shifts? Durable biosecurity infrastructure cannot depend on a single administration's priorities.

The next 12 to 24 months

Watch for CEPI's first public deployment of GPT-Rosalind in pathogen surveillance, and for whether the developer API track attracts independent biosecurity researchers willing to test and publish on its guardrails. The EU is developing its own framework for high-risk AI in biosecurity applications, and regulatory divergence between US and EU approaches will shape how international preparedness programs can access the model. The most important signal to watch is whether the government track produces any verifiable public-health outcome before the next major outbreak scenario tests it under real conditions.

Source: Axios

GPT-Rosalind is OpenAI's specialized model for biodefense and pandemic preparedness, launched May 29. It is not a consumer tool. Access is via a curated developer API for credentialed biodefense researchers and a government track for national-security applications at cleared facilities. Launch partners include Lawrence Livermore National Laboratory, Johns Hopkins Applied Physics Laboratory, and CEPI. If you work in public health, outbreak modeling, or biosurveillance, this is the first frontier AI model purpose-built for your field. Information on developer API eligibility is available through OpenAI's safety and research partnership channels.

→ Read more: Axios

The thymus story and the Rosalind story are both about information that was always there, just not being used the right way. Decades of thymus data sitting in CT archives, analyzed for the first time. Biodefense AI is built from existing knowledge about pathogens, organized toward prevention rather than harm.

What piece of existing, underused data do you think AI is most likely to unlock next?

Hit reply. We read every response.