AI model overturns Erdős conjecture & Google rebuilds Search around AI - News (May 22, 2026)

We’ll start in the world of AI and discovery. OpenAI says its newest general-purpose reasoning model has produced an original proof that disproves a well-known discrete geometry conjecture posed by Paul Erdős back in 1946. That’s a big claim in a field that doesn’t hand out credit easily—and it’s especially notable because OpenAI previously took heat after earlier “Erdős problem” claims turned out to overlap with already-known results. This time, the company pointed to supportive remarks from respected mathematicians, including some who had criticized the earlier episode. If the broader community validates the proof, it’s another sign that AI may be moving from summarizing knowledge to genuinely extending it—at least in some corners of mathematics.

Staying with AI, Google says it’s about to remake how people use Search. At its 2026 I/O conference, the company described a “complete reimagining” of the search bar, shifting the experience toward AI-generated answers and a conversation-style interface powered by the Gemini 3.5 Flash model. Google is also leaning into richer inputs—like images and files—and even hints that Search can work more like an assistant that keeps an eye on topics and reports back. The immediate, real-world question is who wins and who loses when the answer shows up directly on the results page. Publishers, creators, and small businesses have been warning for months that AI summaries can mean fewer clicks, and fewer clicks can mean less revenue. This isn’t just a product tweak—it could change the economics of the open web.

Google is also pushing AI deeper into the smart home—though with a different strategy. The company says it wants Google Home to become a “full-stack AI offering,” combining its Home platform with Gemini-powered features that partners can build into their own devices and services. Read between the lines, and it looks like Google wants an ecosystem where internet providers, security companies, and device makers sell ongoing subscriptions on top of Google’s foundation. That could make smart homes more capable, but it also nudges the category further toward recurring fees—and potentially away from Google making as many of its own first-party devices.

Now to medical research, where a set of studies this week put the spotlight on immune cells and precision biology. In Australia, researchers at the Garvan Institute and UNSW Sydney captured striking footage: macrophages—immune cells typically known for cleaning up debris—actively attacking and engulfing live melanoma cells in real time. Using advanced microscopy in mice, they saw a particular macrophage subset, marked by a protein called CD169, patrolling tumor edges and essentially “eating” cancer cells. When those macrophages were depleted, tumors grew larger, suggesting these cells play a protective role. What makes this especially interesting is what it didn’t rely on. The anti-tumor activity appeared to occur independently of T cells and B cells—the immune players most current immunotherapies focus on. The team also found this macrophage population in healthy human skin and around the margins of human melanoma tumors, hinting it’s not just a mouse phenomenon. If future work holds up, it could point toward therapies aimed at boosting these macrophages—potentially helping in so-called “cold” tumors that don’t respond well to today’s checkpoint treatments.

Another immunotherapy update comes from UCLA, where scientists reported a preclinical CAR-T approach against glioblastoma—one of the toughest brain cancers to treat. Their engineered CAR-T cells target a glioblastoma-associated marker and are also designed to release immune-activating signals, aiming to pull more types of immune cells into the tumor. That matters because glioblastoma often dodges single-target therapies: some cancer cells simply don’t display the target, survive, and come roaring back. In mouse models with intact immune systems, the “cytokine-armored” CAR-T strategy improved tumor control and appeared to reduce one of the key barriers in solid tumors—escape through tumor diversity. Safety is the big caveat here, because immune-boosting signals can also trigger dangerous inflammation. The team tested additional design strategies to limit toxicity while keeping anti-tumor effects, and they say they’re working toward a first-in-human trial, pending further preclinical work and funding.

Next, a different kind of precision medicine: AI-designed proteins that act more like custom-made switches. Researchers at the University of Washington’s Institute for Protein Design, working with startup Skape Bio, reported tiny “miniproteins” that can selectively turn certain G protein-coupled receptors—GPCRs—on or off. GPCRs are among the most important drug targets in medicine, involved in everything from hormones to mood to metabolism, but they’re notoriously tricky to control precisely because their binding regions can be deep and flexible. The team’s approach uses AI-driven design to build very small proteins that fit those hard-to-reach spots and stabilize either the active or inactive state of the receptor. They also built a high-throughput screening method that tests huge numbers of candidates directly in living human cells, which could speed up the hunt for effective designs. In a mouse study, one candidate performed comparably to an existing drug, with fewer side effects—a promising signal, even if it’s early days.

On the business-and-health intersection, Eli Lilly announced phase 3 results for its experimental weekly obesity injection, retatrutide, suggesting larger average weight loss than current leading drugs. In a trial of more than two thousand adults without diabetes, the highest dose group lost, on average, roughly 28 percent of body weight over 80 weeks—an eye-catching figure in a field that’s already moving fast. Side effects were common and mostly gastrointestinal, and they increased with dose—an important reminder that real-world success depends on what people can tolerate for the long haul. If the full data holds up under peer review and regulators sign off, these results could raise expectations for medical weight loss yet again—and intensify competition in a market that’s already reshaping healthcare, from cardiology to orthopedics to insurance coverage decisions.

And obesity also leads us to a surprising science finding from Germany: an AI-driven “body atlas” that maps how obesity affects the entire body at cellular resolution, rather than looking at organs one at a time. The platform—called MouseMapper—combines whole-body imaging with AI analysis to build detailed 3D views. In mice fed a high-fat diet, researchers saw widespread inflammation and tissue changes across multiple organs. The standout, unexpected result was evidence of nervous-system damage: the trigeminal nerve showed fewer branches and endings, and the mice had reduced sensitivity in behavioral tests. Even more notable, when scientists checked human samples, they found similar molecular changes in the trigeminal ganglion in people with obesity. It’s not the final word, but it reframes obesity as a whole-body condition in a very literal way—and could help researchers connect symptoms that don’t always get linked to metabolism alone.

Now to geopolitics. President Donald Trump says he plans to speak with Taiwan’s President Lai Ching-te about a potential US arms sale—something that would break decades of diplomatic practice. Since the US switched formal recognition to Beijing in 1979, American support for Taiwan has been substantial but carefully managed, often avoiding direct leader-to-leader contact. China condemned the idea and warned against what it calls signals to separatism. Trump hasn’t decided whether to proceed with a reported multi-billion-dollar arms package, and Beijing is reportedly tying other US–China engagement to that decision. Coming right after Trump’s summit with China’s leader Xi Jinping—where Taiwan was described as a major flashpoint—this is the kind of move that can quickly raise temperatures across the Taiwan Strait, and across global markets watching for stability.

Finally, a major global health development: the World Health Organization has declared the Ebola outbreak in the Democratic Republic of the Congo a Public Health Emergency of International Concern after confirmed cases appeared in neighboring Uganda. The outbreak is being driven by the Bundibugyo strain, and that detail matters because existing Ebola vaccines were largely built around a different strain. Health officials emphasize that Ebola spreads through direct contact with infected bodily fluids—not through the air—and people aren’t contagious before symptoms begin. But cross-border movement can outpace detection, especially in regions with frequent travel and limited resources. With no approved Bundibugyo-specific vaccine or targeted treatment, the response depends on fast diagnosis, isolation, contact tracing, safe burials, and supportive hospital care. The emergency declaration is meant to unlock funding and coordination quickly—and it’s also a reminder that preparedness needs to cover more than one viral “version.”