OpenAI claims new Erdős proof & Google Search goes full AI - Tech News (May 22, 2026)

OpenAI is taking another swing at “AI does real math,” and this one is getting more careful attention. The company says a general-purpose reasoning model produced a new proof that disproves a long-standing discrete geometry conjecture first posed by Paul Erdős in 1946. What makes this notable isn’t just the claim—it’s the context. OpenAI previously got burned when an executive suggested earlier models had solved multiple Erdős problems, only for researchers to point out the results weren’t actually new. This time, OpenAI is citing feedback from established mathematicians, and the alleged breakthrough is that the model found a better construction than the grid-like patterns people assumed were best. If the wider community validates it, it’s a genuine milestone for AI-assisted discovery, not just AI-generated math-flavored text.

Google, meanwhile, is moving fast to make AI the front door to the internet—whether publishers like it or not. At I/O 2026, Google said it will “completely reimagine” the search bar, calling it the biggest change to Search in over 25 years. The new experience is built around Gemini, with AI-generated summaries and a more conversational flow designed to keep you asking follow-up questions. Google also wants Search to accept richer inputs—think images, files, videos, even what you currently have open—and to proactively track topics through agent-like features that can send updates. The big reason this matters: if answers increasingly live inside Google’s interface, the web’s traditional bargain—search sends traffic out, sites provide content—gets renegotiated in a way that could squeeze publishers and small businesses that rely on referrals.

Staying with Google, there’s a new laptop storyline brewing: the so-called “Googlebook.” It’s still early and partly speculative, but the idea gaining traction is that Google wants to go beyond the cloud-first Chromebook identity and turn laptops into an AI-first platform with Gemini integrated at the system level. The pitch, at least as described, is tighter continuity between phone and laptop, more consistent Gemini workflows across everyday apps, and new interface ideas like a more context-aware pointer that can trigger actions without digging through menus. The unanswered questions are the practical ones—what the hardware looks like, what it costs, and how much runs on-device versus in the cloud. But the strategic point is clear: Google appears to want a coherent “AI computer” story that competes directly with Microsoft’s Copilot+ push.

Google’s Chrome team also used I/O to highlight how quickly AI is reshaping software creation—this time in the world of browser extensions. Google says developer registrations are up sharply over the past year, and a noticeable share of new extensions are now being built with help from AI. To make that less chaotic, Chrome is packaging best practices into guidance that AI coding agents can follow, and Chrome DevTools is becoming more agent-friendly for extension debugging and testing. There are also changes for teams and enterprises: more granular roles in the developer dashboard, and a way to distribute private extensions to organizations that explicitly approve them. And in a nod to cross-browser reality, Chrome is smoothing compatibility so developers don’t have to maintain as many workarounds when targeting more than one browser.

On the infrastructure side, NVIDIA says its new Vera CPU platform is now in full production, and that it’s delivering CPU racks to major AI customers, including OpenAI, SpaceX, Anthropic, and Oracle. NVIDIA is positioning this as more than just “we also make CPUs now.” It’s framing Vera as a control-and-orchestration companion to its GPU-heavy AI stacks, and it’s even floating eye-popping revenue expectations for CPUs as demand for AI hardware keeps ballooning. The other key detail: NVIDIA is warning it expects supply constraints to continue, and it pointed to memory availability as a bottleneck. Translation: the AI buildout is now constrained by the broader supply chain, not just by who can design the fastest chip.

Robotics had the most internet-ready moment of the day. Figure AI drew huge attention by livestreaming its humanoid robots placing packages onto a conveyor belt—then turning what was supposed to be an eight-hour demo into an always-on endurance run. Viewers treated it like a spectator sport, complete with robot nicknames and prediction-market bets, and the company leaned into the hype. Figure even staged a “man versus machine” throughput contest where an intern narrowly beat the robots over ten hours—an unintentionally useful reminder that consistency isn’t the same as dexterity. Skeptics are also asking a familiar question in robotics: is it truly autonomous, or is there unseen human help? Without independent verification, that debate will continue. Still, long-duration operation is a real signal: even narrow, repetitive humanoid labor is starting to look plausible in controlled settings.

A cluster of stories today all point to the same reality: as AI spreads through engineering orgs, the hard part isn’t generating code—it’s managing the consequences. One concrete example comes from Buildkite, which updated how it talks about pricing and metered usage, with clearer rules around how CI costs scale with concurrency, test volume, and hosted compute. That kind of transparency matters because teams are running more tests and more automation than ever, and “surprise bills” are a fast way to lose trust in tooling. At the same time, the human side is getting louder. ClickUp says it cut staff significantly while claiming the business is strong, arguing that AI is pushing it to restructure around a smaller number of highly leveraged builders and AI workflow owners. And several essays making the rounds are pushing back on the fantasy of “agentic engineering” as a clean conveyor belt—tickets in, perfect code out overnight. The more sober take: agents can speed up output, but they also amplify weak specs, brittle tests, slow review cycles, and fuzzy decision-making. In other words, AI can make the bottlenecks more obvious—and more urgent to fix.

In health and biotech, there were multiple noteworthy updates—some early, some more mature, all attention-getting. First, researchers shared early human trial results for an experimental therapy aimed at the Parkinson’s-linked LRRK2 pathway. The study was focused on safety and whether the drug hits its biological target, and it reported a sizable reduction in LRRK2 protein levels in cerebrospinal fluid. That’s not the same as proving symptom improvement, but it’s the kind of “target engagement” result that justifies larger trials. Second, AI-driven protein design keeps pushing into drug territory. A team from the University of Washington and Skape Bio reported miniproteins designed to switch major cell receptors on or off—receptors that many medicines target, but which are notoriously tricky to control precisely. The work, published in Nature, suggests a new route to more selective therapies. Third, Eli Lilly says its experimental obesity injection retatrutide produced very large average weight loss in a big trial. The caveat is important: the company has released results before peer review and before seeking approval, and side effects increased at higher doses. Still, if validated, it raises the ceiling for what drug-based weight loss might achieve. And finally, UCLA researchers reported a preclinical “cytokine-armored” CAR-T approach that improved control of glioblastoma in mice, aiming to tackle the long-standing challenge of using CAR-T against solid tumors without unacceptable toxicity. It’s early-stage, but it’s a creative attempt at widening CAR-T’s reach.

In mobility news, Stellantis and China’s Dongfeng announced plans for a new joint venture headquartered in Europe, with Stellantis taking the controlling stake and Dongfeng EV models set to be assembled at a Stellantis plant in France. The big driver here is regulation: the EU’s “Made in Europe” requirements are pushing automakers to localize production to qualify under the rules. For Dongfeng, it’s a pathway into Europe under tighter trade conditions. For Stellantis and French industry, it’s a chance to keep an existing facility busy as the market shifts toward electric. It’s another sign that EV supply chains and partnerships are being redesigned by policy, not just by consumer demand.