Subscribe
Share
Share
Embed
Daily AI news and research, distilled. UpNext AI breaks down the most important developments in artificial intelligence—from major industry moves to cutting-edge papers.
Welcome to the UpNext AI podcast. It's Tuesday, May 19th, 2026, and here's what matters in AI today.\n\nFirst up, the biggest practical story today is infrastructure.\n\nThe Financial Times reports Google is making a chip push with a Blackstone-backed AI cloud group, tied to a $5 billion investment that is set to help bring 500 megawatts of data-center capacity online next year. Even with the details here still fairly high level, the signal is clear: the AI race is still being fought not just in models and products, but in power, hardware, and physical capacity. Five hundred megawatts is the kind of number that reminds you this is now an industrial buildout story as much as a software story. For Google, it also suggests the chip strategy is being tied directly to where inference and training actually happen: in large-scale compute environments that can come online fast enough to meet demand.\n\nFrom there, to the legal beat.\n\nArs Technica reports Elon Musk lost his lawsuit against OpenAI, with a nine-person jury unanimously deciding that he waited too long to file the case. The suit accused Sam Altman and OpenAI of effectively betraying the nonprofit mission Musk says he helped fund with $38 million. According to the reporting, the jury found Musk was aware of OpenAI’s restructuring plans as early as 2021, which meant he missed the statute of limitations. The jury also found no liability for Microsoft, which Musk had argued aided OpenAI’s shift toward a for-profit structure. Musk plans to appeal. The immediate takeaway is straightforward: one of the highest-profile legal attacks on OpenAI has, for now, failed on timing grounds rather than on a broader re-litigation of OpenAI’s mission.\n\nNow to the research section.\n\nA paper in npj Digital Medicine looks at a problem that matters a lot in healthcare AI: how do you evaluate language models without burning huge amounts of time and money on giant static benchmarks? The researchers argue that computerized adaptive testing — basically the same family of ideas used in some standardized exams — can do the job much more efficiently. Instead of giving every model the full bank of questions, the test adapts based on how the model is performing, choosing the next items to measure ability more efficiently.\n\nIn their study of 38 language models, the adaptive approach achieved a near-perfect correlation with full-benchmark results while using only 1.3 percent of the items. Evaluation time fell from 6.85 hours to 8.4 minutes per model. Token usage dropped from 1.77 million to 0.03 million. And at current API pricing, the cost went from about $1,475 to under $5 per model, while preserving model rankings.\n\nThe researchers are also clear about what this is not. It’s not a substitute for clinical validation or prospective safety studies in the real world. But as a screening and monitoring tool, it looks potentially very useful. Bottom line: if this holds up, teams working in regulated domains may be able to evaluate models far more often, and far more cheaply, without giving up much measurement quality.\n\n...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. In fact, you're listening to one of their voices - right - now. If you are a developer looking to elevate user experience with natural voice interfaces, this is your solution. Visit up next dot fm slash eleven to check out their latest offerings. ...\n\nForbes takes a broad look at five benefits and risks of using AI for cybersecurity. The framing is familiar but still useful: AI can help automate monitoring, anomaly detection, incident response, and routine analysis, while also creating fresh vulnerabilities and a larger attack surface. The practical point is that AI in security is increasingly a double-edged tool — valuable, but not neutral.\n\nAnd Forbes also reports that Anthropic has bought Stainless, the SDK tooling company used to turn APIs into production-ready software development kits across multiple languages. In that reporting, the notable wrinkle is that Anthropic is shutting down the hosted product for other companies, while existing customers keep what they’ve already generated. If that holds, this is less a flashy model story and more an infrastructure control story: owning a piece of the plumbing that helps developers actually connect AI systems to the rest of their stack.\n\nBefore we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers.\n\nIf you enjoyed this episode, don't forget to subscribe, rate, and leave us a review! And that's your briefing for today. Full source links are in the episode notes, and we'll be back tomorrow with what's up next!