ChatGPT tackles open math problems & QUIC vs WebRTC for voice AI - Hacker News (May 9, 2026)

First up: a story that’s going to make a lot of researchers pause. Mathematician Timothy Gowers recounts testing ChatGPT 5.5 Pro on open questions in additive number theory, originally raised by Mel Nathanson. With minimal prompting, the model quickly produced a new construction that improves a known bound for a key case, and then extended the approach to a related variant. The bigger twist came when Gowers pushed toward the general case: after iterations and feedback involving MIT student Isaac Rajagopal’s work, the model delivered an argument that Rajagopal believes likely upgrades an exponential bound to something polynomial in k for fixed h. Why this matters: it’s a concrete example of AI plausibly contributing new ideas—like using so-called dissociated sets to imitate geometric-series behavior while keeping numbers from exploding. But the fallout isn’t just mathematical. Gowers raises practical issues: how do you archive results that an AI helped generate, how do you assign credit, and what happens to training pipelines when “entry-level open problems” become rarer—especially if the best models are expensive or gated?

Staying with AI, but shifting to product engineering: a developer critique argues that WebRTC is the wrong transport choice for low-latency voice agents. The core claim is that WebRTC was built for real-time human calls, so under network stress it tends to drop or degrade audio in ways that might be tolerable for conversation—but can be brutal for speech-to-text accuracy and agent reliability. On top of that, the post argues WebRTC is operationally awkward at scale: lots of session setup, tricky routing patterns, and workarounds that can create fragility when clients move across networks. The proposed direction is to lean into QUIC-based approaches—think WebTransport or Media over QUIC—because QUIC is designed around faster connection setup and more resilient connections even when NAT mappings change. Why it matters: voice AI is quickly becoming a mainstream UI, and the plumbing choices teams make now will shape reliability, cost, and global scaling later.

Now to the web’s quiet gatekeeper: reCAPTCHA. Reports say Google’s newer reCAPTCHA verification flow on Android has effectively been tied to Google Play Services. In suspicious cases, the challenge can turn into a QR-code scan flow that expects Play Services running and up to date in the background. For users on de-Googled phones or custom ROMs—GrapheneOS is the example that keeps coming up—this can mean you simply can’t pass the check. Why it matters: reCAPTCHA isn’t a niche login feature; it’s a choke point for huge parts of the web. If passing basic human verification increasingly requires Google’s proprietary stack on Android, that’s ecosystem lock-in by another name—especially notable when the iOS path reportedly works without installing extra Google components.

On the preservation front, Internet Archive Switzerland has launched as an independent non-profit foundation based in St. Gallen, with the familiar goal of “universal access to all knowledge.” The pitch is straightforward: digital information is fragile—formats rot, storage fails, content gets deleted, and paywalls narrow what people can learn. What’s new is the early focus. One initiative is a ‘Gen AI Archive’ with the University of St. Gallen, aimed at preserving today’s generative models for future research. Another is an ‘Endangered Archives’ effort focused on vulnerable cultural and historical materials threatened by conflict, disasters, or suppression, working with partners including UNESCO. Why it matters: preservation isn’t just about old web pages anymore. If AI models shape science, policy, and culture, then archiving them—and the context around them—becomes part of keeping the public record intact.

A quick software-engineering reset next. Martin Fowler revisits The Mythical Man-Month, Fred Brooks’s classic on managing large software projects. Fowler’s point isn’t nostalgia; it’s that the hard parts didn’t disappear. Brooks’s Law—the idea that adding people to a late project makes it later—still bites because communication overhead grows faster than teams expect. But Fowler emphasizes an even more durable idea: conceptual integrity. A system with a coherent design tends to age better than one that just accumulates features. Why it matters in 2026: with AI-assisted coding and faster scaffolding, teams can produce more code than ever, but the constraint is still clarity—what the system is, and what it refuses to become.

Related, but down at the performance layer: a BYU FLOW Lab post shows Julia approaching C++ performance on a compute-heavy numerical kernel. The headline isn’t “Julia is slow” or “Julia is fast.” It’s that you can absolutely get near C++ speed, but you often have to write in a way that avoids hidden allocations, keeps types predictable, and treats memory layout like a first-class design decision. Why it matters: more teams are using high-level languages for scientific computing and simulation, and the trade-off isn’t just runtime. It’s maintainability versus mechanical sympathy—how much you need to think like the compiler to hit your targets.

Let’s jump to science for a minute. A Quanta feature argues lightning research is being reshaped by instruments that can peer into storm clouds—and the results keep undermining the simplest textbook story. Measurements often show storm electric fields that look too weak to kick off a spark in ordinary air. So researchers have been building a more complex picture, where high-energy particles help close the gap. Recent observations include storms emitting X-rays and gamma rays, and NASA’s 2023 ALOFT campaign finding frequent gamma activity even without obvious lightning. Another thread: radio measurements suggest some lightning initiates in ways that don’t neatly align with the local electric field, reviving the idea that cosmic rays might sometimes provide the initial ionization. Why it matters: lightning isn’t just a curiosity—it impacts aviation safety, wildfire risk, power infrastructure, and atmospheric chemistry. Understanding triggers could improve forecasting and hazard modeling.

Now, a major public-health and accountability story. An investigation by the AP, The Atlanta Journal-Constitution, and FRONTLINE reports extensive PFAS contamination—so-called “forever chemicals”—in the Conasauga River system tied to decades of stain-resistant carpet manufacturing in northwest Georgia. The reporting describes how wastewater and sludge containing PFAS moved through systems not designed to remove these chemicals, and how land application spread contamination across large areas, with downstream communities reporting elevated PFAS levels and worrying health outcomes. Why it matters: PFAS is a case study in how persistent chemicals, weak oversight, and long corporate timelines can externalize costs to the public for generations. Even when lawsuits produce big numbers, the hard problem remains: cleanup, long-term exposure, and the fact that the science and testing burdens often land on residents first.

Finally, something more practical for everyday tech: an updated explainer on Wi‑Fi generations argues that the way routers are marketed leads people to chase the wrong upgrades. The takeaway is that real-world speed usually isn’t limited by the router’s headline number—it’s limited by client devices, interference, distance, and how your home network is designed. It also reframes the usual ‘buy the newest standard’ instinct. You may get more from better placement, wired backhaul for access points, or simply reducing congestion than from swapping one shiny box for another. Why it matters: Wi‑Fi has become critical household infrastructure, and consumers deserve realistic expectations—especially as more devices compete for the same airwaves.