Natural quantum spin liquid crystals & Apple accessibility with on-device AI - Hacker News (May 19, 2026)

Let’s start with that science headline. Physicist Aaron Breidenbach says he’s found large amounts of natural herbertsmithite crystals in waste tailings from an abandoned mine in Chile’s Atacama Desert. That mineral is a leading candidate for studying a “quantum spin liquid,” a strange magnetic state that doesn’t settle into a neat pattern even at very low temperatures. Why it matters is less about the treasure-hunt vibe and more about sample quality. A lot of the debate in this field comes down to imperfections—tiny defects that can mimic or mask the signals researchers are looking for. Breidenbach argues these natural crystals may be unusually pure compared to many lab-grown ones, which could make upcoming neutron-scattering experiments cleaner and help settle an ongoing dispute about whether the spin liquid is truly “gapped.” If that claim holds up, it’s not a product tomorrow—but it’s a sturdier foundation for the physics that future devices would depend on.

From fundamental physics to practical computing: Apple previewed a broad set of accessibility updates slated for later in 2026, and Apple Intelligence is at the center of it. The standout theme is richer understanding of what’s on screen and in the camera view—without forcing users to jump through menus. VoiceOver is getting deeper image descriptions and a mode where you can ask natural-language questions about what the camera sees. Magnifier is getting a similar AI boost, plus voice control for common actions. And Voice Control itself is shifting toward “say what you see” interactions, which is a subtle but big deal: it reduces the burden of memorizing exact button names, and it can help compensate for apps that never bothered to label their UI well. Apple also mentioned systemwide subtitles for uncaptioned videos and new reading modes that handle complex layouts, including summaries and translation while keeping formatting intact. The interesting part isn’t that Apple is adding AI—it’s that accessibility is becoming one of the most compelling, socially valuable places to apply it, especially if the on-device processing story holds up in practice.

Staying in AI, there’s a small developer-focused release that’s oddly important if you build LLM tools at scale. A new open-source JavaScript library called id-agent generates token-efficient identifiers—basically, IDs that are cheaper to include in prompts and easier for models to reproduce correctly. The argument is straightforward: UUIDs and other long IDs eat context window space, and models can miscopy them or “helpfully” hallucinate a character. id-agent swaps that for short, human-readable word-based IDs designed to be tokenizer-friendly. It also supports deterministic IDs—so the same input can map to the same identifier—and includes tooling to temporarily replace long IDs in text before you send it to a model, then restore them afterward. Why it matters: as agentic systems grow, prompt hygiene starts looking like performance engineering. If you can save tokens and reduce ID-related mistakes, you get more reliable workflows and more room in the context window for the content that actually matters.

On the broader “where are LLMs headed” question, Simon Willison shared annotated slides from a PyCon US 2026 lightning talk summarizing the last six months of model progress. His big claim is that late 2025 was an inflection point, not just because the “best” frontier model kept changing, but because coding performance made a noticeable jump. He frames it as coding agents moving from “often works” to “mostly works,” thanks to stronger training techniques paired with tool-using setups that let models plan, run actions, and iterate. He also points to the rise of personal assistants—he mentions the fast-growing OpenClaw ecosystem—as both powerful and a little risky, because autonomy amplifies both productivity and mistakes. The other thread: open-weight and local models are getting more competitive, which changes deployment choices. For developers, that’s not just academic. It affects cost, privacy, latency, and whether you can run useful AI features without betting your entire product on a single vendor’s API.

Switching gears to creative tooling: PhotoGIMP is a community-made patch that reshapes GIMP’s interface to feel more familiar to Photoshop users. It rearranges tools, adopts Photoshop-like shortcuts, and tweaks defaults to prioritize canvas space. The key point is what it doesn’t do: it doesn’t fork GIMP or change the core editor. Instead, it swaps configuration—shortcuts, panels, layout, preferences—so a newcomer doesn’t have to unlearn years of muscle memory on day one. That’s valuable because the hardest part of switching tools is often not features; it’s friction. Projects like this lower the barrier to adopting open-source creative software, especially for people who just want to get work done without a week of retraining.

Also in the maker-and-media corner: a creator uploaded a new 3D “splat” scene called “Strawberry” to SuperSplat. It’s a macro photogrammetry capture turned into a Gaussian splatting asset, with unusually detailed documentation about how the source was captured and reconstructed. This matters because Gaussian splats are still young enough that shared, well-documented reference assets are genuinely useful. If you’re researching reconstruction quality, testing rendering pipelines, or comparing workflows, you want something realistic and high-detail—not a toy example with mystery settings. The file is available under a CC BY license, so it’s reusable in projects and experiments, and it gives the community a concrete benchmark for macro-scale capture where small errors are obvious.

For hardware folks: the kv4p HT project is an open-source VHF radio that turns an Android phone into a handheld transceiver through USB-C. It’s aimed at licensed amateur radio use, and it’s positioned as a pocketable, DIY option for off-grid voice and text. What’s notable here is the design philosophy: use the phone for compute, battery, and UI, and keep the radio hardware minimal and open—app, firmware, board designs, and printable parts. That combination makes it both hackable and approachable for tinkerers who want a functional device without an entire RF engineering lab. As always with DIY comms gear, reliability is on you—but for experimentation, emergency kits, and learning, open designs like this keep the ecosystem honest and inventive.

Finally, a sad note from the computing community. Messages on the TUHS and Multics mailing lists report that computer security pioneer Peter Neumann has died, following complications from a fall and surgery. Colleagues describe him as a deeply valued presence, and SRI is expected to host a memorial service in the weeks ahead. Neumann’s influence goes beyond any single paper or project. He helped shape how generations of engineers talk about risk, reliability, and the uncomfortable gap between what systems are supposed to do and what they actually do under real-world pressure. When people like that pass, the loss is personal—but it’s also a reminder that today’s security culture was built by long-term, patient thinkers who kept pushing the field toward rigor.