HackerNews Digest

Brown University chemists provided direct spectroscopic evidence that relativistic effects alter triple‑bond structures in heavy elements. Using photoelectron spectroscopy on carbon‑bismuth molecules cooled near absolute zero, they observed binding energies inconsistent with the conventional sigma + 2 pi model. Instead, the data indicate a bond arrangement of one pi bond and two hybrid sigma‑pi bonds, reflecting spin‑orbit coupling that merges sigma and pi character when electrons move at relativistic speeds in heavy nuclei. This demonstrates that the textbook sigma/π separation breaks down for elements such as bismuth, where nuclear mass induces significant relativistic behavior. The findings suggest revisions to chemical‑bonding curricula and highlight potential applications of bismuth in lead‑free solar cells and quantum‑material research. The work was supported by NSF (CHE‑2403841) and the U.S. Department of Energy (DE‑SC0008501).

QuadRF is a handheld phased‑array radio built around a Raspberry Pi 5 and an FPGA board, offering picosecond‑level timing and beamforming in the 4.9‑6 GHz band. Its open‑source firmware streams raw I/Q data over the Pi 5’s camera/display MIPI connectors at >5 Gbps, enabling low‑latency, full‑duplex SDR without USB bottlenecks and allowing multiple modules to be daisy‑chained for distributed phase control. The system boots a Wi‑Fi hotspot and provides a web‑based VNC interface to launch GNU Radio, SDR tools, and a custom augmented‑reality visualizer that maps detected signals as colored blobs (e.g., 5 GHz Wi‑Fi appears light‑blue). A Mobile Expansion Pack adds battery power and a handheld mount for real‑time C‑band analysis. In field tests the unit tracked a DJI Mini Pro 4 drone and visualized nearby Wi‑Fi networks; gain adjustment was manual, and the UI remains rough. A larger “Moon‑scale” version aims for up to 1.15 MW EIRP for Earth‑Moon‑Earth experiments. Kits cost $499, with production still in pre‑release.

Apple has filed a lawsuit in the U.S. District Court for the Northern District of California accusing OpenAI of trade‑secret theft by former Apple employees. The complaint names former Apple VP of product design Tang Tan and former senior system electrical engineer Chang Liu as defendants, alongside OpenAI and its acquired startup io Products. Apple alleges Tan used Apple’s confidential project codenames and directed interview candidates to bring actual Apple hardware, CAD files, and prototype components to OpenAI interviews, extracting detailed design, subsystem, vendor, and manufacturing information. Liu is accused of exploiting a security bug to download a 1,000‑page technical file set, failing to return an Apple laptop, and coaching another recruit on confidential materials. Apple also claims OpenAI misled an Apple partner into performing Apple’s proprietary metal‑finishing process and queried a long‑time Apple supplier about battery components. The suit seeks injunctive relief and damages, noting over 400 former Apple staff now work at OpenAI.

The guide shows how to build an iroh‑powered smart fan with an ESP32‑WROVER (or compatible) devkit. Starting from the iroh‑esp32‑examples echo server, the firmware is flashed via `cargo run --release` (optionally speeding the flash with `ESPFLASH_BAUD`). Wi‑Fi credentials are supplied via `WIFI_CONFIG`. The device stores a secret key in non‑volatile memory to generate a stable endpoint ID, then registers with the nearest iroh relay using QUIC. A DHT22 sensor is wired to 3.3 V, GND and GPIO 26; readings are logged and exposed via an irpc protocol defined in a separate `smart-fan-proto` crate (`GetLatest`, `GetStatus`, `SetThreshold`). The server holds the latest reading in an `Arc>>` and controls a 5 V PWM fan (or LED/relay) on another GPIO, using hysteresis to avoid rapid toggling. A WebAssembly GUI runs in a browser, accepts a ticket, displays temperature, humidity, fan state, and (with a secret) allows threshold adjustment. The project includes a 3‑D‑printed enclosure, STL files, and full source in the `iroh-smart-fan` repository.

Unicode includes decorative characters such as the Manichaean Punctuation Fleuron (U+10AB1). Because this fleuron is classified as a right‑to‑left (R) strong type, it influences the layout of surrounding weak characters, causing numbers that follow it to appear before it (e.g., “𐫱400” renders as “400𐫱”). This behavior follows the Unicode Bidirectional Algorithm, which assigns each character a bidi type (L, R, EN, WS, etc.) and resolves ordering based on strong types, with weak types inheriting direction from nearby strong characters. To prevent unwanted reordering, the fleuron can be isolated using the HTML  element or CSS `unicode-bidi: bidi-override;`, which forces the character to retain its intended position regardless of surrounding text direction. These methods work across modern browsers, though legacy support (e.g., IE) is limited.

The article details the ongoing wave of web‑scraper attacks that leverage residential and mobile IP addresses. Attackers control compromised devices—often via malware on PCs, media‑streaming hardware, or VPN‑based services such as Bright Data—to issue millions of short, human‑like requests that rarely load images or CSS, making IP‑based blocking ineffective. The shutdown of Google’s IPIDEA botnet temporarily reduced traffic, but new networks (e.g., NetNut) have revived the threat. Operators range from outright criminals to companies advertising “ethically sourced” proxy pools, and some AI model developers use their own scrapers with identifiable user‑agents while adhering to robots.txt. Defenses employed by sites include proof‑of‑work systems (Anubis), commercial human‑verification services, login or paywall gating, and data‑poisoning tools like iocaine. LWN has optimized its infrastructure and selective throttling to mitigate impact without hindering legitimate crawlers (search engines, Internet Archive). Recent coordinated takedowns by Google and law‑enforcement have offered brief relief, but the article warns that the arms race will persist.

None

SpaceX has filed an FCC application to launch up to 100,000 third‑generation (Gen3) Starlink satellites, expanding the current ~11,000‑satellite constellation. Each Gen3 satellite would weigh over 2 tonnes, requiring launch on Starship (or, in the interim, Falcon Heavy) rather than Falcon 9. The Gen3 system is pitched as a “ultra‑low‑latency” network delivering multi‑gigabit symmetrical broadband with latency under 20 ms and a claimed 100‑fold increase in total bandwidth. The filing requests an unusually wide spectrum allocation, spanning Ku, Ka, V, E, W and D bands, with downlink frequencies of 10.7‑13.4 GHz, 17.3‑21.2 GHz and 37.5‑42.5 GHz and uplink bands up to ~275 GHz, along with waivers to assemble larger contiguous channels. New user terminals and antennas would be needed to exploit the higher speeds. Pricing for Gen3 is not set but is expected to exceed the current $130/month residential plan, possibly $200‑$300/month. Competitors include Amazon Leo, Eutelsat‑OneWeb, Telesat Lightspeed, Blue Origin’s TeraWave, and legacy GEO providers such as HughesNet and ViaSat. The proposal will undergo FCC review, public comment, and possible conditions on debris mitigation, spectrum coordination, and interference, amid concerns from astronomers about the impact of large satellite constellations.

The oral history details ILM’s small but rapidly expanding CG team during Terminator 2 (1991), focusing on the creation of the liquid‑metal T‑1000 effects. Key figures—including Dennis Muren, technical directors like Stefen Fangmeier and John Schlag, and software developers Eric Enderton and Michael Natkin—described how they repurposed tools from The Abyss (e.g., the monolithic pseudopod system) into modular utilities such as “Make Sticky” and “Body Sock.” Development relied on SGI workstations (240/340 VGX) and custom batch‑processing scripts written in C‑shell; a processor‑allocator GUI (PA) was built to share render farm resources before commercial solutions existed. Early animation packages included Alias and Wavefront, with ILM ultimately favoring Alias for its usability. Artists were expected to perform multiple roles—modeling, animating, texturing, lighting, rendering, and compositing—on each shot, reflecting a DIY workflow. The department grew from ~12 to ~300 staff, investing millions in hardware to deliver roughly 50 CG shots over six months, establishing a precedent for modern digital characters.

The MiMo‑V2.5 and MiMo‑V2.5‑Pro models use Hybrid Sliding Window Attention (Hybrid SWA) and sparse Mixture‑of‑Experts (MoE) to cut KV‑cache storage and compute to ~1/7 of full‑attention models, enabling efficient long‑context and multimodal inference. Key engineering advances: - **KV‑cache redesign**: Dual‑pool KV system separates Full‑Attention (O(N)) and SWA (O(W)) storage, tiered across GPU, host, and L3 caches; layer‑wise prefetch and a SWA‑aware prefix tree enforce “window‑safe” hit matching, preventing invalid KV reuse. - **GCache**: A distributed, consistent‑hashing cache with memory/disk tiers, RDMA‑optimized I/O, and co‑deployment on GPU nodes provides low‑cost, high‑throughput L3 KV storage and robust fault handling. - **Scheduling**: Stateless LLM‑Router uses prefix affinity and load‑aware scoring to boost L2 hit rates (~25 %) and throughput (~30 %); TTFT prioritization reduces P90 latency for long requests by up to 30 % without starving short jobs. - **Prefill/Decode optimizations**: Reduced expert parallelism, three‑tier length bucketing, and MTP support cut prefill time (~40 %); decode KV expansion, PD‑disaggregated preallocation, and CUDA‑graph tuning raise KV capacity ~5×. - **Multimodal pipeline**: Overlapped embedding transfers, GPU‑based preprocessing, cross‑request batching, and consistent‑hash encoder routing improve encoder throughput and cache hit rates (≈30 % increase). These combined changes realize the theoretical efficiency of Hybrid SWA in production, delivering higher throughput, lower memory use, and reduced API costs.