Published on July 11, 2026 at 06:00 CEST (UTC+2)
Einstein's relativity rules chemical bonds in heavy elements, new research shows (142 points by hhs)
Einstein's relativity rules chemical bonds in heavy elements, new research shows
Brown University chemists have provided the first direct experimental evidence that Einstein’s theory of relativity alters the traditional triple-bond structure in heavy elements like bismuth. Using photoelectron spectroscopy, they showed that the textbook model of triple bonds—distinct sigma and pi bonds—breaks down when atomic nuclei are sufficiently heavy. This overturns a long-held high-school chemistry concept and confirms theoretical predictions from the 1970s. The findings have implications for understanding the chemistry of heavy elements, which is relevant to materials science and nuclear chemistry.
QuadRF can spot drones and see WiFi through my wall (492 points by speckx)
QuadRF can spot drones and see WiFi through my wall
Jeff Geerling highlights the QuadRF, an open-source phased-array radio built on a Raspberry Pi 5 and FPGA that can detect drones and intercept WiFi signals through walls using advanced beamforming and signal processing. The device can stream and decode RF data, exposing network vulnerabilities that were previously only accessible to governments. The project is part of a larger ambition to build a Moon-scale antenna array for Earth-Moon-Earth radio experiments. Geerling argues that such tools should be public to encourage better security practices rather than banned.
Apple sues OpenAI, accuses ex-employees of stealing trade secrets (699 points by stock_toaster)
Apple sues OpenAI, accuses ex-employees of stealing trade secrets
Apple has filed a lawsuit against OpenAI, alleging that former Apple employees—including ex-VP of product design Tang Tan—stole confidential information about unreleased technologies for OpenAI’s benefit. The lawsuit names specific individuals and claims significant evidence of trade secret theft related to Apple’s hardware and software innovations. OpenAI has not yet publicly responded beyond an initial statement. This case underscores the intensifying legal battles between Big Tech and AI startups over intellectual property and talent poaching.
An iroh powered smart fan (46 points by surprisetalk)
An iroh powered smart fan
This blog post demonstrates how to build a cloud-free smart fan using an ESP32 microcontroller and the Iroh peer-to-peer networking library. The fan measures temperature and can be controlled remotely via a WebAssembly-based website, with no central server required. The project showcases how Iroh enables direct device-to-device communication, making IoT devices more private and resilient. It is a practical example of decentralized edge computing for home automation.
An update on residential proxies and the scraper situation (125 points by chmaynard)
An update on residential proxies and the scraper situation
LWN.net reports on the ongoing and worsening problem of web scraping for AI training data. Attackers use residential proxy networks to disguise bot traffic as legitimate human visits from millions of unique IPs, making it nearly impossible to block effectively. The article describes methods to detect scrapers (e.g., missing CSS/image fetches) but notes that by the time detection occurs, the IP is already discarded. The open web is increasingly strained under this load, raising questions about sustainability and the ethics of data collection for AI.
GPT-5.6 Sol Ultra produces proof of the Cycle Double Cover Conjecture [pdf] (385 points by scrlk)
GPT-5.6 Sol Ultra produces proof of the Cycle Double Cover Conjecture [pdf]
OpenAI’s latest model, GPT-5.6 Sol Ultra, has purportedly generated a full proof of the Cycle Double Cover Conjecture, a long-standing open problem in graph theory. The PDF link contains the proof, which is a significant milestone in AI-driven mathematical reasoning. If verified, this would demonstrate that large language models can now tackle high-level theoretical mathematics that has eluded human mathematicians for decades. The result could accelerate AI’s role in scientific discovery and formal verification.
SpaceX wants to launch 100k more Starlink satellites for 100x the bandwidth (112 points by CrankyBear)
SpaceX wants to launch 100k more Starlink satellites for 100x the bandwidth
SpaceX has applied to the FCC to launch up to 100,000 third-generation Starlink satellites, promising “ultra-low-latency” multi-gigabit symmetrical broadband for rural users. Current real-world speeds are far lower (around 145–170 Mbps down), but the new constellation aims to dramatically increase capacity. Critics raise concerns about space debris, light pollution, and regulatory hurdles. For AI/ML, such a network could provide high-bandwidth internet anywhere on Earth, enabling distributed AI inference and edge computing at a global scale.
Inference Optimization for MiMo v2.5: Pushing Hybrid SWA Efficiency to the Limit (55 points by theanonymousone)
Inference Optimization for MiMo v2.5: Pushing Hybrid SWA Efficiency to the Limit
Xiaomi’s MiMo-V2.5 and V2.5-Pro models use a Hybrid Sliding Window Attention (Hybrid SWA) mechanism that compresses KV cache to about 1/7 of full attention, combined with sparse MoE activation and multimodal encoders. The blog details the full-pipeline inference optimization that makes long-context and multimodal reasoning efficient. This is a practical guide for deploying resource-efficient large models on limited hardware, balancing reasoning quality with compute cost. It represents a key trend in making advanced AI models more accessible for real-world applications.
The tech of 'Terminator 2' – an oral history (2017) (189 points by markus_zhang)
The tech of 'Terminator 2' – an oral history (2017)
This oral history recounts how Industrial Light & Magic (ILM) invented new CGI tools—like “Make Sticky” and “Body Sock”—to create the iconic liquid-metal T-1000 in Terminator 2. The article features interviews with the small team of artists and engineers who pushed the limits of early 1990s computer graphics. It highlights the iterative, collaborative process behind groundbreaking visual effects. While not directly about modern AI, it offers a historical perspective on how creative problem-solving drives technological innovation.
Good Tools Are Invisible (380 points by theanonymousone)
Good Tools Are Invisible
GingerBill argues that the best tools—like text editors or programming languages—should fade into the background, allowing the user to focus on the task at hand. He criticizes the tendency to celebrate a tool’s limitations as “fun puzzles,” using vim as an example where users spend excessive time on macros instead of simpler alternatives. The piece advocates for tooling that is efficient, intuitive, and unobtrusive, warning against cult-like devotion to complex workflows. This philosophy is highly relevant to the design of AI-assisted developers tools and user interfaces.
AI as a mathematical theorem prover reaches a milestone
The claim that GPT-5.6 Sol Ultra produced a proof of the Cycle Double Cover Conjecture shows AI is moving beyond pattern matching into formal mathematical reasoning. If verified, this could accelerate discovery in combinatorics and other fields, and may eventually lead to AI systems that generate and verify proofs autonomously. Implication: Expect more AI-driven breakthroughs in mathematics and science, and a growing need for verification frameworks to trust AI-generated proofs.
Legal battles over IP and talent are reshaping the AI landscape
Apple’s lawsuit against OpenAI is part of a broader trend where tech giants use litigation to protect trade secrets and curb AI startups from poaching key personnel. Such cases will likely increase as AI talent becomes scarce and proprietary data (e.g., training datasets, model architectures) becomes a competitive advantage. Implication: Startups must invest in clear IP policies and legal safeguards, while large companies may leverage lawsuits to slow down competitors and control the talent flow.
Inference efficiency is the new frontier for scaling AI
Xiaomi’s MiMo v2.5 optimization report illustrates a major trend: making large models practical for deployment. Techniques like Hybrid Sliding Window Attention, sparse MoE, and multimodal encoders reduce compute and memory requirements while preserving quality. This is critical for running AI on edge devices, reducing cloud costs, and enabling real-time applications. Implication: Future AI progress will depend as much on efficient inference engines as on raw model size; expect more hardware–software co-design (e.g., custom ASICs, optimized kernels) to dominate the field.
Data scraping for AI training is creating an unsustainable web
The continued rise of residential proxy-based scraping, as described in the LWN article, threatens the open web’s viability. Sites are being hammered by bots that mimic humans, making it harder to maintain free content and driving up costs. This is leading to more aggressive anti-bot measures, paywalls, and legal actions. Implication: The AI industry must shift toward legitimate data partnerships, synthetic data generation, or federated learning to avoid a “data famine” and regulatory backlash.
Edge AI and decentralized computing are becoming practical
The Iroh-powered smart fan and QuadRF projects demonstrate that peer-to-peer networking and edge compute can work without cloud dependency. Combined with SpaceX’s Starlink expansion, this points to a future where AI inference and IoT control happen locally or via low-latency satellite links, reducing reliance on centralized data centers. Implication: Developers should explore frameworks that support offline-first, peer-to-peer architectures (e.g., Iroh, IPFS) to build resilient AI applications for remote or privacy-sensitive environments.
The tooling philosophy “good tools are invisible” applies directly to AI assistants
GingerBill’s critique of overhyped, friction-filled tools resonates with the design of AI coding assistants and chatbots. Many current AI tools add cognitive overhead—requiring prompt engineering, output validation, or context switching—rather than disappearing into the workflow. Implication: The next wave of AI tooling must prioritize seamless integration and minimal user friction, possibly through agentic systems that anticipate intent or via voice/gesture interfaces that reduce explicit prompting.
Connectivity infrastructure is being supercharged for AI at the edge
SpaceX’s plan for 100,000 Gen3 Starlink satellites aims to provide multi-gigabit speeds globally, which would enable distributed AI inference, real-time video analytics, and large model fine-tuning from remote locations. Combined with open-source hardware like the QuadRF for RF sensing, this infrastructure could support autonomous drones, precision agriculture, and disaster response. Implication: AI applications that were previously limited to urban data centers can now be deployed anywhere, but this also raises concerns about security, spectrum allocation, and space debris that need coordinated governance.
Analysis generated by deepseek-reasoner