Published on July 14, 2026 at 06:00 CEST (UTC+2)
Japan develops a method to recover up to 90% of lithium from used EV batteries (74 points by donohoe)
Japanese scientists have developed a new method to recover up to 90% of lithium from used EV batteries, a significant improvement over traditional methods that recover less than 50%. The breakthrough could dramatically improve battery recycling efficiency and reduce waste as electric vehicles proliferate. If scalable, this technology has the potential to reshape how EV batteries are manufactured and reused, addressing a key sustainability challenge.
Fundamentals of Wireless Communication (49 points by teleforce)
This is a textbook on wireless communication fundamentals by David Tse and Pramod Viswanath, published by Cambridge University Press in 2005. The online version provides free PDF chapters covering topics from wireless channels and MIMO to multiuser capacity and detection theory. It is widely used as a graduate-level reference for physical-layer wireless communication theory.
The Git history command deserves more attention (116 points by turbocon)
The article highlights the experimental git history command (available in Git 2.54 and 2.55) which provides three subcommands: fixup, reword, and split. These allow users to edit old commits and automatically rebase branches, offering some of the benefits of the jj version control system without switching tools. The author argues this command deserves more community attention as it simplifies parallel branch management while staying within core Git.
What will be left for us to work on? (58 points by randomwalker)
This is a transcript and slides from Arvind Narayanan's keynote at ICML 2026, addressing anxiety about AI capability growth and future work. He argues that the "AI as normal technology" framework is useful unless recursive self-improvement occurs, but no single lab milestone will suddenly eliminate jobs. Instead, future jobs will require adaptation, and he proposes a vision of human-AI "co-superintelligence."
Building and shipping Mac and iOS apps without opening Xcode (352 points by speckx)
The author describes how to build and ship Mac and iOS apps entirely from the command line, never opening Xcode's GUI. Using tools like xcodebuild, notarytool, and stapler, along with LLM-based coding assistants (e.g., Claude Code), developers can automate the full build, signing, notarization, and deployment pipeline. One-time setup steps (Apple ID sign-in, certificate creation) are the only GUI-required tasks.
The Economics of Recursive Self-Improvement [pdf] (30 points by apsec112)
This PDF paper explores the economics of recursive self-improvement in AI systems, likely analyzing scenarios where AI can autonomously improve its own capabilities. It examines the economic incentives, potential for rapid acceleration, and implications for market dynamics and policy. The content is technical and focused on future AI development trajectories.
Apple's new SpeechAnalyzer API, benchmarked against Whisper and its predecessor (488 points by get-inscribe)
Apple's new SpeechAnalyzer API, introduced in iOS/macOS 26, was benchmarked against Whisper models and the legacy SFSpeechRecognizer. On the LibriSpeech dataset, SpeechAnalyzer achieved a test-clean word error rate of 2.12% (versus Whisper Small's 3.74%) while running roughly three times faster. It outperformed all tested Whisper variants and is the most accurate on-device speech engine tested.
An Englishwoman who sketched India before photography took hold (76 points by 1659447091)
This BBC article profiles Emily Eden, an Englishwoman who sketched people and scenes across India in the 19th century, before photography became widespread. Her detailed drawings captured the dress, jewelry, and daily life of diverse communities, offering a unique visual record of an empire in transition. The article highlights her curiosity and precision as an artist.
Is x86 ready to ACE it? (21 points by mfiguiere)
This technical article examines the x86 Ecosystem Advisory Group's new ACE accelerator specification, which adds a second accelerator type to Intel's AMX (Advanced Matrix Extensions) instruction set. ACE is designed to accelerate matrix multiplication for machine learning workloads, complementing the existing TMUL accelerator. The piece discusses how x86 is evolving to better handle modern AI inference and training tasks.
MorphoHDL: A minimalistic language for growing circuits (33 points by jacktang)
MorphoHDL is described as a minimalistic domain-specific language for hardware description, focused on "growing circuits" through iterative exploration. The article appears to present an interactive tool for hardware design that allows designers to experiment with circuit structures. It likely targets custom accelerator design and educational use.
1. On-Device Speech Recognition Reaches Parity with Cloud Models
Apple's SpeechAnalyzer API now beats Whisper Small in both accuracy and speed on a standard benchmark (LibriSpeech), running locally on an M2 Pro. This marks a tipping point where on-device models can match or exceed cloud-based alternatives for key tasks. Implications: edge AI is becoming viable for high-accuracy applications, reducing latency and privacy concerns. Developers should reassume on-device-first architectures for speech interfaces.
2. Hardware Accelerators for ML Continue to Specialize
Intel's AMX extension is evolving with a second accelerator (ACE) for matrix operations, alongside the existing TMUL. This reflects the industry's push to build dedicated silicon for AI workloads directly into general-purpose CPUs, not just GPUs. The trend means that ML inference will become more efficient on commodity hardware, lowering the barrier for deploying models without dedicated accelerators. Developers should watch for new ISA extensions and optimize for them.
3. The Economics of Recursive Self-Improvement Becomes a Policy Focus
The PDF paper on recursive self-improvement economics signals growing academic attention on scenarios where AI systems can autonomously improve themselves. This is a core concern in AI safety and governance, especially around risks of rapid capability jumps. The implication is that research on alignment, control, and economic disruption will intensify. Practitioners should consider how self-improving systems might affect their own workflows and the broader job market.
4. AI Anxiety and the Reimagining of Human Work
Arvind Narayanan's ICML keynote directly addresses the widespread fear among AI professionals that their work will be automated. His argument for "human-AI co-superintelligence" reframes the narrative from replacement to augmentation. This trend suggests that the AI community needs to actively shape the future of work through retraining, new interface paradigms, and ethical guidelines. Actionable takeaway: invest in skills that combine AI tool usage with domain expertise.
5. Vibe Coding and LLM-Assisted Development Reshape Software Tooling
The article on building Mac/iOS apps without Xcode explicitly leverages LLM coding assistants (Claude Code) to automate build configurations, while another article discusses using git history to simplify version control. This reflects a broader shift where developers use AI to scaffold and automate tedious parts of the development lifecycle. The implication is that traditional IDEs and manual scripting may decline as AI agents handle more boilerplate. Teams should adopt LLM-friendly CI/CD pipelines.
6. Sustainability Considerations Enter AI Hardware Discussions
The lithium recovery breakthrough is not directly AI/ML, but the massive energy and material demands of AI training/inference hardware (GPUs, battery-backed data centers) make battery recycling increasingly relevant to the AI supply chain. As AI scales, sustainable sourcing of materials—including for energy storage in renewable-powered data centers—will become a strategic concern. This trend points to an intersection of AI hardware and circular economy innovations.
7. Benchmarks and Transparency Demands in AI APIs
The independent benchmark of Apple's Speech API against Whisper, conducted because Apple published no accuracy figures, highlights a growing demand for third-party evaluation of proprietary AI services. Developers need transparent, reproducible benchmarks to make informed decisions. The implication is that companies that fail to publish performance data risk losing trust, while open benchmarks will drive competition. Practitioners should always run their own evaluations against representative data.
Analysis generated by deepseek-reasoner