CATL signed a three-year strategic cooperation to deliver 60 GWh of sodium-ion batteries to HyperStrong, the largest such order to date. The deal covers technology development, product application, and project delivery and is linked to HyperStrong's broader plan to procure 200 GWh from 2026 to 2035. CATL says the arrangement shows it has mature, large-scale delivery capabilities for sodium-ion cells and has overcome key manufacturing challenges.
New Stories
The article references Fedora 44 and describes the Fedora desktop as familiar, with an added layer of security and reliability. It mentions Atomic Desktops and invites readers to Learn More.
Amazon says it will offer OpenAI models on Amazon Bedrock in limited preview, including the latest OpenAI models and Codex, along with Bedrock Managed Agents powered by OpenAI. The offerings mark a broader expansion of the AWS-OpenAI partnership, providing access to frontier models through Bedrock APIs with existing security, governance, and cost controls. Bedrock Managed Agents enable production-ready AI agents on AWS, while Codex on Bedrock supports enterprise software development within AWS environments.
Ars Technica reports that former FBI director James Comey was indicted in North Carolina on charges that he publicly threatened the life of the President, based on a 2025 Instagram post featuring seashells arranged to read ‘86 47.’ The article notes a prior Virginia indictment against Comey that was dismissed and includes critical commentary about DOJ actions regarding the case.
King Charles III addressed the U.S. Congress during his state visit, stressing the importance of a united U.K.-U.S. partnership amid global uncertainty and highlighting NATO ties. He invoked Magna Carta-style checks and balances to please Democrats and mixed formal remarks with light humor, without mentioning Jeffrey Epstein or his victims. President Trump called the King a “fantastic person” and appeared pleased with the visit.
Guinea worm disease is nearing eradication, with human cases reduced from an estimated 3.5 million in 1986 to about 10 in 2025, a drop of more than 99.99%. The campaign relies on education, behavior change, safe drinking-water practices, surveillance, and partnerships with governments, focusing on six African countries: Angola, Cameroon, Chad, Ethiopia, Mali, and South Sudan. Remaining challenges include animal infections, particularly in dogs in Chad, and security-related access issues that complicate monitoring.
The material covers vLLM-compile, a project that integrates Torch's torch.compile with the vLLM inference engine to automate kernel generation and graph-level optimizations for faster, more portable LLM execution. It explains piecewise CUDA graphs and custom fusion/transformation passes—such as Attention with Quant FP8 or FP4, RMSNorm with Quant, and SiLU-Mul with Quant—aimed at improving end-to-end performance, with reported improvements up to 11% in some cases, though not all operators are supported by CUDAGraphs. Additional goals include reducing startup and compilation time and applying optimizations independently of model definitions, as well as exploring vLLM IR to separate semantics from implementation.
RecipeScape is an interactive tool for analyzing cooking instructions at scale, offering visual components to cluster, compare, and analyze recipes for a given dish. It uses a computational pipeline that converts recipe text into a tree representation via POS tagging and human annotation, calculates weighted tree-edit distances guided by word embeddings, and builds a distance matrix for hierarchical clustering. Data are collected by crawling schema.org Recipe entries from recipe sites, and a study with 500 recipes across two dishes shows the system can identify distinct procedural approaches and usage patterns of ingredients and actions.
pman is a command-line tool that initializes and manages project directories, generating boilerplate files such as .gitignore, LICENSE, and README while creating language-specific entry points. It maintains a central registry of registered projects to support auditing, listing, pruning, and exporting, and supports custom templates via scripts placed in ~/.config/pman/templates/.
Researchers at the University of Oregon repurposed a potentiostat, a tool used to test batteries, to measure coffee flavor by applying electricity to a brew. They focused on roast color and beverage strength, tested four coffee samples, and could identify a batch that failed quality control. The findings were published in Nature Communications in 2026 and aim to help baristas reproduce flavor profiles more consistently.
Security engineer Andy Nguyen released PS5-Linux, a project that lets PS5 Phat consoles run a functional Linux gaming PC. It uses a patched hypervisor vulnerability to unlock the hardware, delivering 8 CPU cores (16 threads) at 3.5 GHz and a GPU at 2.23 GHz with 4K at 60 Hz over HDMI and support for Steam games and emulators on firmware 3.xx–4.xx. Booting Linux requires running the exploit each time and it is not a dual-boot; a 64 GB USB drive, a keyboard and mouse, and a network adapter are needed, with optional M.2 storage and a Bluetooth dongle for the DualSense.
An article explains that PTFE (Teflon) is highly nonstick and difficult to glue. It describes a chemical pre-treatment using sodium naphthalenide in THF that partially substitutes surface fluorine with naphthalene, enabling adhesion to the outer layer while preserving bulk properties. The piece concludes there is no universal adhesive for Teflon; bonding is achieved only by chemically modifying a surface region rather than using a standard glue.
The Kuleshov effect is a film-editing phenomenon demonstrated by Lev Kuleshov in the 1910s and 1920s, where viewers derive more meaning from the interaction of sequential shots than from a single shot alone. In the original experiment, the same actor's expression was paired with different images—such as a bowl of soup, a coffin, or a woman on a divan—leading audiences to attribute hunger, grief, or desire to the face, illustrating montage's power in shaping interpretation. Subsequent studies have shown context can influence emotion attribution and perception of facial expression, with discussions by filmmakers like Alfred Hitchcock.
Published as a 2026 preprint, the essay argues that languages are converging toward a common Latin-based script, with English functioning as a default pidgin. It identifies three strategies for societies: absorb loanwords as-is, Latinize with local extensions, or resist and incur growing costs; it also contends that linguistic purity offers no advantage and suggests Hebrew was constructed rather than revived. The work discusses emoji as a modern sign of logographic influence and contrasts script with music notation to explain why script can change without erasing content, noting that the author offers ideas rather than implementations.
Structured Prompt-Driven Development (SPDD) is an engineering method developed by Thoughtworks that treats prompts as first-class, version-controlled artifacts used to govern AI-assisted software changes. It centers on the REASONS Canvas—a seven-part structure for Requirements, Entities, Approach, Structure, Operations, Norms, and Safeguards—and uses a defined workflow and the openspdd tool to keep prompts in sync with code. The approach aims to convert business input into a closed loop of requirements, analysis, prompt generation, coding, testing, and release, and emphasizes alignment, abstraction-first thinking, and iterative review, as illustrated by an example of enhancing a billing engine.
The paper evaluates NVIDIA's CUDA Tile (CuTile), a Python-based, tile-centric GPU programming abstraction, across Hopper and Blackwell GPUs and compares it with cuBLAS, Triton, WMMA, and raw SIMT. Benchmarks cover GEMM, fused multi-head attention, and end-to-end LLM inference in BF16/FP16 on H100 NVL, B200, and RTX PRO 6000 Blackwell Server Edition, finding CuTile's performance is highly workload- and architecture-dependent, with Blackwell B200 achieving up to 1007 TFLOP/s for fused attention (2.5x FlashAttention-2) using about 60 lines of Python. GEMM runs at 52-79% of cuBLAS with 22 lines of CuTile code versus 123 for WMMA; on RTX PRO 6000 the CuTile attention kernel reaches only 53% of FlashAttention-2, while Triton sustains 62-101% of cuBLAS across platforms, indicating better portability.
A now page is a page on a personal website that communicates what the person is focused on at the moment. It is intended to reveal current priorities and help the person say no to unrelated requests rather than function as marketing. The article explains how to create one, how to get listed on nownownow.com, and identifies Derek Sivers as the project's founder.
A study examining housing markets in 23 OECD countries from 1990 to 2019 finds that booms and busts are more dramatic than the long-run trend, with prices rising about 6% a year during booms versus 2.6% on average. The study attributes cycles to institutions and policies that encourage speculative behavior, notably low capital gains taxes and landlord-friendly rental regimes. Social housing, rent controls, and tenancy protections are linked to more stable prices, while mortgage credit alone is not the main driver of volatility; housing supply elasticity dampens cycles only modestly.
Apple plans to drop AFP support in macOS, which could leave Time Capsule devices that rely on AFP and SMB1 unusable after macOS 27. Time Capsule hardware last sold in 2013 and was discontinued in 2018, but some enthusiasts have added Samba 4/SMB3 support to run on NetBSD and extend the devices' life. This workaround enables SMB access and Time Machine usage via a Samba server.
The article argues that useful open-source contributions start before writing code, emphasizing context over quick PRs. It advises studying a project’s social architecture, reading documentation and issue histories, and reviewing past PRs to understand decision processes and current pressure points, ensuring changes address real problems. It also recommends using AI to aid learning but not to replace responsibility, and to ask clear, concrete questions when seeking feedback.
Users report a regression in Claude Code where an unconditional malware reminder is injected into every Read and Grep result, causing subagents to refuse legitimate code edits. The issue persists in v2.1.111 despite an earlier fix claim, and it disrupts parallel-agent workflows in benign OSS projects. Proposals to address it include removing the reminder, narrowing its scope, or applying it only to malware files rather than every read.
The video discusses the controversy surrounding the game Pragmata and the idea that male gamers may feel a 'biological clock' driving a desire to marry and have children. It notes that some responses label this interest as incel and situates the debate within broader discussions of masculinity and gender in gaming culture. The video includes a sponsor segment for Hensen Shaving.
Mesa is a POSIX-compatible, durable filesystem with built-in version control designed to connect enterprise agents to the documents they need, enabling parallel execution, human approvals, and audit trails. It provides branches, diffs, durable history, sparse materialization, and fine-grained access control, and can be mounted via FUSE or accessed through SDKs to give agents controlled file access. The project is in private beta with early design partners, with a public beta and v1.0 planned in coming months.
The post surveys coverage-guided fuzzing tailored for smart-contract compilers, explaining how grammar-aware inputs and LLM-assisted tooling help reach deeper compiler passes beyond the lexer and parser. It reports more than 100 bugs across Sui Move, Cairo, Solang, Solidity, and Leo, all triggered by structurally valid programs rather than simple parser crashes. It also details the fuzzing framework, including AFL++–based harnesses, custom mutators, corpus and dictionary strategies, and a triage workflow for deduplication and reporting, plus evaluation results and future work.
The article explains that ChatGPT injects ad units into the conversation stream and ties each ad to a four-part Fernet-based attribution loop (ads_spam_integrity_payload, oppref, olref, ad_data_token) for reconciliation of impressions and clicks. A merchant-side OAIQ SDK in the user’s browser collects post-click data via first-party cookies to support attribution, with ads loaded from OpenAI-hosted creatives. The piece provides observed examples across several brands and notes contextual targeting by chat topic, as well as the domains involved in ad data collection.
Researchers Felicia Fang-Yi Tan and Oded Nov presented CHI’26 findings that delaying chatbot responses by 2, 9, or 20 seconds influenced 240 adults to rate answers higher, with longer delays interpreted as thinking. They described Context-Aware Latency and 'positive friction' as design variables—faster replies for simple questions and slower ones for complex or moral queries—and warned against over-trusting slower AI, noting ethical concerns about deceiving users into thinking the AI is thinking.
Anthropic announced a set of connectors that let Claude work directly with major creative tools, including Blender, Autodesk, Adobe Creative Cloud, Ableton, SketchUp, Affinity by Canva, and Splice, to assist across design, 3D, and audio workflows. The connectors enable Claude to learn tools, generate code, translate formats across pipelines, and automate repetitive production tasks, with Blender’s MCP connector now available and Anthropic joining Blender Development Fund; the company is also collaborating with design programs at RISD, Ringling, and Goldsmiths to test the tools.