Xinzhi Zhang, a researcher at Microsoft Research, along with collaborators, proved the Kannan-Lovász-Simonovits (KLS) conjecture proposed in 1995, establishing a dimension-independent constant bound on the bottleneck of hyperplane cuts in high-dimensional log-concave distributions; this breakthrough was achieved through human-led AI agent collaboration, exploring over 100 mathematical paths, ultimately eliminating dimensional dependence via recursive error control methods, and receiving parallel verification from multiple independent teams as well as formal verification in Lean 4.
The GitHub engineering team announced a critical upgrade to its Git infrastructure, shifting write serialization granularity from repository-level to individual reference-level, and introducing virtual write replicas in Azure to decouple read and write workloads; these changes enabled high-load repositories to complete migration to Azure, with internal tests showing a 3–4x increase in write throughput for different references; currently, most repositories have two replicas deployed in Azure, which handles approximately half of Git traffic.
Switchboard is a free, open-source macOS menu bar application for visualizing and managing all services running on localhost. It groups services by project and supports one-click start/stop, log viewing, port jumping, and dependency graphs. It offers installation via brew and a command-line tool named switchboard, requiring macOS 14+. The project is hosted on GitHub.
On October 21, 2026, the PHP Foundation released versions 8.5.12, 8.4.27, 8.3.36, and 8.2.35 to fix CVE-2026-109618 and CVE-2026-109619, two unauthenticated remote code execution vulnerabilities with a CVSSv4 score of 9.2; government websites in Finland, the Netherlands, France, and other countries were temporarily taken offline, and Cloudflare observed early attacks hours before the patch release and pushed WAF rules; the vulnerabilities affect PHP 5.4 and above, but official complete updates are only provided for versions 8.2+, while versions 8.1 and earlier rely on community patches or upgrades; PHP core developers emphasized that the only mitigation measure is to immediately upgrade PHP.
Tencent is internally testing 'Dongle Ya', an AI interactive community for children and adolescents. The official website dongleya.qq.com and WeChat mini-program are now live, currently open to a small group of parent-child users and frontline teachers. Users can generate 'Wonder Cards' via natural language input—interactive content units including 3D demonstrations, dynamic courseware, interactive picture books, lightweight games, and memory cards—without programming.
Cohere has released the Rerank 5 model, available in Pro and Fast versions. It supports document-level and paragraph-level reranking and introduces a Context Focus feature to reduce LLM input tokens. It provides calibrated relevance scores (e.g., 0.7 indicates a 70% probability of usefulness). The model is now live on the Cohere API, Model Vault, Microsoft Foundry, Amazon SageMaker, and the North platform. The Pro version leads competitors on benchmarks such as ViDoRe V3, nanoMTEB, BRIGHT, and BEIR, while the Fast version achieves 2.9 times the throughput and nearly 3 times the latency performance of the Pro version. API pricing is $0.05/1M tokens for Pro and $0.02/1M tokens for Fast.
IBM released the Granite Speech 5.0 series voice models, including the new model granite-speech-5.0-2b (2 billion parameters, supporting multilingual transcription and translation for English, French, German, Spanish, Portuguese, Japanese, Italian, and Chinese, with language ID hints, 2000-word keyword biasing, punctuation and capitalization control), the updated granite-speech-5.0-470m-turboctc (multilingual, native punctuation, reverberation-resistant, RTFx reaching 19870), and granite-speech-4.1-2b-plus (optimized for speaker attribution, timestamps, long audio performance, and keyword biasing).
Inception has released the diffusion language model Mercury 3, claiming it outperforms competitors in 8 of 12 benchmarks, with an inference speed reaching 1068 tokens/s, significantly faster than GPT-6 Luna (approx. 130 tokens/s) and Claude Haiku 5.5 (approx. 240 tokens/s); it supports a 1M context window and is API-compatible with OpenAI; current pricing is $0.10/M tokens for input and $0.35/M tokens for output, and it is now available on OpenRouter, Baseten, and Vercel.
ElevenLabs has launched its Realtime Translation product, targeting telecom operators, call centers, and live events. It supports real-time bidirectional voice translation in over 90 languages. The solution employs single-WebSocket end-to-end streaming processing, maintains speaker voice consistency (with options for official voice libraries or custom cloned voices), synchronously outputs subtitles and text, and allows adjusting the trade-off between speed and accuracy based on scenarios.
The Solana mainnet activated the Alpenglow consensus protocol on October 21, 2026, replacing TowerBFT which had been in use since March 2020; the switch was completed at slot 459653000. Transaction finality (irreversibility) was reduced from approximately 12.8 seconds to approximately 150 milliseconds.
ModStart has released the open-source desktop tool QuotaPanel, developed with Go + Wails + Vue3. It supports 16 types of AI subscription and API quota sources, including Claude, ChatGPT, GitHub Copilot, Kimi, etc. Adopting a local-first architecture, data is stored by default in ~/.quotapanel/. It supports macOS/Windows/Linux and offers features such as browser login, API Key management, quota visualization, PNG export for cards, system tray residency, and bilingual support (Chinese/English).
Nvidia unveiled the DSX POD at GTC Berlin, a modular AI data center reference design based on the Vera Rubin system. Partners including Dell, HPE, Lenovo, and Schneider Electric complete the pre-installation and testing of computers, network equipment, cables, power supply, and liquid cooling systems in the factory; on-site installation only requires connecting power, cooling, and networking. Deployment time is reduced from six months to as little as 14 days. Its closed-loop liquid cooling system claims to reduce evaporative cooling water usage by up to 100%, though additional cooling is still required in extreme high-temperature climates. The first units will ship globally in early 2027.
Uberblok is an interactive web museum showcasing the internal mechanisms of the ZFS file system. All animations and code snippets strictly correspond to the OpenZFS 2.4.4 version (commit 71a9f9578616) source code, focusing on parsing core designs such as the uberblock structure, transaction groups (txg), block pointer rings, vdev architecture, and ARC cache.
Claude Code has launched the Claude Mods feature, enabling users to write plugins in TypeScript/JavaScript to customize behavior and interface rendering on CLI and desktop clients, achieving real-time intervention in Agent execution flows and human-machine interaction points; Anthropic includes six built-in Mods: agents-md, diff, plugin-authoring, sec-default, telemetry, and you-should-know; the author created a Next Steps Supervisor Mod and open-sourced it via ModSmith; this feature shifts the Agent Harness from machine-oriented to human attention-oriented, constituting a programmable Attention Interface.
Opposable is an Android app that reads the screen and operates other apps via accessibility services; when not connected to a cloud model, it can run GUI-Owl 1.5 2B locally on the phone (4-bit quantization, totaling 1.05 GB language model + 445 MB visual encoder), supporting about 40 fixed tasks such as setting alarms and sending SMS, with other actions generated by the model; single-step latency is 7.3 seconds on Mac M2 CPU, while visual encoding takes only 236 ms on Galaxy S25 NPU; the model is a multimodal model fine-tuned by Alibaba's Mobile-Agent team based on Qwen3-VL, taking screenshots as input and outputting action instructions; the app is free, but the screen model requires separate download.
Sumi.css is a low-class-first CSS framework weighing 6.7 KB (minified + gzipped), with no JavaScript, no dependencies, and only a single file; it defaults to HTML element styles, supports automatic dark mode, full accessibility, and zero-specificity (:where()) design; all colors, fonts, sizes, etc., are controlled via CSS custom properties prefixed with --s-, facilitating theme overrides; it provides CDN and local installation methods and is compatible with modern browsers.
NVIDIA and the Institute for Protein Design at the University of Washington jointly released the open-source protein design model RFD4-Proteina. Based on the BioNeMo framework, this model supports the joint generation of protein sequences and all-atom structures, and introduces reward-guided inference-time search to focus on high-scoring candidates. It offers two deployment methods: an open-source model (with sampling code control) and a NIM microservice. Measured results show it is 3.8 times faster than RFD3 in 200-residue tasks, with peak memory usage reduced by 84%. It supports integration with external oracles such as OpenFold3, accompanied by the AtomWorks data preparation tool and Agent skills like rfd4-proteina-campaign.
Timeminator is a self-hosted, project-based time tracking tool that supports running on PHP shared hosting (MySQL/MariaDB) or locally with SQLite without configuration, requiring no Composer or build steps; it uses a client → project → work package → task hierarchical data model, offering real-time timers, CSV bulk import with rollback, team mood tracking, role-based permission management, and built-in offline chart rendering; currently in pre-1.0 alpha/beta version, the Community edition is open-source under MIT license, while the Pro edition is a commercial extension under private license.
Mistral released Mistral Large 4 (codename Le Chonk), which has approximately 1 trillion total parameters and 49 billion active parameters. It is a multimodal model trained on 3800 NVIDIA Grace Blackwell GPUs in European data centers, with weights scheduled for release at the end of October; this model will integrate with NVIDIA's accelerated computing, Omniverse, CUDA-X, and Warp tools to accelerate engineering simulation and closed-loop AI-driven design, and jointly launch public benchmarks and open datasets for physical modeling and agentic AI.
The article analyzes the root causes behind DeepSeek and OpenAI taking diametrically opposite paths in embodied AI: DeepSeek locks in interfaces and acceleration capabilities by investing in Unitree Robotics and Winverse AI, avoiding the high fixed costs of real-robot data collection and body manufacturing; OpenAI, having lost access to physical data after terminating its partnership with Figure, has shifted to developing actuators in-house and building data collection factories, undertaking heavy asset investments.
Zijing ZhiKang (incubated by Tsinghua AIR) has launched the 'Zijing AI Hospital' system, which enables same-department support and cross-departmental collaboration through dedicated doctor AI avatars, offering two interaction forms: on-demand reminders and consultation mode. The system is based on the SEAL dual-layer evolutionary architecture, trained in a virtual medical world and fine-tuned with real outpatient data. As of July 2026, it covers 18 clinical departments at Beijing Tsinghua Changgung Hospital, with over 40 doctors using it routinely; diagnostic efficiency has improved by 30%, and MedQA accuracy exceeds 96%.
Hugging Face and NVIDIA jointly released SimReady Central, an open-source library containing over 2000 simulation-ready (SimReady) OpenUSD assets, covering robot arms, grippers, warehouse equipment, racks, cooling facilities, etc.; it supports OEM vendors in submitting verified assets according to the SimReady Foundation specification, and enables developers to quickly build robotic pick-and-place scenarios or digital twin layouts for AI factories through natural language search, USD Search, and AI assistance (such as GPT-6 Astra).