yuxi1989 developed the macOS utility Codex-RetryPilot to automatically click the "Retry" button when Codex displays 'Selected model is at capacity. Please try a different model.'; it supports a three-layer mechanism involving Accessibility APIs, in-process event delivery, and coordinate-based clicking, operates without internet connectivity or telemetry, and requires Accessibility and Screen Recording permissions; installation options include Homebrew, dmg download, and source code (MIT License).
Ethereum developers postponed the Glamsterdam upgrade on the Hoodi testnet because ChainSecurity discovered that the combination of EIP-8037 and EIP-8038 could alter how existing smart contracts handle failed transactions: some contracts that should have reverted entirely due to gas exhaustion might now execute failure-handling logic due to gas refunds; other contracts might unexpectedly fail during valid operations due to increased gas costs (e.g., creating a new storage slot rising from 22100 to 110020 gas).
Alibaba Qoder announces support for Linux ARM64 and Loongson LoongArch64 platforms, compatible with UnionTech UOS and Kylin Desktop Operating System V10; this upgrade brings the complete agent workflow to domestic desktop systems, supporting tool invocation, file operations, and multi-agent collaboration; Qoder has completed adaptation for HarmonyOS PC, HarmonyOS mobile, UnionTech UOS V20, and Kylin V10.
WorldScape R²SI from Manifold AI (Manifold Space) has taken the top spot on the RoboDojo Agent Track, achieving a macro progress score of 46.58, a fine manipulation success rate of 15.5%, a long-horizon task performance of 34.25%, and generalization capability of 35.83%. Its core is the R²SI recursive self-improvement framework: an outer-layer Reflective Agent reviews execution logs to revise Skills or update the WSP model, while an inner-layer Reactive Agent handles execution calls—requiring no retraining or manual code changes throughout. The accompanying ICL capabilities (WorldScape Policy 2.5) support learn-by-demonstration, action transfer, skill composition, and long-term memory. The synergy between ICL and R²SI enables 'learn it in one look, improve it in one try.'
Pristine is an open-source agent skill released by independent developer maddhruv. It supports agent tools such as Claude Code, Codex, GitHub Copilot, and Cursor, offering four types of commands: /pristine init, /pristine critique, /pristine design, and /pristine polish. It includes 16 composable lenses (such as bolder, typeset, adapt) that can score and automatically optimize web pages across devices/themes. The project is hosted on GitHub; documentation and examples are available at https://pristine-ui.pages.dev.
GAV (Generator-Agent-Verifier) is a scientific experiment environment for partial differential equation (PDE) programming agents. The Generator learns continuous reference solution distributions based on numerical simulations or experimental data and reverse-constructs matching PDE problems using manufactured solutions, making new problems inherently verifiable; the Agent iteratively designs experiments and improves programs; the Verifier provides numerical feedback on accuracy and computational cost based on reference solutions and can guide the generation of targeted new problems.
SWE-Verifier is a lightweight generative model used to selectively trigger candidate action sampling and verification before a coding agent executes an action. It is trained based on Qwen3.5-4B, using preference data to learn scoring and ranking actions. On MiniMax-M2.7, it improved the resolution rate from 75.47% to 77.20% (at a cost of 1.09x), and on the untrained DeepSeek-V4-Flash, it improved the resolution rate from 72.60% to 74.27% (at a cost of 1.13x). Its verification steps account for less than 4%, and the code and data are open-sourced.
The paper proposes an automated method for screening individual patient privacy risks in medical AI models. It conducts membership inference attack risk assessments on 6 clinical datasets, 4 modalities, and 7,789 models, finding strong consistency among high-risk patients across different training configurations, architectures, and initializations. Based on this, a lightweight screening tool using 4 small models was developed to identify vulnerable patient records before production model training, and protection pathways without data deletion were explored.
This paper proposes a human-AI collaborative discovery loop that combines a fast numerical pipeline, an exact symbolic verifier, and large-scale distributed search to construct a five-parameter generalized Boyd algorithm. It achieves, for the first time, the generation of irreducible integer-coefficient polynomials with Mahler measure strictly below the Boyd limit point α₁≈1.255434 for any even degree d (≤2000). The method maintains the best known record on all tested even degrees with 99.6% success rate, and open-sources a verified dataset of minimal Mahler measure polynomials.
The paper proposes the Latent Interconnect Protocol (LIP), which maps high-dimensional representations from expert models such as proteins and small molecules to the internal states of Gemma 3 (4B/12B) and Kimi K2.6 (1T total parameters, 32B active) via learnable connectors, while freezing all pre-trained models. Independently trained connectors support zero-shot cross-domain task transfer and multi-expert combination, but current LLMs are not natively adapted for expert representation inputs. Code, connectors, and data have been open-sourced.
This paper presents the first staged empirical benchmarking of the handshake processes for three major AI agent communication protocols: MCP (2025 initialization version and 2026 discovery version), A2A, and ACP. Using official SDKs for each protocol, it measures round-trip counts, per-stage latency, network bytes, and token consumption under scenarios such as cold start, hot reconnection, capability mismatch, and metadata degradation (N=50/scenario). Results show that round-trip counts dominate overhead; the old MCP version requires 3 round trips taking 230ms (50ms RTT), while A2A requires only 1 round trip taking 61.0ms. When an LLM reads peer capability metadata, token consumption reaches 24–107 tokens/capability, meaning 200 capabilities occupy 11% of a 200k context window. The MCP 2026 version reduces round trips to 2, and hot reconnections can drop to 0. All testing tools, raw data, and analysis are open-sourced at https://github.com/pvarshh/handshake-benchmark.
This paper compares two external system interaction paradigms in AI agents: Command Line Interface (CLI) and Model Context Protocol (MCP). Based on an evaluation across 14 dimensions, it indicates that CLI is more suitable for single-user local workflows, while MCP can meet requirements that CLI cannot achieve, such as fine-grained access control, credential isolation, structured auditing, and multi-tenant governance. Through progressive disclosure techniques, the context window overhead of MCP can be mitigated, bringing its token efficiency close to that of CLI. It recommends adopting a hybrid architecture for production systems.
At the 2026 World Physical AI Conference, Unitree founder Wang Xingxing stated that the ChatGPT moment for embodied intelligence has not yet arrived, and it will take 1-2 years or 2-3 years to break through physical AI. The core bottleneck lies in aligning AI models with real robots at the level of physical contact; currently, only scenarios with high fault tolerance and fixed tasks, such as commercial services, are deployable. Its latest product, H2-X, still focuses on non-generalizable tasks such as performance and security. While industry financing has surged, there is a lack of application scenarios, leading to winter signals such as the bankruptcy of Yinghe Robot, unpaid wages at DeepBlue, and a halving of Unitree's market value.
qxmqh has released WorkSkill.tools, an independently developed skill store for coding agents, offering over 400 skills and supporting Claude Code, Codex, Cursor, etc.; 60 free skills (36 skills + 24 templates) require no registration and can be installed via one-click linking; paid skills deliver finished videos and editable source code; accompanied by two GitHub repositories: mailes/awesome-agent-skills and mailes/awesome-claude-5-5-videos.
Media AI, an iOS multimedia player developed by myqonnt, has been released on the App Store. It uses Whisper on-device to generate subtitles for local videos and audio files. Features include real-time segment-by-segment display, priority processing of the current playback position, automatic saving of progress, bilingual comparison, speaker identification, subtitle export in SRT/VTT/HTML/JSON formats, as well as access to podcasts via RSS, WebDAV, and SMB file protocols. Translation relies on system capabilities, and some language models need to be downloaded upon first use.
Google has launched the Agent Compute Engine (ACE) service, which provides AI agents with hardware-isolated micro virtual machines (microVMs). It supports fast startup, independent kernels, root privileges, and automatic suspension during idle periods with millisecond-level recovery (pause/resume time was under 500 milliseconds in testing). This service is already being used to support Gemini, Antigravity, and internal model evaluation processes, and will be made available to developers via Google Cloud in the future.
sone47 has open-sourced cinetty, a TUI architecture diagram generation Skill designed specifically for terminals. It includes 6 built-in templates, natively supports Claude Code and Codex, optimizes the rendering and alignment of Chinese characters in terminals, and can convert READMEs, knowledge structures, or workflows into dynamic TUI interfaces with one click. Project URL: https://github.com/JosssphZhou/cinetty
Kirahowe released editmd 0.1.0, a native macOS Markdown editor built on AppKit and WebKit. It uses Y.js CRDTs to enable conflict-free real-time editing of the same file between users and AI agents. It supports GitHub Flavored Markdown, Mermaid diagram previews, and secure handling of external documents and remote images. Requires macOS 14+ and is installed via Homebrew.
gnakzz has open-sourced wx-publish-studio, which integrates the entire workflow of writing, formatting, cover creation, and publishing for WeChat Official Accounts. It supports one-click startup via docker compose up; users need to configure their Official Account information. The tool offers preset layout templates, custom CSS, and reusable cover templates. Article rendering relies on milkdown with extended syntax components, but the image rendering plugin does not currently support all syntax features. It is recommended to use skills to distinguish between article and image content to avoid draft box recognition failures.
XiaomiAlbumSyncer is an album synchronization tool developed by author cofan. It now supports dual platforms: Xiaomi Cloud (albums + recordings) and Apple iCloud (photos). Features include Web UI configuration, scheduled incremental synchronization, multi-account management, Exif time filling, and Passkey authentication. The project is positioned for archiving; it does not synchronize delete operations. It runs on NAS or homelab environments, does not interact with phones, and requires only an internet connection.
On October 22, 2026, a16z crypto officially renamed itself to a16z Frontier, announcing an expansion of its investment scope to frontier technologies such as AI and robotics, while continuing to invest in Crypto. This move marks its transition from a vertical fund focused on crypto to a broader frontier tech investor, echoing the trend in the first half of 2026 where leading crypto VCs like Paradigm, Haun Ventures, and Framework Ventures simultaneously downplayed their Crypto labels. Background factors for the rename include the flagship fund shrinking to $2.2 billion, the failure of heavily invested project Farcaster, large-scale LP capital flows into the AI sector, and the reality that blockchain technology has 'faded into the background' in scenarios like stablecoins and tokenized assets, diverging from Web3 fundamentalist narratives.
sobinyuan developed and launched brascale.com, a purely static bra size calculator targeting US users. Deployed on Cloudflare Pages, it supports multiple calculation logics including band size, cup size, sister size, and plus-four algorithms, with all calculations performed client-side in the browser.
NVIDIA Founder and CEO Jensen Huang visited Astana and jointly announced multiple AI collaborations with Kazakh President Tokayev and Deputy Minister of AI and Digital Development Madiyev: Firebird obtained a U.S. export license and will build the first phase of an AI factory in Kazakhstan, deploying 45,000 NVIDIA Grace Blackwell GPUs for commercial use in Q3 2027; Kazakhtelecom has already deployed 400 H200 and 480 B300 GPUs, with plans to add over 2,000 Vera Rubin GPUs next year; Higgsfield AI received investment from NVIDIA; inDrive and Pony.ai are exploring robotaxi services based on the NVIDIA Hyperion platform; Madiyev emphasized developing Kazakh-language AI models and physical AI (for mining, agriculture, logistics, etc.), while Huang stated that 'domestic data is a natural resource' and proposed 'converting energy into intelligent exports.'