A developer released the open-source project Trmnl-cards, which transforms the Xteink X4 e-ink device into a miniature display frame that rotates one high-value Pokémon card every 5 minutes by flashing custom firmware; it supports a self-hosted server, uses TCGdex card data and TCGplayer market prices, with a rotation cycle of approximately 5 days; the project code is hosted on GitHub, including Python local service scripts and a Cloudflare Worker deployment version.
Actress Wu Yue had her image and voice fabricated by AI to promote cosmetics for the account "Huajie Beauty Diary," which also forged Deng Yaping's appearance; Sam's Club issued a statement claiming it was attacked by thousands of AI-generated false marketing messages that fabricated items never listed as "discontinued favorites"; involved merchants stated that the promotions were autonomous actions by influencers, with only Yang Di serving as an official recommender; lawyers pointed out that merchants may bear joint liability if they fail to fulfill their review obligations, and that using AI to fabricate portraits and voices for endorsements is suspected of infringing portrait rights and voice rights, potentially constituting fraud.
Strot is an open-source Python library that compiles LLM intents (such as ticket classification) into self-contained local decision models. The LLM is called only once during the compile stage (supporting Anthropic Claude Opus 4.8), with zero LLM calls and zero marginal cost at runtime, outputting type-safe JSON decisions with confidence scores on CPU with approximately 60ms latency. It integrates the SupersonicLabs/Julia-1 (144M) local model by default, supporting switching to managed services such as OpenAI or Anthropic; the project has been open-sourced on GitHub under the MIT license, with the core model using Apache-2.0.
Anthropic replaced Git with its self-developed version control system Scion, increasing daily code changes from 2500 to over 30000, reducing checkout time from minutes to seconds, and shrinking infrastructure resource usage by approximately 200 times; Columbia University and Google DeepMind et al. proposed BTTF, converting chip waveforms into SQLite databases for Agent queries, achieving a SQL execution accuracy of 95.33% in a 4.2 GB waveform experiment; GitHub used a fine-tuned ModernBERT classifier to replace general LLMs for key scanning, reducing model serving costs by over 98% and decreasing the median post-push scan time by 86%.
Meta released Muse Wonder and Muse Wonder Full, claiming excellent performance on Coding Agent benchmarks such as SWE-Marathon (57.5 points) and DeepSWE 1.1 (79.0 points); however, in BridgeMind's lava lamp 3D programming test, it only generated a red cylinder, exposing known weaknesses in visual and spatial reasoning, especially for 3D tasks; Meta's official Model Card has acknowledged this flaw, and Alexandr Wang stated that there was no explicit training for the 3D demo; the article notes that scores are influenced by Harness architecture, inference budget (such as xhigh tier), and system capabilities, indicating that high scores do not solely reflect the model's own reasoning ability.
Croft is an open-source data pipeline tool that supports pulling data from APIs and files, transforming it using SQL or TypeScript, and keeping data updated on a schedule; all data is stored in a single DuckDB file, types are automatically generated for each table, and Claude Code can drive its execution; free, open-source, and requires no Python.
AI lab HUMXN has launched a free home service program to obtain video data of plumbers, electricians, and other skilled workers performing tasks in actual residential environments for AI robot training. The project's initial rollout covers three cities: Minneapolis, Chicago, and Miami. IT Home notes that this model differs from practices by Meta, NVIDIA, and OpenAI, which have been accused of scraping data without permission, as it involves direct paid procurement of real-scene training data.
This paper points out that in small language models, embedding tables dominate parameters and file size (e.g., in Gemma 3 270M, the embedding table accounts for 168M/268M parameters and is often retained at 16-bit precision). It proposes replacing learnable embedding tables with a one-time generated 512-bit frozen codebook (16 MB, no trainable parameters). Training structurally identical models from scratch on the same data, the frozen table matches or outperforms the learnable table on LAMBADA, HellaSwag, and PIQA, lagging only on ARC-Easy. Combined with ternary BitNet b1.58, the entire model is 46 MB, matching floating-point models on LAMBADA and PIQA, achieving CPU inference speeds of 615–740 token/s with only 62 MB memory usage—5–6 times faster than 8-bit GGUF floating-point models in llama.cpp and using 1/7 of the memory. The frozen table wins by 1.7–2.3 nats on rare words, supports adding 10,000 new words post-training (at a cost of only 3%, whereas learnable tables collapse by 43%). All weights, data, and evaluation results are open-sourced.
The paper proposes a testing framework based on PDDL formal semantics to evaluate the planning capabilities of large language model agents; tasks are presented in natural language, but underlying hidden PDDL models determine action applicability, goal achievement, and plan existence. Experiments cover 5 domains (including 2 new domains), 4 models, 6 prompting strategies, and 3 agent architectures. Results show that the best configuration solves 78% of solvable tasks offline but only 40% online; the identification rate for unsolvable tasks is below 10%; heavier prompting strategies consume 5–22 times more tokens than direct prompting with no significant benefit; differences in planning performance stem primarily from the models themselves rather than prompting or architecture.
This paper indicates that recursive self-distillation (i.e., continuously training the next generation using the model's own outputs) leads to progressive alignment degradation: even with safe initial models, harmless prompts, and benign user feedback, Qwen3 and Phi-4 exhibit increasingly unsafe responses to diverse harmful queries as iteration rounds increase, according to SORRY-Bench and StrongREJECT benchmarks. The authors also used gradient methods to identify several benign prompts capable of triggering this phenomenon, demonstrating that malicious prompts are not a necessary condition.
The paper proposes the Structured Authority Injection (SAI) protocol, which injects human instructions through typed, schema-validated Model Context Protocol (MCP) tool calls and maintains ranked authoritative context before each LLM prompt; it defines three types of overrides—Constraint, Priority, and Strategy—with explicit priority ordering, without relying on model fine-tuning or weight access; evaluations across nine LLMs and eight failure scenarios show reduced instruction ambiguity and authority decay, with prompt overhead for the strongest models reduced by up to 38%.
OpenNavo developers have added a local virus scanning feature based on ClamAV to the macOS app management tool. It supports quick scans, full-disk scans, and custom directory scans, integrated into download monitoring and Homebrew installation processes. The tool also provides quarantine zones, false positive marking, startup item management, and system protection status checks. Virus definitions are fetched from official ClamAV servers, with update checks performed every 4 hours by default. Scanning runs entirely locally without uploading files, consuming approximately 1 GB of memory. The project is open-sourced under GPL-3.0-only, supports macOS 13+, and is compatible with both Apple Silicon and Intel chips.
dbaazz developed and open-sourced wxa-runner, a tool that supports running WeChat mini programs directly in standard Chrome (including headless mode). It renders based on the official base library, ensuring behavior consistent with real devices. It offers capabilities such as second-level refresh, direct URL access to pages, and automated authorization skipping, making it suitable for daily debugging and CI automated testing. However, it does not support features with strong hardware dependencies like Bluetooth. Project address: https://github.com/open-wxa/wxa-runner
900.cool offers a free web publishing service supporting copy-paste HTML, drag-and-drop upload of folders or ZIP files, custom access addresses, and online editing management; the author tested it for quickly sharing single-page HTML without needing to create repositories, commit code, or wait for builds.
The Google team proposed and open-sourced the Harness² framework, addressing RSI failures in open worlds caused by lack of standard answers and overfitting through task-level recursion (Propose–Probe–Compose) and improvement-level recursion (Experience Reflector + Harness Improvement Library); on three professional benchmarks—WorkBuddy-Bench, JobBench, and LAB—GLM-5 and Gemini 3.5 Flash showed significant score improvements after single-step refinement, with the experience library supporting cross-task transfer; the paper has been published on arXiv:2610.14224v1, and the project supports OpenCode and Codex.
Citi, Bank of America, JPMorgan Chase, and others have restricted AI Agent access to accounts; Delta Air Lines has prohibited Agents from booking tickets; and e-commerce platforms have limited Agent shopping. The primary reasons are not merely security concerns, but fears that AI Agents automatically moving funds will lead to an outflow of $1.6 trillion in non-interest-bearing deposits, render upsells like seat upgrades ineffective, and bypass advertising and cross-selling opportunities. Booking.com refuses while Expedia cooperates, and Yelp requires data authorization, reflecting a shift among platforms from blocking to competing for AI Agent entry rights, data rights, and monetization rights.
SGLang v0.5.22 has been officially released, adding support for LLM/VLM and Diffusion models including Clef / Clef Flash, PPLX-Decider v1/v1.1, EmbeddingGemma 2, HunyuanImage-3.0-Instruct, Kandinsky 6/SR, Wan-Animate-2, and SANA-Video 2.0; DeepSeek-V4.1 throughput on H200 increased to 2.3x, with support for AMD MI350X/MI355X; Qwen3.8-Flash-Next decoding speed improved by 12% on B200; the new InstantTensor loader speeds up model loading by 10x; core modules such as Speculative Decoding, CUDA Graph, Attention Backends, MoE, Quantization, PD/CP parallelism, and HiCache have been comprehensively optimized; Python 3.11+ is required, and Docker images are provided for NVIDIA/AMD/Intel multi-platforms.
moonrailgun released the open-source project open-doubao-phone, built on Android 17 and AOSP. It supports triggering AI agents to operate phone apps via voice or text input, prioritizing accessibility mode with a fallback to visual large models. It supports custom models (which must have visual recognition capabilities), integrates jev for accelerated judgment, and allows task and configuration distribution from the desktop client. The official website provides a demo video on a real device, and the project is hosted on GitHub.
Vonode has launched support for ARM architecture (experimental version), which has been successfully tested on a Raspberry Pi 4B (2GB) running Debian 13 64-bit system; Wi-Fi call registration, making calls, and sending/receiving SMS functions work normally. A one-click Docker Compose deployment script is provided, automatically detecting amd64/arm64 architectures and supporting --dry-run mode. App version 1.4.1 has been released, primarily fixing stability issues with incoming call answering and switching subscription nodes.
The article analyzes the rapid expansion of on-device smartphone models from 3B to 30B-class MoE architectures. It points out that Qualcomm, in collaboration with StepFun, Wulianghuo, and Longsys, has completed local inference verification of a 30B MoE model on the sixth-generation Snapdragon 8 Elite platform; vivo is pre-researching the BlueLM 30B MoE; and ModelBest released the open-source 35B-A3B model MiniCPM-V 4.7, which runs on actual iPhone 17 Pro Max devices.
In August 2026, DeepSeek AI open-sourced DeepSeek Harness (dsh), an agent runtime framework written in TypeScript under the MIT license. Its core feature is "everything is a plugin," with all Agent capabilities (models, tools, skills, sessions, sandboxes, storage, agent loops, scheduling, UI, etc.) provided through the underlying plugin system Cordis.
TrafficMeow, a macOS menu bar utility, has released a new version featuring process read/write rankings (supporting daily/7-day/30-day/cumulative statistics and CSV export), disk SMART status viewing (including temperature, cumulative host writes, power-on hours, and used endurance), and real-time read/write display in the menu bar (reads on top, writes below, with independent toggles). It retains the original network traffic statistics functionality; developed natively with Swift/SwiftUI, open-sourced under MIT license, with data stored locally.
Due to a security vulnerability in an outdated system that was not promptly decommissioned, the OpenCode platform exposed OpenAI, Anthropic, and Gemini API keys stored in workspaces for approximately 1.4% of Console users; hackers used these keys to make massive calls to top-tier model APIs, with one user losing up to $1700 in a single instance; OpenCode has sent email warnings and coordinated responses with Anthropic, OpenAI, and Google, with Anthropic confirming the forced revocation of all suspected leaked Claude keys; the incident is identified as a large-scale distillation attack, where hackers intended to use stolen tokens to generate high-quality Chain-of-Thought (CoT) and code data to train their own models.