Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

AI Trending

0113

MachinePulse AI open-sources Karpo decision models and browser-agent tools

MachinePulse AI open-sourced six Karpo Decision models, the Karpo Decision 27B Act web-action model, the TypeScript karpo-browser framework, and the KBB browser-action benchmark. The browser system uses a general-purpose model for planning and the dedicated action model for web interactions, while karpo-browser handles page reading, execution, page changes, and errors. The decision models use a Readout Head to return candidate probabilities in one forward pass without token-by-token generation. QbitAI reported that the team’s tests found an 89% average success rate across 102 Browserbase Benchmark v2 tasks, with an average duration of 201 seconds and model cost of $0.096 per task.

026

App Store Compliance Guard targets app rejections

The mjmirza/app-store-compliance repository provides a playbook and pre-submission guard for iOS and Android apps, including checks for privacy manifests and Data safety forms. Its author says the project documents 117 rejection patterns and fixes and tracks 95 regulatory deadlines. The guard can run through AI coding tools and block submission commands when it detects critical risks. The repository uses the OpenRoots Agent License 2.3 for its current release, and earlier releases have different license terms.

038

Tony Zhang open-sources Linger personal AI agent

Simplex AI founder Tony Zhang has open-sourced Linger, a personal agent designed to run on a user's computer, reuse existing browser and application sessions, and connect third-party or local models. Its Another Me component can monitor email, calendars, chats, and meetings, maintain a local knowledge base, issue proactive reminders, and learn notification preferences from feedback; the project also includes the PolyStrata local inference engine. Vendor-reported claims include 95% accuracy on LoCoMo and local inference of Qwen3.8-Flash-Next on consumer hardware, while GitHub stars reportedly exceeded 2,000 shortly after launch.

046

sakurai-skill packages Sakurai’s game-design tutorials for AI agents

Japanese developer kociden published the unofficial “sakurai-skill” project on GitHub, organizing subtitles from all 257 videos in Masahiro Sakurai’s game-development YouTube series into files that AI agents can consult selectively. The roughly 27,000-word main file covers 14 chapters on game design, planning, action design, effects, UI, sound, and control-feel adjustment. A sample game, Hajiki, demonstrates the skill’s application with Godot 4.6.3. The developer says the material was summarized by AI from automatically generated subtitles, is not Sakurai’s verbatim words, may contain speech-recognition or interpretation errors, and has no association with Sakurai.

053

AWS Labs project coordinates AI coding CLIs

The AWS Labs repository describes CLI Agent Orchestrator (CAO), an open-source project that coordinates multiple AI coding CLIs so a supervisor can delegate work in parallel or sequence. Its documentation says CAO runs a local cao-server and starts provider CLIs in isolated terminal sessions while preserving their native authentication and capabilities. Supported tools include Claude Code, Kiro CLI, Codex CLI, GitHub Copilot CLI, Cursor CLI, and others. The supplied source does not establish a specific release date.

063

Phone Harness Lets AI Agents Control Real Phones

Phone Harness is a tool that lets AI agents such as Claude Code and Codex control a real phone by scrolling, tapping, typing, and reading the screen. The project supports iPhones through a Mac's iPhone Mirroring window and Android devices over adb, with Phone Harness Cloud offering hosted Android and iPhone options. Its capabilities and demonstrations are described by the project maintainer.

073

mini-swe-agent v2 launches as a minimal software-engineering agent

The SWE-agent project released mini-swe-agent v2, a compact command-line AI agent whose agent class contains roughly 100 lines of Python and uses only Bash for actions. The project says it scores above 74% on SWE-bench Verified, now powers Ramp SWE-Bench, and outperforms Claude Code and Codex on DeepSWE, though these performance claims are not independently verified in the supplied report.

080

Tencent releases WeMM-Video-9B multimodal checkpoint

Tencent's model card describes WeMM-Video-9B as the dense 9B checkpoint in the bilingual WeMM-Video family for joint audio-visual understanding and reasoning. It accepts text, images, audio, and video and generates text responses. The release includes BF16 Safetensors, quantized variants, Transformers support, and serving instructions for patched vLLM and SGLang versions.

093

Cognition Reports Devin Bot Helping Maintain Manim

Cognition reported that Grant Sanderson set up a Devin Bot to help maintain Manim, the open-source animation system behind every 3Blue1Brown video. The report cited 96.6k GitHub stars, 7,936 forks, and 250,000 monthly downloads. Cognition also claimed that Devin Bots can lead teams of agents in tasks ranging from writing firmware to merging thousands of pull requests per week; the report did not provide independent verification of this broader capability.

100

Agnes Harness open-sources dual-system agent architecture

Agnes Harness is an Apache 2.0-licensed open-source agent system that pairs a 9B Agnes Sys1.0 model for fast decisions with a large language model for complex reasoning. Its test-time training and recursive self-improvement mechanisms adapt to new tasks and turn validated procedures into reusable Playbooks. Plugins, MCP connectivity and the Model Hardware Standard are designed to integrate software systems and physical devices. Developer-reported tests cited by QbitAI showed up to 14.3 times faster decisions, about one-twenty-seventh of the per-call cost and a 32.9% reduction in total task time in a robot-dog simulation. No independent evaluation was identified in the supplied report.

113

cheap-watts open-sources AI-assisted home energy control

Developer Romin Irani open-sourced cheap-watts on GitHub, using Google Antigravity with Gemini 4 Hydrogen and Gemini 4 Carbon to coordinate rooftop solar, a home battery, two vehicles, and dynamic electricity tariffs. The project includes offline tests across more than 25 commits, hardware abstractions, CI safety checks, and a dry-run mode that queries live inputs without writing to the inverter. Irani reports that the models handled requirements, coding, and testing while he focused on architecture and reviewing changes.

123

Ant inclusionAI open-sources LingSante-Pulse-1.0 encoder

Ant inclusionAI released the MIT-licensed LingSante-Pulse-1.0 physiological-signal encoder on Hugging Face, according to AIbase AI新闻. The separate lingsante_pulse package applies fixed 128-Hz preprocessing, handles observation and storage-padding masks, and supports temporal, channel-temporal, and batched inputs. The model returns CLS and observed-valid PatchMean features shaped [B, C*256], with optional ordered patch features and 0.25-second validity masks, but includes no prediction head for specific downstream tasks.

133

FreeVideo v0.5.0 adds local H3 image generation

FreeVideo v0.5.0 adds local MiniMax H3 image generation for images up to 16 megapixels, according to the project’s release notes. The open-source local inference engine runs as a ComfyUI plugin and supports text prompts, first and last frames, and image, video and audio references. The project says FreeVideo can run MiniMax H3 on consumer GPUs with as little as 8 GB of VRAM and 16 GB of RAM. Its code is licensed under Apache License 2.0, while the model weights use the MiniMax H3 Community License, which includes territorial and acceptable-use restrictions.

143

TorchTitan unifies PyTorch-native generative AI training workflows

TorchTitan is a PyTorch-native platform for pretraining and post-training generative AI models, spanning supervised fine-tuning and reinforcement learning. Its documentation says FSDP2, tensor, pipeline, context, and expert parallelism can be composed through DeviceMesh and DTensor, while TitanRL shares model definitions and kernels between the trainer and vLLM generation. The supplied report is a project-maintained repository snapshot and does not independently validate the stated performance.

153

Ant’s OceanBase team open-sources PowerContext agent memory service

Ant’s OceanBase team has open-sourced PowerContext, a context service designed for AI agents. It moves durable information such as decisions, constraints, evidence, and progress outside individual conversations so later sessions, agents, or people can continue from the current state.

160

Tencent releases WeMM-Video-2B GGUF multimodal bundle

Tencent published a GGUF Q4_K_M bundle for WeMM-Video-2B, providing language, vision, and audio components for the WeMM native runtime. The package supports text, image, audio, video, and combined video-audio prompts under the Apache 2.0 license. Tencent says the CPU runtime loaded the bundle and passed generation checks across those input modes.

170

Tencent releases WeVisForensics-4B AI video detector

Tencent’s WeVisForensics repository releases five LoRA adapters for WeVisForensics-4B, built on Qwen3.5-4B and trained with DTPO. According to the repository, a main investigator agent coordinates inspector sub-agents that examine visual appearance, physical consistency, common sense, on-screen text, speech content, and voice profiles. Deployment options include merged models for multi-GPU use, five adapters served through vLLM on one GPU, and an in-process Transformers demo requiring at least 24 GB of GPU memory.

180

LiberNT adds a new kernel with AI-assisted development

Tom’s Hardware reports that the ReactOS-derived LiberNT project has added a new bootloader, graphics stack, memory manager, and kernel. The GPLv3-licensed operating system targets amd64, arm64, and riscv64 platforms, with its source code and ready-made bootable disk images available. Developers say AI helps review and write code, while a person reviews the output and validates end-to-end tests before it enters the main branch.

190

Sellbird 0.4.0 Launches as an Open-Source AI Outreach Platform

Sellbird 0.4.0 is the new name for the open-source outreach platform previously called Harvey, with existing configurations, data, and dashboard sign-in carried over. According to the project README, it runs locally through Claude Code CLI or Codex, and can find prospects, verify email addresses, draft personalized outreach, send approved emails, and handle replies. A hosted version with sending-domain, inbox, and warmup setup is available in early access.

200

Fuxi Intelligent Decision Platform v1.0.0 Adds Multi-Agent Orchestration

开源中国报道称,伏羲智能决策平台发布 v1.0.0,升级多智能体编排与知识检索,并加入历史回归、业务结果回传、偏差归因、漂移监测和优化评估等流程。该版本据称用于连接自然语言建模、确定性决策执行与发布后运营,平台采用 Rust 技术栈,并支持标准 DMN 决策模型和项目级治理。相关能力目前主要来自该二手报道对项目版本的描述。