Microsoft and Nvidia partner on AI PCs running autonomous agents locally via OpenClaw framework, replacing Copilot+. Dell and Surface will unveil first models at Computex and Build next week.
Anthropic releases a public repository for Agent Skills, reusable components for AI agents. The project enables development and sharing of standardized agent capabilities.
Terence Tao argues AI could introduce division of labor in mathematics for the first time. Currently, researchers master every step alone (problem framing, verification). Tao foresees "industrial mathematics": AI-supported teams replacing lone geniuses, with humans remaining essential for "inspired guesses."
Gryphe releases Pantheon-Reasoning-27B, an uncensored Qwen 3.6 27B model fine-tuned on roleplay data with full reasoning traces. Trained on Pantheon corpus (~28%), Claude Opus 4.6 reasoning traces (~21%), WorldSim narrative data (~16%), and text adventure content (~16%), the model experiments whether reasoning improves roleplay quality. GGUF quantizations available.
CVE-Bench is a benchmark for evaluating LLM agents on real-world vulnerability patches. The study tests models' ability to identify and fix security flaws in existing code.
OpenAI upgrades GPT-5.5 Instant for more natural responses and removes Canvas feature in favor of direct chat integration. Older models o3 and GPT-4.5 will be retired from ChatGPT by August 2026.
Benchmark of Qwen3.6-27B quantizations on HuggingFace (unsloth, mradermacher, IQ4_XS, Ununnilium) from Q8 to Q2. Measured via llama.cpp: KL Divergence and Same Top P Percentage vs BF16 baseline. 8192 token context, KV cache q8_0. Q6-Q8 nearly lossless.
Google fixes bugs in Gemini usage limits where a single Omni video consumed entire quotas. Ultra members now get twice as many video generations, failed requests are no longer charged, and Google plans increased transparency on usage.
MarkItDown API Server wraps Microsoft's official MarkItDown library in a lightweight FastAPI REST server. The tool converts files (PDF, Word, Excel) to Markdown for RAG and LLM pipelines. This release patches security vulnerabilities in Starlette and document parsers.
Llama.cpp B9406 fixes GGML_ASSERT crash in get_rows/mtmd_helper_decode_image_chunk when using MTP + MoE model + vision with Qwen 3.6-35B-A3B.
A review paper argues the real bottleneck for autonomous AI agents isn't the language model but the software layer around it: tools, memory, testing, and permission boundaries turn a stateless model into a working agent. Deepseek is building a dedicated « Harness » team in Beijing confirming this thesis.
Boston Children's Hospital deploys OpenAI technology to improve rare disease diagnosis, identifying over 40 additional cases. The system reduces operational burden and accelerates patient care.
Open-source platform for reproducible world model research and evaluation. Provides standardized infrastructure to train and test world models on simulated environments.
PaddleOCR is a lightweight, multilingual OCR toolkit (100+ languages) designed to convert PDF and image documents into structured data for LLM consumption.
MinerU converts complex documents (PDFs, Office files) into LLM-ready markdown/JSON for agentic workflows. Open-source tool for document extraction and data structuring.
Open-source platform for reproducible world model research and evaluation. Provides standardized infrastructure for training and testing world models on simulated environments.
Cigref estimates €140 billion in annual cloud and software overruns across European organizations. AI bundled into solutions ranks as the second identified cause. Half of CIOs cannot measure ROI from these integrated AI offerings.
Amazon shuts down internal AI leaderboard after employees inflated scores through meaningless tasks, driving up cloud costs.
llama.cpp PR #23764: use f16 masks in Flash Attention to reduce VRAM consumption. Optimization enabling larger models to fit on GPU memory.
StepFun 3.7 Flash benchmark on M5 Max (128 GB) with llama.cpp. Short contexts (<16k tokens) fast and responsive. 32k-64k contexts usable. Detailed metrics: 65k tokens reaches 360.79 t/s token generation.
Analysis of ClinicalTrials.gov registry shows marked increase in AI-related trials over time, with recent growth in machine learning, deep learning, chatbots, GPTs, and LLMs. China and US lead geographically. Hybrid approach using GPT-5.5 and human review: good agreement on non-AI studies, lower agreement on human-AI interaction classification.
llama.cpp release b9387 includes significant AMD/ROCm performance update. Users encouraged to share initial benchmark results.
Lance model optimization for RTX 2080 Ti 22GB on single and dual-GPU setups. Custom operator configurations for Turing architecture, pipeline/tensor parallelism across 44GB combined VRAM, reproducible open-source scripts.
Beginner's guide to PyTorch profiling using torch.profiler. Covers how to measure performance and identify bottlenecks in AI models, with practical examples for newcomers.
Claude CLI >= 2.1.154 introduces "ctx", "msg", and "system" roles for API messages, breaking vLLM compatibility. A one-line patch in vLLM restores compatibility and enables Claude workflows with local models like MiniMax-M2.7.
Official Compound Engineering plugin for Claude Code, Codex, Cursor and other editors. Native integration to enhance development workflows.
Markdown rendering tool with specialized support for fenced code SVG blocks. Renders the image and provides a tab to switch to code view. Accepts raw Markdown, CORS-enabled URLs, or Gists.
VeritasReason is an open-source Python framework adding structured reasoning and provenance layer to AI agents. It provides queryable context graphs, forward-chaining rule engine (YAML), W3C PROV-O provenance, and policy compliance checking. Works with OpenAI, Anthropic, Groq, Ollama.
Google Cloud launches 'AI Threat Defense', a platform automating detection, assessment, and patching of security flaws in enterprise systems. It integrates technologies from acquisitions.
Hugging Face has developed a fully local experience for Reachy Mini, a conversational robot. A blog post details setup and customization for various use cases, including building voice agents without cloud dependency.
Hugging Face adds a "Base only" toggle on its models page to filter base models and exclude fine-tunes and quantizations. Long-requested feature by the community.
MOSS-TTS is an open-source speech and sound generation model family from MOSI.AI and OpenMOSS. It covers stable long-form speech, multi-speaker dialogue, voice design, sound effects, and real-time streaming TTS.
Firecrawl is an open-source tool to search, scrape, and clean web data for AI agents. It automates web scraping and content preparation for model training or inference.
Official Compound Engineering plugin for Claude Code, Codex, Cursor and other editors. Native integration to enhance development workflows.
MOSS-TTS is an open-source speech and sound generation model family from MOSI.AI and OpenMOSS team. It covers stable long-form speech, multi-speaker dialogue, voice/character design, environmental sound effects, and real-time streaming TTS.
AI data centers consume growing amounts of electricity, threatening power grid stability. Existing infrastructure struggles to provide the power required by large-scale models and massive deployments.
Heterogeneous GPU load balancing optimization for Ollama (RTX 5090 + 3090). Custom implementation weights layer distribution by compute power (SMCount × ClockMHz) instead of free memory alone. Result: faster than RTX 5090 standalone, leverages 3090 VRAM without bottlenecking the 5090.
IGADA-IoT proposes an information gap-guided automatic data augmentation framework for IoT sensor energy optimization in wireless sensor networks. The method employs hierarchical multi-generator collaboration scheduling (HMGCS) and joint information gap-model performance evaluation (IGMP-EC). Results: +7.27% average accuracy improvement, +8.67% vs advanced augmentation methods.
Systematic survey on Mixture-of-Experts (MoE) applied to multimodal learning. Analyzes MoE from three perspectives: efficient engine (scalability, redundancy reduction), representation learner (multi-expert alignment), modular adapter (modality imbalance, missing data). Identifies gaps: interpretable routing, expert communication, modality integration, lifelong learning.
Paper proposing an approach to guide AI agent planning with dynamically changing norms in human-AI settings. Contributions: a defeasible calculus for resolving normative conflicts and using such norms as guardrails on plans. Theoretical validation via formal proofs and empirical validation with SocialBot on natural language dialogue tasks.