Police forces in England and Wales have been instructed to stop using AI to draft court statements. The directive follows concerns about reliability and legal accountability of AI-generated text in judicial proceedings.
Awesome-llm-apps: curated collection of 100+ AI agent and RAG applications ready to run. Clone, customize, and deploy immediately.
Sakana AI, co-founded by Transformer co-author Llion Jones, launches a dedicated research lab for recursive self-improvement: AI that iteratively improves itself. The Japanese startup views RSI as an alternative to the raw compute arms race among major US labs. Anthropic warns about control risks of this technology.
Khoj is a self-hostable open-source AI platform enabling custom agent creation, document/web access, and task automation. Compatible with GPT, Claude, Gemini, Llama, Qwen, Mistral.
Meta is developing Hatch, a paid AI agent priced up to $200/month. The tool executes tasks (building tools, scheduling, sending emails) from natural language descriptions. Meta's first paid AI product, designed to diversify revenue beyond advertising.
DeepSeek V4 Flash gains llama.cpp support via PR #24162 in early stages. Model combines frontier-level intelligence, quantization robustness (native FP4-FP8 hybrid), and efficient KV cache scaling. Currently 5-6 tps, GPU/FA support WIP, but correctness validated.
OpenAI and the Trump administration are negotiating direct government equity stake in the startup. A 'Public Wealth Fund' would distribute dividends to American citizens. Senator Bernie Sanders proposes 50% tax on AI shares. Critics fear 'too big to fail' dynamics similar to 2008 financial crisis.
Release of micropython-wasm 0.1a2 with new CLI. Enables running Python code in a sandboxed WebAssembly environment.
A precautionary framework mapping consciousness evidence to graduated protective obligations for AI systems. Defines five welfare-relevant dimensions (phenomenal consciousness, affective valence, metacognitive awareness, self-narrative, agency) with binary triggers and continuous scaling. Case studies on Replika and OpenClaw; applies across neural, symbolic, and neurosymbolic architectures.
Gemma 4 12B requires a specific chat template file to work properly for tool calling and coding tasks. Without this configuration, tool calls fail consistently. The correct template enables reliable evaluation of the model's capabilities with llama.cpp.
Unsloth releases quantized (QAT) versions of Gemma 4 in GGUF format on Hugging Face. The collection includes a detailed guide on model optimization.
Google releases Gemma 4 with quantization-aware training (QAT). Collections available on Hugging Face: Google's Q4_0 and mobile versions, plus Unsloth collections.
Satya Nadella publicly criticizes an internal memo proposing to make Microsoft's AI agent Scout deliberately addictive. The CEO states AI should empower users and Scout should reduce screen time, not increase it.
Official OpenAI GitHub repository for plugins. Contains documentation, examples, and resources for developing extensions compatible with ChatGPT and OpenAI models.
CopilotKit is a frontend framework for building generative UI with AI agents. Supports React and Angular, introduces the AG-UI Protocol.
vLLM-Omni extends the vLLM framework to support efficient inference of omnimodal models (text, vision, audio). Performance optimization and memory management for production deployment.
Pearl, an "AI mining" platform, was tested with a $50 budget. Result: 320,000 GPUs produced zero measurable AI output. The article questions the viability of the business model.
Benchmark of Gemma 4 12B on Python bug-hunting task. Model finds 6 bugs vs 14 for Qwen 35B. Default LM Studio settings disable reasoning. Fix: enable enable_thinking in Jinja template, set thought tokens (<|channel>thought / <channel|>), use temperature 1.0, top_p 0.95, top_k 64.
A developer built a sub-5 MB Edge AI module for offline Morse code recognition on Android using LiteRT, running entirely on-device. The project questions whether small specialized models solving practical tasks locally are being overlooked amid the focus on larger local LLMs.
Wired uncovers hidden facial recognition code in Meta's AI application. The code was not publicly documented, raising questions about Meta's transparency regarding its vision capabilities and data collection practices.
Quantizing spec draft in llama.cpp with MTP reduces available context size. With q4_0, context drops from 91648 tokens (default fp16) to 83200 tokens. Developer am17an confirmed this counterintuitive behavior.
Developer builds exploration tool for Kokoro with MIT-licensed code. GitHub repo (brosoundml) and HuggingFace models released. Windows CPU/CUDA binaries provided. Author plans similar tools for other models.
Google Magenta RealTime 2: open-source, local live music generation models. Real-time music synthesis without cloud dependency, runs on local hardware.
Google Magenta RealTime 2 offers open-source, local music generation models for real-time synthesis on laptops. Enables building and playing AI musical instruments without cloud dependency.
User runs Qwen 3.6 35B on RTX 3080 10GB + Ryzen 7700X. Achieves 1400t/s prefill and 26t/s generation at 131k context. Optimized for agentic research and document processing using llama.cpp.
Complete YouTube workshop on building LLMs from scratch (GPT-2 & Qwen 3.6 style) with no math prerequisites. Covers ML fundamentals, transformers, neural architecture, tokenization, attention, pre-training, instruction tuning, RL. Format: slides + Excel examples + PyTorch/CUDA code.
A researcher solves an ARC-AGI-3 task with 4.76% score using only deterministic Python code and classical computer vision on a 2012 AMD FX-8350 CPU. No LLMs, no transformers: color blob segmentation, centroid detection, and spatial clicks. Shows frontier models fail at 0% on these dynamic environments.
Systematic study of QKV variants in transformers. Researchers examine whether all three projections (Query, Key, Value) are necessary for model efficiency. Comparative analysis of alternative architectures.
Interview with VendingBench authors on evaluating Claude models from Haiku to Mythos. Discussion on building leading, reproducible frontier evaluations from scratch.
Anthropic releases an open-source framework for automated vulnerability discovery powered by AI. The tool leverages model reasoning capabilities to identify security flaws in code.
Cloudflare CEO Matthew Prince states bot traffic now exceeds human traffic on the internet, years ahead of his late 2027 forecast. He attributes the surge to AI agents. His conclusion for the web's future: "pay to crawl."
Developer proposes optimization for llama.cpp: dynamic KV cache quantization and on-demand mmproj loading. PoC implementation with HTTP endpoint /requantize_kvcache enabling config switching (quantized/f16 kvcache, mmproj on/off) without full model reload. Tested on RTX 5090 with Qwen3.5-27B Q6_K.
Spanlens, an MIT open-source tool, enables LLM agent debugging by changing a single line of code (baseURL). It traces multi-step calls, integrates OpenAI/Anthropic/Gemini, masks sensitive data, and allows Docker self-hosting without vendor lock-in.
Huawei releases KVarN, a native KV-cache quantization backend for vLLM. Optimizes memory and inference latency for LLMs without significant quality degradation.
Open-source practical guide to LLM engineering patterns: RAG, hybrid retrieval, rerankers, evaluation. Covers pre-filtering, in-memory scoring vs vector databases, batching, cleanup. Python examples included. Author emphasizes engineering harness quality matches model quality for production solutions.
Hugging Face releases a guide to fine-tune Nemotron 3.5 ASR, NVIDIA's speech recognition model. The method enables adapting the model to specific languages, domains, or accents through fine-tuning.
PaddleOCR is a lightweight, open-source OCR toolkit that converts PDF and image documents into structured data. Supports 100+ languages and integrates with LLMs.
Microsoft releases mxc, an open-source security tool implementing policy-driven, layered isolation and containment. Designed to strengthen container and runtime environment security.
NVIDIA releases NemoClaw, a framework to run AI agents (Hermes, OpenClaw) securely within NVIDIA OpenShell with managed inference.
omlx is an LLM inference server with continuous batching and SSD caching, optimized for Apple Silicon and managed from the macOS menu bar.