Older news from the AI Haven archive.
Google Research announces TurboQuant, a memory compression algorithm that reduces LLM KV cache size by 6x with zero accuracy loss and up to 8x speedup on H100 GPUs.
ATLAS enables a frozen 14B model to achieve 74.6% on LiveCodeBench using a single $500 GPU, outperforming Claude Sonnet 4.5 at 71.4%.
OpenAI shuts down Sora video app 8 months after launch, citing resource constraints and declining user interest after 1M downloads.
Anthropic launches auto mode for Claude Code, letting the AI autonomously approve its own actions without constant user permission. The feature uses built-in AI safeguards to evaluate safety risks in real-time.
A practical guide to running local LLMs with Ollama in 2026. Covers installation, Python integration, tool calling, and OpenAI-compatible API setup.
Warren calls the Pentagon's decision to blacklist Anthropic "retaliation" for the AI company's ethical stance on AI development, demanding documentation by April 6.
USCC warns China's open-source AI models now lead global usage, surpassing Meta's Llama in downloads and powering 80% of US AI startups.
Gimlet Labs raised $80M to build multi-chip AI inference software that runs across NVIDIA, AMD, Intel, ARM, Cerebras and d-Matrix—delivering 3-10x faster performance.
Cursor AI admits Composer 2 coding model was built on Moonshot AI's Kimi 2.5 without initial disclosure, sparking transparency questions.
California's SB 243 law regulating companion chatbots is now enforceable as of 2026, requiring AI girlfriend apps to disclose their non-human nature and implement suicide prevention protocols.
OpenClaw agents now get automated persistent memory through Hindsight integration, solving context retention for autonomous AI agents.
OpenAI releases GPT-5.4 mini and nano with 54.4% on SWE-Bench Pro, nearly matching flagship performance at half the cost.
Build a fully private, cost-free RAG system that runs on your own machine using Ollama and LangChain. Step-by-step guide with code.
arXiv, the pioneering preprint server hosting nearly 3 million papers, will become an independent nonprofit on July 1, 2026 after 35 years at Cornell, driven by $6.7M annual costs and AI-generated "slop."
NVIDIA releases Nemotron-Cascade-2-30B-A3B, an open-weight 30B MoE model outperforming Qwen3.5 on coding and math benchmarks with only 3B activated parameters.
WordPress.com now lets AI agents write, edit, and publish posts directly. Posts default to drafts requiring approval.
Anthropic submitted sworn declarations to federal court refuting Pentagon claims that the AI company poses an unacceptable national security risk, revealing negotiators said sides were "nearly aligned" just days after Trump ended ties.
WordPress.com now lets AI agents write and publish posts directly on websites via MCP, raising concerns about machine-generated content proliferation and security risks.
California's AI companion chatbot law took effect January 1, 2026. Weeks later, Chat & Ask AI exposed 300 million messages—including suicide queries and illegal content.
Nvidia launches NemoClaw, an enterprise AI agent platform built on OpenClaw, addressing security and privacy concerns as adoption accelerates.
