Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Local LLMs have improved sharply, with Simon P. Couch, senior software engineer at Posit, reporting that the April Qwen 3.5 and Gemma 4 releases each scored 90% on his agentic coding eval, up from 0% for any MacBook-runn…
Elon Musk, CEO of SpaceXAI, said on X that training for Grok 4.8 will complete this week, Grok 4.9 will be 'Astra/Fable class,' and Grok 5 will achieve AGI, while simultaneously endorsing Anthropic CEO Dario Amodei's cal…
A developer has documented the architecture for an AI voice sales agent that lets dealers place product orders through a phone conversation, combining Twilio telephony, the OpenAI Realtime API, and LangGraph for stateful…
A developer outlined a split-architecture approach for turning text prompts into manufacturable printed circuit boards, arguing that large language models should handle component selection and connectivity decisions whil…
A developer outlines a three-pillar engineering approach to managing generative AI hallucinations in production, combining self-correction architectures, standardized metrics like faithfulness and answer correctness, and…
Recent reinforcement learning advances for Mixture-of-Experts (MoE) large language models have focused on improving optimization stability and training efficiency while treating expert selection as a fixed component, acc…
The Open Source Initiative's Open Source AI Definition (OSAID 1.0) is under dispute because it requires model parameters including weights to be released under OSI-approved terms but does not mandate disclosure of traini…
A developer has published a role-based guide to building MCP agents, arguing that every useful agent consists of four parts: a model, two or three MCP servers, a system prompt, and a write guardrail. The guide maps speci…
Google Threat Intelligence Group (GTIG) reported that state-affiliated cyberespionage groups and cybercrime gangs are targeting enterprise AI assets — including proprietary models, model weights, source code, API credent…
As of August 2, 2026, the EU AI Act's enforcement regime for general-purpose AI obligations took effect, imposing penalties of up to 15 million euros or 3% of global annual turnover on non-compliant model providers. The …
A seven-stage benchmark comparing Claude (Opus 5 Medium) and ChatGPT (5.6 Sol Medium) on a live Zendesk ticket dataset via the Geckoboard Metrics MCP server found the two AI assistants performed differently on a realisti…
Fyxer's executive-assistant agent uses OpenAI models combined with fine-tuning, persistent memory, and a real user-feedback loop to sort inboxes and draft email replies in each user's own voice, according to OpenAI. The …
Researchers introduced Grouped Value Attention, a method that reduces the Transformer KV cache by reconstructing keys on demand rather than storing them at every decoding step. The technique builds on grouped-query atten…
Everpure is positioning its FlashBlade//S and FlashBlade//EXA storage platforms as the fix for GPU starvation in large-scale AI deployments, with FlashBlade//EXA delivering 400 to 450 million IOPS, 220 GB/sec bandwidth a…
A Medium post in the "AI in Plain English" series argues that BM25 keyword ranking remains essential to retrieval-augmented generation (RAG) systems even as embeddings draw widespread attention. The author walks through …
InclusionAI released Ling-3.0-flash-VL, a hybrid instant/reasoning model with tool calling that builds on Ling 3.0 Flash (124B total / 5.5B active MoE) and adds native visual perception and advanced visual agent capabili…
Quixotic AI released jinfer, an early CPU-only inference engine for the JVM that supports chat, vision, audio, embeddings, reranking, and TTS without a Python runtime, ONNX, Docker containers, sidecar processes, or IPC. …
A practical guide for founders and product teams outlines how to build production-ready custom AI agents in 2026, covering architecture, orchestration, tools, retrieval-augmented generation, memory, and guardrails. It ad…
A developer's cost breakdown argues that teams should default to retrieval-augmented generation when a model lacks business knowledge and to prompt engineering when it answers in the wrong style, reserving fine-tuning fo…
Researchers Xianming Li, Zongxi Li and Tsz-fung Andrew Lee merged ShadowPEFT into the main branch of Hugging Face's PEFT library, adding a stateful shadow network that can be detached with unload_shadow() and run as a sm…