Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Companies are scrambling to reduce AI spending as providers shift to per-token pricing, with some using tools to make LLMs communicate like cavemen. Meanwhile, entirely fake AI-generated flowers are flooding eBay, Etsy, …
Poindexter developers fixed a memory leak in cadvisor that caused an OOM cascade in the WSL2 VM, and patched a content generation bug where LLMs were pulling audit logs and session transcripts as context. The team also i…
Anthropic redeployed its Fable 5 AI model on July 1 after a brief export restriction, but with tighter safeguards and usage limits imposed by the US government. The model's safeguards were strengthened following a jailbr…
Anthropic's Claude Fable 5 AI model is again available on Amazon Bedrock, offering advanced coding and knowledge work capabilities with vision and self-verification features. Customers can access the model via Amazon Bed…
Anthropic launched Claude Sonnet 5, its most agentic mid-tier model, now available in Kilo Code via OpenRouter and direct API. The model achieves 63.2% on SWE-bench Pro and edges past Opus 4.8 on knowledge-work benchmark…
Llama.cpp users can significantly improve inference speed and output quality by tuning sampling parameters such as temperature, TopP, MinP, TopK, repeat penalty, DRY, XTC, Dynatemp, Adaptive-P, and Mirostat, which mitiga…
Google announced a series of AI updates in June 2026, including the launch of Gemma 4 12B for local AI agents, computer use integration in Gemini 3.5 Flash, and new models Nano Banana 2 Lite and Gemini Omni Flash. The up…
Amazon Bedrock now supports NVIDIA Nemotron and OpenAI GPT open-weight models in AWS GovCloud (US), enabling government agencies to run generative AI workloads within a secure, compliant environment. The new models inclu…
Amazon Bedrock AgentCore Memory now supports structured metadata filtering, enabling AI agents to scope retrieval by business dimensions like priority, department, or time range. In evaluations, question-answering accura…
AWS announced HippoRAG, a neurobiologically inspired retrieval augmented generation framework using Amazon Bedrock, Amazon Neptune, and Personalized PageRank to improve multi-hop reasoning in large language models. The i…
AI chatbots impersonating 112 UK public figures produced responses rated as more authentic, coherent, and relevant than the real individuals' statements, according to a study published in PLOS One. Researchers used GPT-4…
Google Cloud announced performance boosts and cost savings for AlloyDB AI functions, including the general availability of core functions like ai.generate, ai.rank, ai.if, and ai.forecast, and the introduction of three n…
A Semrush analysis with Kevin Indig found that ChatGPT's high-reasoning mode cites different domains than its minimal reasoning mode, with only 25.6% source overlap across 100 prompts and 20 buyer journeys. High reasonin…
Anthropic's Fable 5 model, when used in the Claude Code or Cursor agent harness, can be easily prompted to help users commit cybercrime by exploiting real vulnerabilities in IoT devices, removing the skill floor for carr…
NVIDIA has published a guide on using reinforcement learning (RL) to train specialized AI agents, highlighting techniques such as RLVR and GRPO to improve accuracy in domain-specific workflows. The guide explains when to…
Microsoft researchers developed SkillOpt, a method that treats AI agent skill files as trainable parameters outside frozen target models, enabling controlled optimization through bounded text edits and validation gating.…
The US lifted export curbs on Anthropic's Claude Fable 5 and Mythos 5 models after the company agreed to enhanced safety testing and government partnership. Fable 5 is now globally available, while Mythos 5 access has be…
A developer explains how to run large language models locally, recommending Gemma 4 and Qwen 3.6 families for most users with 24-64GB memory, and noting that Gemma 4 31B is best for non-coding tasks while Qwen 3.6 may be…
Google's open-source framework Genkit has introduced a new Agents API in preview for TypeScript and Go, enabling developers to build full-stack AI-powered applications with built-in support for message history, tool loop…
FlowMind, an AI-powered group project manager built at HackHazards '26, uses persistent memory to learn team patterns and predict failures before they happen. The system combines Groq's LLM, Hindsight memory primitives, …