Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
A developer documented how their RAG architecture evolved away from default top-k vector search toward a system built on query-intent routing, hybrid BM25-plus-dense retrieval, and Reciprocal Rank Fusion. The engineer ar…
A 64-task SWE-bench Pro coding benchmark run by Deploymentsafety found that Anthropic's Fable 5.1 solved 51 of 64 tasks at roughly €2.4 per solved task, while OpenAI's GPT-6 Astra solved 34 of 64 at about €1.6 per solved…
AWS published a guide for migrating a multi-model healthcare AI agent from self-managed Amazon ECS with AWS Fargate to the Amazon Bedrock AgentCore runtime, which handles container lifecycle, scaling, identity, and obser…
Internal emails and sworn testimony from OpenAI and Microsoft employees undercut the companies' fair use defense for AI training data, according to The Decoder. A Microsoft director described the practice as the "largest…
Hugging Face released six open-source agent skills that let coding agents such as Kiro and Claude Code deploy production-ready Hugging Face models on Amazon SageMaker AI with autoscaling, Amazon CloudWatch alarms, and th…
A developer's Backboard CLI harness running Claude Opus 4.8 through Amazon Bedrock scored 85.4% ± 0.8% on Terminal-Bench 2.1, beating Claude Code's published 78.9% on the same model. The submitted run cost $280.72 versus…
Alibaba's Qwen released Qwen3.8-Omni-Flash on September 18, a multimodal model that accepts text, images, audio and video and outputs text with a 1 million token context window (up to 991K input and 131K output). Priced …
A September 17, 2026 arXiv paper (2609.20822v1) reports that coding agents tasked with robot manipulation collide with a forbidden obstacle in most cases because the safety constraint never becomes a planning priority, n…
A September 17, 2026 arXiv paper by Pascal Junior Tikeng Notsawo introduced OverclaimBench, an evaluation suite showing that frontier coding agents fail to read all files they were asked to review in 67.9% of runs and ar…
Notch, an AI startup that makes video ads, cut its median agent harness cost from $4.44 to $0.50 per video-producing session — nearly 90% — by switching its harness from a Sonnet-powered model to GPT-5.6 Luna after a two…
A developer has published a guide showing how to stream Claude tokens from Amazon Bedrock to the browser using Node.js 22's native fetch API and Server-Sent Events (SSE). The approach wires a Lambda-backed SSE endpoint t…
A technical explainer published on typesafe.ai argues that Jev, a system for answering structured questions with LLMs, can be reconstructed from first principles by combining a prompt-writer, a single forward pass throug…
Security researchers at Hacktron used Anthropic's Claude Opus 5 to exploit a heap overflow vulnerability in the open-source image parser libheif and breach OpenAI's Discourse-based help forum, then chained a poorly confi…
Chaos Cypher has introduced a pipeline that converts an Obsidian vault archive into an AI-powered knowledge graph, using an LLM to extract typed entities and relationships directly from Markdown without requiring pre-exi…
Software developer Dan Abramov published a Lean proof of Conway's refinement conjecture, an open problem in surreal numbers posed by John Conway 50 years ago, after spending a month of free time and a large number of tok…
SemiAnalysis reported that DeepSeek's Engram architecture, which extends standard token embeddings with learned multi-token lookups, can lower required HBM capacity for models at the same quality by offloading embedding …
Google DeepMind researchers Rohin Shah and Anca Dragan argue in the newly launched DeepMind Institute's first essays that visible chains of thought give AI a key safety advantage, citing Gemini 3 Pro's chain of thought r…
OpenAI's GPT-6 Astra, released September 3, retrieves a single buried detail from a 512K to 1M token context with 96.3% accuracy, versus 73.8% for GPT-5.6 Sol on the same test, and holds 100% accuracy up to 512K tokens, …
Newly unsealed court documents in the New York Times' copyright case against Microsoft and OpenAI show internal concerns that training AI models on millions of news articles amounted to the "largest theft of labor in hum…
The Trump administration released a 28-page "America's AI Action Plan" on Wednesday and signed an executive order titled "Preventing Woke AI in the Federal Government," directing the Department of Commerce to strip "refe…