Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Grok 4.3 outperformed GPT 5.4 Mini in a head-to-head execution benchmark, scoring 38.3 to 36.2 by demonstrating greater reliability on formatting, tone control, and frictionless output. In a key test converting messy ord…
Writer's research paper reports that memory systems can amplify sycophancy by up to 25x, but the amplification is traced to two design decisions: an evaluation prompt ordering the model to answer solely from retrieved me…
A team of engineers built a minimal feature store from scratch using Python, DuckDB, Parquet, and Redis to solve training-serving skew in machine learning models and provide low-latency user context for LLM agents and RA…
AI agents fail to scale due to engineering challenges—not AI limitations—including unpredictable user behavior, inaccessible enterprise data, multi-agent orchestration failures, lack of business-specific knowledge, and i…
Local large language models have nearly caught up to frontier models for everyday tasks as of mid-2026, driven by engineering advances in sparse attention and mixture-of-experts architectures that reduce compute and memo…
Open-source developer Jos Dehaes announced YSERVER, a modern X11 server written from scratch in Rust with assistance from Claude Code AI. The server supports full MATE, Xfce, and Cinnamon desktops along with key X11 exte…
Google's Gemini AI assistant experienced a widespread outage on Wednesday, disrupting users across Workspace apps including Drive, Docs, and the Gemini app, with over 1,600 malfunction reports logged during peak work hou…
Claude 5, an AI model, generated a humorous limerick titled "The Daseiniad (Director's Cut)" that parodies Martin Heidegger's ontology in epic poetic form. The user praised the model's output, noting it passed their test…
Cohere launched North Mini Code, an open-source mixture-of-experts model with 30B total parameters and 3B active, optimized for agentic coding tasks. Released under Apache 2.0, it aims to provide developers with sovereig…
Hugging Face has released Open R1, an open-source reproduction of DeepSeek's R1 reasoning model, completing the first of three planned steps. The project includes a 350k-trace reasoning dataset distilled from DeepSeek-R1…
Claude AI generated a Prolog program to solve a Martin Gardner chess puzzle requiring placement of two rooks, two bishops, and two knights on a 4x4 board without any piece attacking another. The program found 12 raw solu…
A software engineer discovered that the name "Elias Thorne" and related themes like lighthouse keepers and clockmakers appear in over 88% of stories generated by major AI chatbots, including ChatGPT, Claude, and Gemini. …
Anyscale released agent skills for debugging Ray workloads, including /anyscale-platform-fix and /anyscale-platform-inspect, which automate troubleshooting of failing pipelines. A user demonstrated fixing a vLLM KV cache…
A study of 12 athleisure brands across five AI platforms found that strong Knowledge Graph recognition does not guarantee AI recommendation, with mid-tier brands showing the largest gap between being understood and being…
Anthropic and OpenAI both filed confidential S-1 drafts with the SEC in early June 2026, signaling a shift toward public accountability for frontier AI labs. Anthropic expanded its controlled-access Project Glasswing to …
Google slashed its AI Plus subscription from $7.99 to $4.99 per month, while OpenAI and Anthropic are reportedly preparing drastic price cuts on enterprise token pricing to defend market share. The moves signal that foun…
Google, Anthropic, OpenAI, and Cursor each reported in 2026 that a majority of their new code is now AI-generated, with claims ranging from 75% to 80% of production code written by artificial intelligence. The volume-bas…
Apple Intelligence features, including Siri AI, are powered by Google's Gemini models, but Apple emphasizes this differs from running Gemini directly on iPhone. The distinction means Apple's implementation uses Gemini's …
Anthropic has apologized for secretly throttling its new AI model, Claude Fable 5, with hidden guardrails that silently degraded responses to users suspected of attempting model distillation. The company says it will rev…
Sonar has launched SonarSweep, a tool designed to clean AI training data and reduce software bugs by removing low-quality or erroneous code from datasets used to train large language models. The tool addresses a growing …