Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
OpenAI and Broadcom co-developed Jalapeño, a custom AI inference chip designed to replace general-purpose GPUs for large language model serving. The chip, built in nine months, targets memory and networking bottlenecks i…
A developer built Carbon.Ledger, a full-stack Django web app that logs user activities, computes CO₂ impact using database-stored emission factors, and provides AI-powered reduction tips via NVIDIA NIM's Mistral Large mo…
NVIDIA released NeMo AutoModel, an open-source library that accelerates fine-tuning of Mixture-of-Experts (MoE) transformer models by 3.4-3.7x in training throughput and reduces GPU memory usage by 29-32% compared to nat…
Prompt injection remains the top vulnerability for large language models, according to OWASP's Top 10 for LLM Applications, because it is an architectural property of how LLMs process input rather than a model bug. Unlik…
A developer built RationaleVault, a memory platform designed for cognitive continuity rather than simple document retrieval, after finding that traditional RAG systems struggle to help AI continue work. RationaleVault pr…
The Motion Mind Team has launched LibPilot, an AI-powered search engine and discovery hub for NPM packages. It indexes over 4,000,000 packages in real time, offering intent-based discovery, continuous AI chat, and granul…
Vanta's Staff Software Engineer Noam Rubin outlines the company's agent development principles, emphasizing betting on model intelligence, enabling composition for emergent behaviors, and building strong verifiers to boo…
Anthropic engineers have demonstrated that iterative agent loops, which allow AI models to self-correct through repeated prompting, significantly improve task performance but increase token costs by up to 15x compared to…
OpenAI and Broadcom unveiled Jalapeno, an ASIC accelerator for LLM inference, on June 24, 2026. The chip, co-developed in nine months, is claimed to deliver substantially better performance per watt than current state-of…
Z.ai benchmarked GLM-5.2 against Claude Opus on 45 coding tasks using an identical agent setup, finding both models solved 25 tasks and agreed on 43 of 45 outcomes. With prompt caching enabled, GLM-5.2 cost roughly 46% o…
Apple researchers discovered that a single parameter, termed a 'super weight,' can disproportionately determine an LLM's behavior, with removal causing catastrophic quality loss. The finding enables efficient model compr…
Apple runs a small language model on iPhones in about a gigabyte of memory and a larger model on its cloud servers, reflecting a split also seen at Google, Microsoft, and Meta. Small models (0.5B-14B parameters) target d…
A SimilarWeb study found that users who saw a brand recommended by ChatGPT were 2.5 times more likely to visit that brand's website within a week compared to a competitor, with AI-influenced visitors viewing 12 pages and…
Anthropic and Google Cloud are hosting a technical session on June 24 at 10 AM PT to address security and monitoring challenges for autonomous agents built with Claude. The session, led by Roy Arsan and Ivan Nardini, wil…
Zhipu's GLM 5.2 open-weight model, available on Synthorai at roughly one-sixth of frontier per-token prices, achieves frontier-level benchmarks but its per-task cost varies by over an order of magnitude depending on the …
Christof, a software engineer at Sovaro Consulting, shipped an llms.txt file for client California Metals and initially attributed a rise in Google AI Mode visibility to it. However, Google's AI optimization guide states…
The Trump administration has grown frustrated with Anthropic CEO Dario Amodei and replaced him with cofounder Tom Brown in meetings about re-releasing the Claude Fable 5 AI model, which remains under export controls due …
AWS launched Kiro, a spec-driven AI coding IDE at Summit New York, which generates structured requirements, design, and task documents from prompts before writing code to prevent technical debt from vibe coding. The tool…
A developer detailed a production-grade setup for streaming Claude's LLM tokens to a browser using Server-Sent Events in a Next.js route handler, emphasizing the critical `X-Accel-Buffering: no` header to prevent nginx b…
Haystack, an open-source AI orchestration framework by deepset, uses explicit directed acyclic graphs (DAGs) instead of implicit chains to build production-ready RAG systems. Its modular, serializable architecture enable…