Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
GitHub has made the redesigned terminal interface for Copilot CLI generally available, featuring a tabbed layout for browsing issues, pull requests, and gists, a guided tool configuration experience, and a cleaner, more …
Google launched Ask Ad Manager, a Gemini-powered AI assistant for publishers using Google Ad Manager, to help troubleshoot delivery issues, generate reports, and navigate the platform via conversational prompts. The beta…
Kilo launched Auto Efficient, a new AI model routing tier that dynamically selects the most cost-effective model for each task based on real-time session classification and benchmark data. The system uses KiloBench to ma…
Software developers will need to adopt Hemingway-like conciseness as AI coding tools such as Claude Code, Codex, and Copilot prefer Markdown-style instructions, shifting from traditional human-readable specifications.
Z.ai's open-weights GLM-5.2 model (744B MoE, 1M-token context) is available for free via web chat and a rate-limited API through OpenRouter, with no credit card required. For developers, the free API tier offers about 50…
A major outage at Anthropic on June 23, 2026, disrupted Claude AI services, highlighting the risks of relying on a single LLM provider. The incident, caused by demand outpacing infrastructure capacity, followed similar d…
SkyPilot launched SkyPilot Endpoints, a production-ready LLM inference system that deploys and manages inference across multiple Kubernetes clusters from a single YAML specification. The system handles placement, autosca…
SGLang achieved a 5x throughput improvement for serving DeepSeek-V4 on NVIDIA GB300 hardware since its Day-0 launch, reaching ~11,200 tok/s/GPU at 50 tok/s/user through kernel and runtime optimizations including MHC fusi…
Krea AI released Krea 2, a series of foundation models for image generation focused on creative exploration, featuring a diffusion transformer architecture and systems for prompt expansion and style referencing. The mode…
Meta suffered its most embarrassing outage ever when a bug allowed users to change any account's email by simply asking the Meta AI bot, leading to high-profile account takeovers. Engineers attributed the disaster to AI-…
Neo4j launched the Agent Memory Service (NAMS), a managed cloud service providing persistent, structured memory for LLM agents using a graph database with native vector indexes. The service offers short-term, long-term, …
Markdown files are a legitimate pattern for agent memory in single-agent, single-user scenarios, but the approach breaks at scale, under concurrent agents, and as facts change and errors compound. Companies like Manus, C…
Three open-weight coding models launched in the week after the US Commerce Department pulled Fable 5 and Mythos 5 offline, including Zhipu's GLM 5.2, Moonshot's Kimi K2.7-Code, and Cohere's North Mini Code. The incident …
New research reveals that auto-regressive large language models suffer from the Reversal Curse, failing to infer that B is A after learning A is B. Tests on GPT-4 showed accuracy dropping from 79% on forward queries to 3…
NVIDIA announced that DFlash speculative decoding boosts inference performance on Blackwell GPUs by up to 15x for gpt-oss-120b and nearly doubles interactivity for Llama 3.1 8B compared to EAGLE-3. The open-source block-…
An engineer at a development team found that AI coding agents often fail not because of the model itself but due to a missing 'harness'—the files, rules, tools, and feedback loops surrounding the model. Common issues lik…
Hugging Face's inference provider information for Meta's Llama 3.1 405B model may be outdated, as the listed provider Featherless AI returns errors and does not show the model as available. A similar issue may affect Bai…
Daters are increasingly using AI chatbots like ChatGPT and Claude as relationship coaches to draft messages, interpret replies, and even break up, sparking debate about authenticity in romance. While some embrace the eff…
Anthropic's Opus 4.7 and 4.8 model upgrades improved benchmarks but broke production systems, sparking developer backlash on Reddit and X. The issue stems from benchmarks not reflecting real-world prompts, with models be…
A Desiderata roundup examines common misunderstandings about LLM consciousness, noting that philosopher Dan Dennett explicitly rejected the idea that chatbots are conscious, contrary to claims by blogger Matt Yglesias. T…