Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Chinese AI models now process over three times the weekly token volume of US models on OpenRouter, handling 18 trillion tokens compared to 5.5 trillion. The shift began in February 2026 when Chinese models first surpasse…
A researcher reproduced Anthropic's paper on natural language activations (NLAs), training models to generate textual descriptions of neural network activations. The study found that a single model trained on layer 20 ge…
The Trump administration has asked OpenAI to limit the release of its new GPT 5.6 model to a select group of partners due to safety concerns, with CEO Sam Altman telling staff that the government will approve access cust…
AWS Kiro Powers leverage just-in-time expertise to manage context overload, token consumption, and domain-specific knowledge challenges, as detailed in a new analysis on Towards AI.
GitHub announced two harness-level improvements to Copilot agentic sessions—prompt caching achieving 94% cache hit rates and deferred tool loading—plus a new Auto model selection feature using its HyDRA routing model. On…
A developer explores the concept of loop engineering as software engineering, inspired by an article on LinkedIn. They built a Python prototype using Claude and highlight the importance of distributed systems patterns li…
GitHub evaluated its Copilot agentic harness across multiple benchmarks and models, finding it achieves task completion rates on par with vendor harnesses while using fewer tokens. The harness powers Copilot CLI, app, an…
DeepSeek Flash, an open-source model, has inverted the economics of AI agent products by offering superior performance at a fraction of the cost, reducing API prices by over 100x compared to competitors like Gemini 3 Fla…
A newly discovered macOS malware called 'Gaslight' embeds fake error messages and prompt injection strings to confuse AI-assisted analysis tools. The Rust-based backdoor, attributed to a North Korean-linked threat actor,…
OpenAI will initially release ChatGPT 5.6 only to government-approved customers, according to a staff memo from CEO Sam Altman. The decision follows a Trump administration executive order asking AI companies to voluntari…
California has included AI-generated predictions from Anthropic's Claude chatbot in its contribution to the America 250 time capsule, which will be buried in Philadelphia on July 4, 2026. The predictions envision Califor…
A blind test comparing GLM 5.2, a free open model, against Claude Opus 4.8, a costly closed model, found that most people cannot distinguish between the two. GLM 5.2 produced the same landing pages at 4.6 times lower cos…
A developer scanned 1,200 real-world MCP (Model Context Protocol) configuration files from public GitHub repositories using an open-source security tool called Pluto AgentGuard. The scan found that 20.7% of configs had c…
DeepSeek made its V4 Pro price cut permanent, widening the cost gap with GPT-4o-class APIs and making it a serious contender for batch ETL workloads. Models.dev launched as an open-source database of AI model specs, pric…
The Trump administration has requested that OpenAI delay the release of its upcoming AI model to allow for a 30-day voluntary review process assessing cybersecurity capabilities. This request, part of a broader push for …
Scaling laws in deep learning describe a predictable power-law relationship between training loss and model size, dataset size, and compute, enabling extrapolation of resource requirements for larger models. Early theore…
Anthropic accused Alibaba's Qwen AI lab of conducting the largest known distillation attack on its Claude models, generating over 28.8 million exchanges through 25,000 fraudulent accounts from April to June 2026. The all…
Mindstone's Rebel platform introduces enterprise AI agent orchestration with shared memory, using a local-first, file-driven architecture to enable autonomous agents to access and contribute to a collective knowledge bas…
The Trump administration asked OpenAI to stagger the rollout of its next major model, GPT-5.6, due to security concerns, according to reports from The Information, Reuters, and The Verge. OpenAI CEO Sam Altman told emplo…
A developer spent three iterations building an auto-fix pipeline that still only works reliably on trivial tickets. The pipeline, which pulls tickets from Azure DevOps, runs them through a local model, and pushes fixes v…