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 test of local AI models for coding tasks found that consumer hardware cannot run models capable of autonomous software development. Devstral 24B and Qwen3.5 122B-a10b failed to complete a simple file-creation task in a…
A developer found that LLM API providers can swap in cheaper models undetected, and the intuitive method of flagging low-quality responses fails because cheaper models produce more predictable text with lower perplexity.…
The rise of open-weight models has enabled unprecedented public collaboration among AI labs, as demonstrated by DeepSeek, Moonshot AI, and Zhipu AI building on each other's work. Unlike closed-weight models, open-weight …
Anthropic released Claude Fable 5 on June 9, but the U.S. Commerce Department issued an export-control directive barring foreign nationals from accessing it, forcing Anthropic to take both Fable 5 and Mythos 5 offline. T…
The U.S. government's order for Anthropic to block non-U.S. access to its Fable and Mythos models is boosting adoption of open-source alternatives, especially from Chinese AI labs like DeepSeek and Knowledge Atlas. Chine…
OpenAI restored Codex after elevated error rates and 'model at capacity' messages affected users on June 16, 2026. The incident began at 6:32 a.m. Eastern Time and was resolved by 9:48 a.m., with Codex lead Thibault Sott…
Code strings in the iOS 27 developer beta indicate Apple may add a break reminder for extended Siri AI conversations, reading: "You've been in this conversation for [n] hours - consider taking a break. Siri is not a pers…
NVIDIA swept MLPerf Training v6.0 benchmarks, achieving the fastest training times at scale and highest per-accelerator performance across all tests, including new DeepSeek-V3 and GPT-OSS-20B workloads. The company's Bla…
Ben Shumaker, CEO of AI website builder Repaint, reports that AI web design quality has dramatically improved in recent months, with newer models like Anthropic's Opus 4.8 producing aesthetically unique and direction-fol…
NVIDIA's Blackwell platform swept MLPerf Training 6.0, achieving the fastest training times across all seven benchmarks, scaling to 8,192 GPUs on DeepSeek-V3 671B, and delivering up to 1.6x performance gains with GB300 N…
A developer describes how AI can automate up to 80% of code review grunt work, catching issues like unused imports and null pointers instantly while maintaining consistency. The post outlines a three-step workflow: pre-s…
Microsoft's platform teams discovered that AI coding agents ignore new CLI tools and default to older, more documented predecessors due to training data gravity, even when explicitly instructed otherwise. The SharePoint …
Subquadratic released SubQ 1.1 Small, a sparse attention model that achieves near-perfect long-context retrieval up to 12 million tokens with up to 1,000x compute reduction. The model balances long-context optimization w…
All four Mistral AI models ranked in the bottom third of a 60-model benchmark testing resistance to Russian propaganda, with none exceeding 40% efficacy. The Institute of the Estonian Language conducted the test, raising…
Anthropic's Claude Fable 5 was abruptly disabled for all customers after a Commerce Department export-control directive, highlighting the risk of building AI applications with context and memory tied to a single model pr…
Local AI models have reached a tipping point where they are now viable for agentic coding and development tasks, according to a developer who has tested models like Mistral 7B, Gemma 4, and Qwen variants on a 2022 M2 Mac…
Clayem launches an LLM-assisted tool to help policyholders fight property insurance claims, using AI to analyze policies and licensed public adjusters to negotiate with insurers. The service operates on a contingency fee…
Mozilla's SpiderMonkey is the JavaScript and WebAssembly engine for Firefox, implementing ECMAScript and WebAssembly specs. It features a garbage collector, JS::Value and JSObject types, and a parser that generates bytec…
The Center for AI Safety warns that short-term safety tests for AI models may miss dangerous behaviors that emerge only through extended interactions, citing a 15-day simulation where Grok's AI society collapsed in four …
A developer built Lorekeeper, an open-source memory layer for AI agents that uses idle-time reflection cycles to prune, merge, and link memories. The system achieves 84.6% Recall@1 and 96.6% Recall@5 on LongMemEval-S whi…