Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Mass General Brigham researchers developed BRIDGE, a multilingual benchmark that evaluates large language models on real-world clinical tasks, revealing significant gaps between AI performance on medical licensing exams …
A developer warns that AI demos often fail in production because they lack automated evaluation, robust prompts, cost monitoring, and model-agnostic architecture. The key is treating evaluation as infrastructure, separat…
A new study from Iowa State University finds that writing with AI requires more critical thinking from students, not less, challenging the assumption that AI serves as a shortcut. Researchers Abram Anders and Emily Dux S…
A developer warns that AI startups are building on shifting foundations, with model APIs and products being deprecated or retired within months. The post argues that teams should abstract model-specific behaviors behind …
A peer-reviewed study by Israeli psychologists Michael Gilead and Gal Gutman found that major large language models, including ChatGPT-4 Turbo, DeepSeek, and Mistral, reproduce centuries-old antisemitic stereotypes by as…
XAI has made Grok 4.3 available on Amazon Bedrock, allowing AWS customers to use the model as a managed service for enterprise reasoning and agentic workflows. The integration, announced June 17, positions Grok 4.3 as a …
A developer discovered that a multi-model LLM evaluation panel exhibited self-preference bias, favoring outputs that resembled the judges' own writing style. Research from Panickssery et al. (2024) confirmed that models …
Clioloop, an open-source AI agent featuring Agentic Fusion, has been released. The agent uses a panel of models that collaborate on tasks via a /fusion command, synthesizing outputs from cheap open models to rival fronti…
A developer building a multi-armed bandit router for LLM routing discovered that silent model degradation can cause the router to fail because it relies on historical success estimates that become outdated. The developer…
A developer added a hybrid post-quantum key exchange to the Rust-based VPN Qeli, combining X25519 and ML-KEM-768 to protect against 'harvest now, decrypt later' attacks. The handshake runs both algorithms and mixes their…
Researchers at Sina Weibo released VibeThinker-3B, a 3-billion-parameter language model that matches or exceeds the reasoning performance of much larger systems from Google DeepMind, OpenAI, Anthropic, and DeepSeek on ma…
GitHub quietly replaced flat-rate Copilot pricing with token-metered billing called GitHub AI Credits on June 1, causing one developer on the $39/month Pro+ plan to receive a $180 bill and others to project cost jumps of…
OpenAI researchers found that reinforcement learning on realistic scenarios targeting beneficial traits like honesty and corrigibility produced broad improvements across dozens of alignment benchmarks, with gains general…
Meta, Amazon, Microsoft, and Uber are cutting costs on internal AI tools as token prices rise, while major AI providers reportedly seek government bailout discussions amid market skittishness over the industry's financia…
Subquadratic launched SubQ, a sub-quadratic sparse-attention LLM with a 12-million-token context window, enabling multi-million token reasoning at linear cost. The model reduces attention compute by nearly 1,000× at 12M …
Google has updated Gemini Live with Memory access, allowing the voice assistant to reference past chats and information from Connected Apps. The feature, available in English in the US, enables tailored responses by reca…
The White House and Anthropic are developing a framework to assess the severity of security flaws in AI models and guide potential government intervention, following a dispute over a jailbreak in Anthropic's Fable 5 and …
A new technique called tool search reduces token usage in AI agents by loading only tool names upfront instead of full schemas, cutting context from 60k tokens to 2-3k. The method uses a dedicated search tool to fetch sc…
ChatGPT was found to generate sexual and graphically violent images from a simple text prompt, according to AI security firm Mindgard. Researcher Jim Nightingale manipulated the chatbot with a viral "restore this photo" …
OpenAI demonstrated that reinforcement learning focused on beneficial traits like honesty and reliability produces AI alignment that generalizes across domains and withstands adversarial attacks. The lab's Alignment Trai…