Toss analyzed the discrepancy between public benchmarks and actual service performance, establishing Toss Benchmark as a domain-specific evaluation standard.
The local LLM agent extension for VS Code has officially launched on the marketplace, adding multi-turn conversation and GitHub integration features.
Researchers at the University of Manchester have developed a high-resolution air pollution prediction model for the entire UK using NVIDIA Earth-2.
AWS Network Firewall has launched a feature that controls traffic for EKS/ECS workloads using attributes, eliminating the need to manage IP lists.
The VisTW benchmark for evaluating VLM performance in Traditional Chinese environments has been released, with Twinkle Eval providing support.
The Rust-based tool llmfit has released a feature that estimates generation speed and compatibility from hardware specs before downloading local LLM models.
The average humor preference agreement rate among LLMs is 75%, with a strong preference for absurd humor confirmed.
The unified Claude app, rolling out to Pro and Max users, introduces beta features for generating documents and presentations directly within the chat interface.
Instinct has launched Concierge, a feature capable of handling phone calls and making reservations.
Devin has released the 'Code Scans' feature, which utilizes an Agentic MapReduce architecture to investigate entire codebases and generate pull requests.
Google has launched a private preview of the Agent Anomaly Detection feature on the Gemini Enterprise Agent Platform to detect abnormal agent behavior.
Google DeepMind has established the 'DeepMind Institute' to research and discuss the technical and societal impacts of AGI.
Ant Group has released the finance-specialized open-weight model 'Ling-3.0-flash-Fin', achieving high efficiency with 5.1B active parameters.
Transluce has proposed four embedded evaluator pilots to monitor misalignment and internal manipulation risks in undisclosed AI models.
Anthropic has integrated Claude Cowork and Chat into a unified Claude, launching Docs and Slides features in beta for paid plans.
TypeSafe AI has released Jev, a decision model that provides typed probabilistic answers without text generation, on AI Gateway.
Apple researchers confirmed that injecting specific values into LLMs improves safety but also increases sycophancy.
A 4B parameter RL model has been released, reducing Postgres query planning latency by 44.7% and achieving up to 1.81x acceleration.
A study featured on the Apple Research blog proposes TS-DFM to improve the trajectory quality of discrete flow matching, reducing perplexity by 32% compared to a 1,024-step teacher model with 8-step inference and improving speed by 128x.
Apple researchers have released DACA-GRPO, a method that reduces reinforcement learning bias in diffusion language models.
Noam Brown identified recursive self-improvement as OpenAI's top priority, predicting that AI will surpass human researchers within the next 1-2 model releases.
Odyssey has released Odyssey-3, a general-purpose physical intelligence model designed to control robots, autonomous vehicles, and games.
OpenArm has released an open-source humanoid arm for teleoperation and imitation learning.
Reallocating computation from easy problems to hard ones improved LLM reinforcement learning performance on difficult tasks.
AI Briefing