News Hub

LangSmith Preview Builds enable teams to test agent changes in temporary production-like deployments before merging.

Generative recommenders shift recommender systems from traditional approaches using LLM-inspired methods for scale.

Mistral AI offers agentic search to help AI systems navigate, read, and verify complex documents.

OpenAI launched AI Futures blog to explore how transformative AI could reshape power, governance, economy, and individual freedom.

Apple research uses lexical interventions for cross-lingual knowledge transfer with scarce target language data.

LFM2.5-DSpark achieves up to 3.2x faster inference from H100 to MacBook.

Apple research addresses scaling laws for mixture pretraining when target data is limited.

Mooncake converts fragmented rollout data into efficient bulk I/O operations.

Stampli used ChatGPT to compress weeks of production launch work into days.

Apple researchers applied iterative pseudo-labeling to improve speech recognition for Mandarin-English code-switching using unlabeled data.

NVIDIA Holoscan is a platform for building real-time AI applications at the edge with CLI tools and AI coding agents.

SPADE uses self-play RL to generate adaptive goals for language agents, expanding goal distribution as the learner scales.

ADEPT is an RL framework that pretrains dexterous robot policies on generic tasks then post-trains for specific long-horizon tasks.

Group-calibrated on-policy distillation uses task-level verifiers instead of token-level guidance for long-context reasoning tasks.

Two fine-tuning strategies for sound search by vocal imitation: contrastive learning and joint contrastive-triplet learning.

NVIDIA FLARE enables federated training of vision-language models across distributed data held by multiple institutions.

ChildSafeAds shared task detects commercial content in YouTube videos reaching children using 3,360 videos from 939 channels.

VLA framework monitors and steers hidden communication channels in language-model agents to prevent covert coordination.

Multiple Intel AI PCs with integrated GPUs and NPUs can serve large LLMs through pipeline parallelism over network.

Precision of model outputs, not capability, should be the frontier metric for comparing AI systems.