Hugging Face

Hugging Face has published a new framework addressing how language models handle sensitive queries, arguing that blanket safety refusals are too blunt and often censor harmless information. * AI safety models traditionally block entire topics based on a single trigger word, killing useful nuance along with potential harm. * Hugging Face proposes

Hugging Face has released @huggingface/kernels, a collection of over 200 WebGPU kernels designed to accelerate machine learning directly inside web browsers. By leveraging the local GPU via WebGPU, developers can execute heavy AI workloads client-side without constantly pinging cloud servers. * WebGPU kernels shift heavy computing tasks from expensive cloud

IBM Research has rolled out its enterprise time series models natively integrated with Confluent, aiming to help businesses process streaming data into predictive insights without losing their minds in the pipeline. * IBM time series models now plug directly into Confluent for continuous data stream analysis. * Enterprises can supposedly forecast business

Hugging Face has introduced Funes, a new framework designed to give coding agents a persistent, user-owned memory. As developers increasingly rely on AI to write and debug software, the inability of these models to retain context across sessions has remained a persistent friction point. Funes attempts to solve this by

Hugging Face has demonstrated that a modest 350-million parameter model can be successfully fine-tuned for better structured outputs using just 100 GRPO steps, proving that brute-force size isn't always required for specific formatting tasks. * Using GRPO steps allows smaller models to achieve complex formatting without demanding massive compute

Hcompany has introduced NeoMME, a multimodal-native and multilingual encoder engineered to handle diverse data types across various languages with supposedly unprecedented efficiency. In a crowded landscape of universal AI models claiming to bridge global divides, this latest entry promises streamlined processing without the usual heavy computational baggage. * NeoMME integrates text

Hugging Face has demonstrated an unusual training pipeline, using TRL and OpenEnv to teach a standard coding model how to render watercolour paintings through executed code rather than direct pixel generation. * It forces text-based language models to reason spatially and visually via code execution. * The novelty highlights how reinforcement learning

The Allen Institute for AI has dropped BenchMIRT, a diagnostic framework designed to figure out what LLM benchmarks are actually measuring. As the industry drowns in endless claims of superhuman AI performance, this tool attempts to separate genuine capability from sophisticated data contamination and test memorisation. * AI labs keep inventing

Researchers at Multiverse Computing have introduced Quantization-Aware Healing, a clever technique that compresses heavy models down to 4-bit precision while somehow managing to outperform their full-sized, resource-hogging originals. * Quantization-Aware Healing defies the usual rule where compressing models means sacrificing accuracy for speed. * By shrinking memory footprints drastically, this technique makes

IBM has released a technical breakdown detailing the architecture and training methodology behind its Granite 4.2 LLMs, offering a rare peek behind the corporate curtain of enterprise AI development. * IBM Granite 4.2 LLMs continue the corporate tradition of wrapping standard transformer architectures in layers of compliance-friendly marketing. * The