Hugging Face

The UK AI Safety Institute and EvalEval have teamed up to tackle the notoriously slippery problem of AI benchmark reproducibility, attempting to inject a bit of verifiable truth into an industry built on vibes and marketing claims. * Brings actual methodology to AI evaluation instead of just trusting vendor press releases.

Multiverse Computing has published a new approach to pruning large language models, framing transformer block removal as an Ising optimization problem. By treating neural network layers like interacting magnetic spins, researchers claim they can strip out redundant architecture with mathematical rigour rather than just guessing which layers to drop. * Brings

IBM Research has introduced ALTK Evolve, a new framework designed to tackle a fundamental flaw in modern artificial intelligence: the fact that autonomous agents can pass a benchmark once and then completely fail the exact same task five minutes later. By focusing on consistency rather than just peak performance, the

Hugging Face has released a new setup for Async GRPO with LoRA across HF Jobs that bypasses the usual headaches of distributed GPU synchronization. By swapping out NCCL for a bucket and a proxy, they have made scaling Reinforcement Learning from Human Feedback slightly less painful for those without a

Liquid AI has released LFM2.5-DSpark, boasting up to 3.2x faster inference speeds for AI workloads that allegedly refuse to slow down. As the hype cycle demands perpetual acceleration, this latest drop promises performance gains without requiring you to completely remortgage your server rack. * Faster inference numbers always look

A recent analysis from Hugging Face has taken a hard look at how automatic speech recognition models are evaluated, revealing that impressive leaderboard scores often owe more to clever benchmark optimization than genuine technological leaps. * Evaluation datasets are increasingly treated as a study guide rather than a blind test. * Small

Papers with Code has overhauled its search functionality, relying on Hugging Face inference endpoints, jobs, and storage buckets to index and retrieve machine learning literature. Instead of forcing researchers to sift through endless pre-print servers, the platform now uses modern infrastructure to match queries against datasets and benchmarks with actual

Hugging Face has released a new guide aimed at simplifying how developers wire together, run, and deploy machine learning workflows using Gradio. The documentation focuses on turning disparate python scripts into interactive web applications that can handle complex data pipelines without requiring a complete frontend rewrite. * Developers can finally stop

Open-source staple AUTOMATIC1111 is getting a structural makeover courtesy of a Gradio Workflow rebuild, bringing a fresh coat of paint to everyone's favourite clunky Stable Diffusion interface. * Open-source developers continue reinventing the wheel to make local AI generation slightly less hostile to human eyes. * The Gradio Workflow integration

IBM Research has released the Granite Time Series PatchTST-FM-r2 model, bringing state-of-the-art forecasting capabilities to the open-source community paired with a surprisingly permissive commercial-friendly license. * IBM Granite Time Series offers state-of-the-art performance for enterprise forecasting tasks without the usual restrictive licensing hurdles. * The PatchTST-FM-r2 architecture allows businesses to reliably predict