Intel Labs

Universal Assisted Generation (UAG) is a groundbreaking method for extending assisted generation support to small language models from any model family! Developed in collaboration with Hugging Face, UAG makes it possible to accelerate inference from any decoder or Mixture of Experts models by ~2x with almost zero overhead. Learn more from the researchers behind the method in this blog. https://intel.ly/40GwOnN Daniel Korat, Oren Pereg, Moshe Berchansky, Jonathan Mamou, João Gante, Lewis Tunstall, Nadav T. and Moshe Wasserblat #AI #AssistedGeneration #LLMs

Maintaining data integrity, privacy, and accuracy is at the heart of our security research. From innovations that help protect sensitive workloads to the development of AI methods that will help restore the public’s trust in media, Intel Labs researchers remain at the forefront of security and privacy research and development. Learn more here: https://lnkd.in/g6Q8ik9P   #Developer #ArtificialIntelligence #Security

LLMs are all the rave - but scaling foundation models in a computationally and memory efficient manner is critical to enabling disruptive applications on consumer grade hardware.  Learn how Intel Labs researchers are tackling these challenges in this video with Somdeb Majumdar, Director of the AI Lab. https://intel.ly/3BZUY25 #LLMs #LCLMs #Innovation

The Intel Rendering Toolkit libraries are now open for community contribution! With the internal development and feature branches now hosted directly on GitHub under Apache 2.0, they're fully open and always up to date. Along with improved features and functionality, we hope these changes bring increased development transparency and cross-platform CI for validation of contributors. You can see the full list of libraries here: https://intel.ly/3Aki8je #Graphics #Research #OpenSource

While quantum computing holds promise for breakthroughs in drug discovery, materials design, and more, it also poses a threat to current cryptography. Classical public-key algorithms must be replaced by the National Institute of Standards and Technology (NIST) post-quantum cryptography (PQC) standards to stay secure. In support of this, Intel Labs research scientist Christoph Dobraunig, together with collaborators from industry and academia, developed the SLH-DSA (SPHINCS+) algorithm for digital signatures. Learn about this new standard and the role it plays in helping the industry prepare for the future. https://intel.ly/3NCa4gT #Research #CyberSecurity #QuantumComputing

Tackle the key challenges in long-context model evaluation with HELMET (How to Evaluate Long-Context Models Effectively and Thoroughly). This comprehensive new benchmark, created in collaboration with Princeton Language and Intelligence, supports ≥128K token lengths across 7 diverse applications, offering more reliable and consistent rankings of frontier LCLMs. Get the code here. https://intel.ly/3YDt19v Minmin Hou, Ke Ding, Daniel Fleischer, Peter Izsak, Moshe Wasserblat, Howard Yen, Tianyu Gao, Danqi Chen #MachineLearning #NLP #LCLMs

Through collaboration, anything is possible. Learn more about the innovative research happening with partners such as the National Science Foundation (NSF) and the Defense Defense Advanced Research Projects Agency (DARPA). Learn more: https://intel.ly/3YkSkgo #AI #CyberSecurity #HPC #Research

Have you heard about Dynamic Speculative Decoding? 🤯

We’re excited to introduce RAG-FiT! This new open-source framework helps improve how LLMs use external information by fine-tuning models on specially created RAG-augmented datasets. The Python-based framework facilitates fast prototyping and experimentation with various RAG techniques. By integrating data creation, training, inference, and evaluation into a single workflow, RAG-FiT assists in the creation of data-augmented datasets for training and evaluating LLMs in RAG settings. Learn more here. https://intel.ly/3BAJAJX #Developer #RAG #LLMs