英特尔研究院

The future of #AI isn’t just bigger; it’s smarter, thanks to ULTRA—created by Intel Labs. See how we use knowledge graphs, zero-shot inference, and fine-tuning to change the game in structured decision-making and LLMs. https://lnkd.in/gAxAcWxM ?? #LLM #Intel

Introducing Cooperative Vectors on DirectX! Cooperative Vectors enable multiplication of matrices with arbitrarily sized vectors to be run on any shader stage. This new feature can be used for AI training and real-time inferencing, specifically per-pixel inferencing. Using Cooperative Vectors, developers can unleash the full power of XMX units in select Intel Core Ultra processors and modern Intel GPUs such as Intel Arc graphics (A series and B series). Learn more about this new feature in the blog from Jonathan Dupuy. https://intel.ly/4j02fiz #Graphics #DigitalDesign #Research

Here’s a novel prompting technique that makes AI interrogate itself before answering: SQuARE. SQuARE, or Sequential Question Answering Reasoning Engine for Enhanced Chain-of-Thought in Large Language Models, makes LLMs generate and solve multiple smaller questions rather than tackling a question directly, leading to deeper, more accurate reasoning. The results? Tested on Llama 3 and GPT-4o across various datasets, SQuARE outperformed traditional chain-of-thought and rephrase-and-respond methods. Read the paper from Intel Labs researchers Daniel Fleischer, Moshe Berchansky, Gad Markovits, and Moshe Wasserblat to learn more about how SQuARE can help AI think more critically with better decision-making, smarter assistants, and next-level reasoning. https://intel.ly/4i4vnWb #AI #LLMs #NLP

“We are pleased to see that the Linux Foundation SMaRT-5G initiative, for which we provide an innovative AI/ML based 5G energy-saving rApp to improve energy efficiency and performance, will have the opportunity to be tested in Rakuten Mobile’s Lab and get one step closer to be deployed in networks in operation. Energy savings of 5G systems is a key usage model called for by Open RAN (ORAN) operators and customers.” -Pranav Mehta, VP, Intel Labs Joining Intel Labs’ efforts with Aether 5G (a Linux Foundation project) and partners like Rakuten Mobile, Inc., Rakuten Symphony, Rimedo Labs, Tietoevry, and VIAVI Solutions is a great way forward under the SMaRT-5G initiative. Read the full article to learn how we’ll create energy savings and focus on traffic steering capabilities to improve efficiency, reduce costs, and enhance user experience through sustainable practices. https://intel.ly/3DBg885 #5G #ArtificialIntelligence #Sustainability

AI is evolving, and we’re now on a fast track to agentic AI systems capable of autonomously performing actions without direct commands. This advancement is dramatically increasing the need for new approaches to responsible AI development, as current external processes will not be enough to keep these new agentic AI models in check. Through a series of blog posts Gadi Singer, VP and Director of Emergent AI Research at Intel Labs, will dive into the topic of intrinsic AI alignment and explore the role it plays in guiding, monitoring, and aligning rapidly improving machine intelligence. Read Part 1 now. https://intel.ly/43wslp1 #AgenticAI #AISecurity #ResponsibleAI

Locally-adaptive quantization is a critical ingredient for high performance in our Scalable Vector Search library. This industry-first method overcomes the challenges of existing compression techniques, lowering the memory footprint and preserving the efficiency of the similarity computation. Our latest blog has all the details: https://intel.ly/3XtRLzH #VectorSearch #Developer #VectorQuantization

Happy to see the growing excitement around collegiate overclocking!

DeepSeek-R1, the largest open-source mixture of experts (MoE) model, demonstrates reasoning capabilities comparable to proprietary frontier models. This prompted a question from our team: Given the model’s leading performance on reasoning benchmarks, could we observe any evidence of semantic specialization in the experts? Read the full paper to see what Intel Labs researchers discovered: https://intel.ly/4h0RtY9 #OpenSource #DeepSeekR1 #ArtificialIntelligence

Agentic AI is quickly becoming the next evolution of AI. Learn more about the benefits and potential pitfalls in the blog coauthored by Intel and the National Institute of Standards and Technology (NIST). #AgenticAI #Innovation #Research